Intelligent question and answer method and system based on large language model in post-loan risk management

Abstract

This invention discloses an intelligent question-answering method and system based on a large language model for post-loan risk management, belonging to the field of artificial intelligence technology. The method generates standardized indicators, user scores, and risk signals by constructing a unified access and standardized indicator processing pipeline for multi-source heterogeneous data, and outputs structured early warning results through an early warning rule engine. The above data is stored in a query database and then encapsulated into structured data interfaces for use by the large language model according to risk dimensions. After obtaining user questions, the question-splitting agent performs intent recognition and task decomposition, and intelligent routing assigns the tasks to corresponding professional analysis agents. Each agent calls the structured data interface to obtain enterprise data and integrates it with knowledge base document recall, while simultaneously calling the large language model to generate structured analysis results for each dimension. Finally, these results are aggregated into a unified structured comprehensive assessment report and streamed out. This invention improves the efficiency and accuracy of post-loan risk analysis.

Description

Technical Field

[0001] This application relates to the fields of financial technology and artificial intelligence, specifically to intelligent question-answering methods and systems based on large language models in post-loan risk management. Background Technology

[0002] Post-loan risk management is a crucial link in the entire process of commercial bank lending. Its core lies in continuously monitoring the borrower's operational status, financial health, and various risk signals to enable timely risk mitigation measures. With the continuous enrichment of enterprise data dimensions and increasingly stringent regulatory requirements, account managers need to comprehensively analyze multi-source, heterogeneous data in post-loan management, including enterprise business registration information, legal records, financial data, guarantee information, public opinion dynamics, and credit records. However, the diverse sources, inconsistent formats, and fragmented access methods of this data present technical challenges for post-loan risk analysis.

[0003] To address the aforementioned issues, some practitioners have attempted to leverage the semantic understanding and generation capabilities of large language models to improve the efficiency of information retrieval and analysis. For example, patent application CN121681780A proposes a credit review knowledge-based question-answering scheme based on a large model. The core technical path of this scheme is as follows: First, the credit review question to be processed is obtained; core keywords are extracted using natural language processing algorithms; and the question is converted into a deep semantic vector using a semantic vector model. Simultaneously, the question is input into an intent classification model to generate filter tags for full-database or specific-database retrieval. Then, the core keywords, deep semantic vectors, and filter tags are input into a distributed full-text search engine. Keyword and vector searches are performed in the target knowledge base selected based on the filter tags to obtain a preliminary search result set. Next, a comprehensive matching score is calculated based on weighting factors to select the final reference content. Finally, the final reference content is structurally encapsulated, concatenated with credit prompt words, and input into the large model to obtain a structured credit review answer. This scheme improves the accuracy of knowledge retrieval through dual-path retrieval using keywords and semantic vectors, and standardizes the output boundaries of the large model to a certain extent through prompt word engineering.

[0004] However, the aforementioned existing technical solutions still have the following technical shortcomings when applied to the dynamic, multi-source, and highly regulated scenario of post-loan risk management.

[0005] First, there is a lack of capability for real-time access to multi-source data and standardized indicator processing for dynamic enterprise data. Existing knowledge bases primarily target static historical credit review documents (such as regulations and business rules), with data sourced from offline imported documents. These documents are parsed using batch conversion tools and then stored in a distributed search engine. This architecture cannot access dynamic data from internal credit systems, core transaction systems, and CRM systems in real time, nor can it synchronously acquire changes in external business, judicial, and public opinion data. More critically, this solution lacks a data pipeline for processing multi-source raw data into standardized risk indicators, such as quantifiable and comparable risk indicators like debt-to-asset ratio, interest coverage ratio, loan-to-deposit ratio, and litigation amount. Instead, it directly uses document fragments as search objects. This forces large models to rely solely on qualitative descriptions or scattered raw data fragments when answering post-loan risk questions, making it difficult to obtain accurate indicator values ​​processed by business logic, severely limiting the accuracy of quantitative risk analysis.

[0006] Second, there is a lack of a fundamental mechanism to suppress data illusion in large models. Existing technologies regulate the output of large models through cue word constraints and structured encapsulation of reference content, but these constraints only apply to the input and output ends of model inference, failing to address the illusion problem at the data source. Specifically, the large model in this approach still relies on document fragments stored in the knowledge base to answer specific numerical questions (such as "What is the company's current debt-to-equity ratio?"). If the knowledge base lacks the company's latest financial indicator data, the large model may still fabricate data based on its training memory. In post-loan risk management, corporate financial data, the number of legal cases, and the amount of guarantees are all precise numerical information that the large model cannot accurately obtain based on its own knowledge. Fabrication of such data will directly lead to incorrect risk judgments by account managers. Existing technologies have failed to construct a data acquisition mechanism that "replaces model memory with real-time interface data," thus failing to cut off the root cause of data illusion.

[0007] Third, the risk analysis is limited to a single dimension, failing to achieve multi-dimensional parallel analysis and structured output. Current technologies generate answers by retrieving relevant institutional clauses or regulations from a knowledge base based on user questions, then integrating them into a structured answer using a large model. Essentially, this remains a question-and-answer style of information retrieval and integration, lacking a multi-dimensional, systematic analytical framework for post-loan risk management scenarios. Post-loan risk assessment requires parallel analysis from multiple dimensions, including financial health, legal risks, guarantee relationships, public opinion dynamics, operational stability, and industry trends, and a comprehensive risk judgment based on the conclusions of each dimension. Current technologies do not break down risk analysis tasks into multiple independent analysis units according to professional dimensions, nor do they design mechanisms for collaborative invocation and result aggregation between multiple analysis units. This results in the answer quality being highly dependent on the accuracy of the user's questions, making it difficult to output a comprehensive structured risk assessment report. Furthermore, current technologies lack progressive streaming output and multi-turn conversation management mechanisms, failing to support user follow-up questions and in-depth analysis based on the results, leading to insufficient interactive experience and business adaptability.

[0008] In summary, how to achieve real-time access and standardized indicator processing of multi-source heterogeneous dynamic data in post-loan risk management scenarios, suppress the numerical illusion of large models from the data source, and output structured and interactive risk assessment results through a multi-dimensional parallel analysis framework has become an urgent technical problem to be solved. Summary of the Invention

[0009] To address the aforementioned technical problems in existing technologies, this invention provides an intelligent question-answering method and system based on a large language model for post-loan risk management.

[0010] The intelligent question-answering method based on a large language model for post-loan risk management provided by this invention includes the following steps: Step S1: Construct a unified access and standardized indicator processing pipeline for multi-source heterogeneous data. This involves acquiring internal data from the bank's credit system, core transaction system, and CRM system via a data acquisition layer, and external data from sources such as industry and commerce, judiciary, taxation, credit reporting, and public opinion via external data interfaces. The acquired multi-source raw data is cleaned, aligned, and integrated according to a preset business logic, and processed into multi-dimensional standardized indicators. These standardized indicators include at least financial, deposit, guarantee, credit, transaction, industry and commerce, judicial, public opinion, credit reporting, and operational indicators. Based on these standardized indicators, feature selection is performed using the XGBoost algorithm, indicator selection is performed using WOE and IV value analysis, and a scorecard model is constructed using logistic regression to generate user scores. Finally, according to preset risk signal definition rules, event-based risk assessment is performed on the standardized indicators to generate risk signals.

[0011] Step S2: Based on preset indicator thresholds and scoring segments, the early warning rule engine performs deterministic risk assessment on the standardized indicators and user scores, and outputs a structured early warning result. The early warning result includes the early warning level, triggering indicator, and triggering reason. The judgment logic of the early warning rule engine includes single indicator threshold triggering, multi-indicator combination triggering, and trend change triggering.

[0012] Step S3: Establish a query database, store the standardized indicators, user scores, risk signals, and early warning results into the query database, and encapsulate them into structured data interfaces for use by large language models according to risk dimensions based on the query database; each structured data interface corresponds to a risk dimension, and the interface internally gathers indicator fields, signal fields, and judgment result fields describing the enterprise's risk status under that dimension.

[0013] Step S4: Obtain the question input by the user, and use the question splitting agent to perform intent recognition on the question, parse out the analysis subject and analysis dimensions of the question, and split the question into at least one sub-task according to the analysis dimensions; the analysis dimensions include macroeconomic dimension, policy analysis dimension, industry analysis dimension, financial health dimension, business status dimension, judicial risk dimension, public opinion monitoring dimension, related risk dimension, and early warning assessment dimension.

[0014] Step S5: Based on the analysis dimensions corresponding to each subtask, each subtask is assigned to a corresponding professional analysis agent through intelligent routing. The professional analysis agent predefines its professional role positioning, core analysis framework, and behavioral boundary constraints based on the prompt word engineering. The behavioral boundary constraints include role definition constraints, data usage constraints, and professional knowledge constraints. The data usage constraints limit all analyses to be based on the enterprise data returned by the structured data interface, and prohibit the large language model from fabricating data on its own.

[0015] In step S6, each of the professional analysis agents calls the structured data interface to obtain the enterprise data corresponding to the analysis subject. At the same time, the knowledge base recall module performs semantic retrieval on the user question, retrieves professional knowledge documents related to the sub-task from the knowledge base, and concatenates the retrieved professional knowledge documents, the enterprise data returned by the structured data interface, and the prompt words corresponding to the professional analysis agent according to a preset template to form the complete input of the professional analysis agent.

[0016] Step S7: Each of the professional analysis agents calls the large language model in parallel, and sends the complete input into the large language model, so that the large language model generates structured analysis results of the corresponding dimension based on the enterprise data returned by the structured data interface and the professional knowledge documents recalled by the knowledge base under the behavioral boundary constraints of the prompt words.

[0017] Step S8: By comprehensively summarizing the structured analysis results received from each of the professional analysis agents, the structured analysis results of each dimension are aggregated into a unified structured comprehensive evaluation report according to the preset report structure, and displayed to the user in stages according to the progressive streaming output method; the progressive streaming output includes the output of summary information in the first stage, the output of analysis content in the second stage, and the output of comprehensive risk rating and disposal suggestions in the third stage.

[0018] This invention also provides an intelligent question-answering system based on a large language model for post-loan risk management, comprising: The data acquisition module is used to construct a unified access and standardized indicator processing pipeline for multi-source heterogeneous data. It acquires internal data from the bank's credit system, core transaction system, and CRM system, and external data from external data sources such as industry and commerce, judiciary, taxation, credit reporting, and public opinion through external data interfaces. The acquired multi-source raw data is cleaned, aligned, and integrated according to preset business logic, and processed into multi-dimensional standardized indicators. These standardized indicators include at least financial, deposit, guarantee, credit, transaction, industry and commerce, judicial, public opinion, credit reporting, and operational indicators. The data acquisition module is also used to perform feature filtering using the XGBoost algorithm, indicator filtering using WOE and IV value analysis, construct a scorecard model using logistic regression to generate user scores, and perform event-based risk judgment on the standardized indicators according to preset risk signal definition rules to generate risk signals.

[0019] The early warning determination module is used to perform deterministic risk determination on the standardized indicators and user scores based on preset indicator thresholds and scoring segments through an early warning rule engine, and output structured early warning results. The early warning results include early warning level, triggering indicator and triggering reason. The determination logic of the early warning rule engine includes single indicator threshold triggering, multi-indicator combination triggering and trend change triggering.

[0020] The data interface module is used to establish a query database, store the standardized indicators, user scores, risk signals, and early warning results into the query database, and encapsulate them into structured data interfaces for use by large language models according to risk dimensions based on the query database. Each structured data interface corresponds to a risk dimension, and the interface internally aggregates indicator fields, signal fields, and judgment result fields that describe the enterprise's risk status under that dimension.

[0021] The intent parsing module is used to obtain the question input by the user, and to identify the intent of the question through the question splitting agent. The module parses out the analysis subject and analysis dimension of the question, and splits the question into at least one sub-task according to the analysis dimension. The analysis dimension includes macroeconomic dimension, policy analysis dimension, industry analysis dimension, financial health dimension, business status dimension, legal risk dimension, public opinion monitoring dimension, related risk dimension, and early warning assessment dimension.

[0022] The intelligent routing module is used to allocate each subtask to a corresponding professional analytical agent based on the analysis dimension corresponding to each subtask. The professional analytical agent predefines its professional role positioning, core analysis framework, and behavioral boundary constraints based on the prompt word engineering. The behavioral boundary constraints include role definition constraints, data usage constraints, and professional knowledge constraints. The data usage constraints stipulate that all analyses must be based on the enterprise data returned by the structured data interface, and prohibit the large language model from fabricating data on its own.

[0023] The data assembly module is used by each of the professional analysis agents to call the structured data interface to obtain the enterprise data corresponding to the analysis subject. At the same time, the knowledge base recall module performs semantic retrieval on the user question, retrieves professional knowledge documents related to the sub-task from the knowledge base, and concatenates the retrieved professional knowledge documents, the enterprise data returned by the structured data interface, and the prompt words corresponding to the professional analysis agent according to a preset template to form the complete input of the professional analysis agent.

[0024] The parallel analysis module is used by each of the aforementioned professional analysis agents to call the large language model in parallel, sending the complete input into the large language model, so that the large language model generates structured analysis results of the corresponding dimensions based on the enterprise data returned by the structured data interface and the professional knowledge documents recalled by the knowledge base, under the behavioral boundary constraints of the prompt words.

[0025] The results aggregation module is used to receive the structured analysis results output by each professional analysis agent through the comprehensive summary agent, aggregate the structured analysis results of each dimension into a unified structured comprehensive evaluation report according to the preset report structure, and display it to the user in stages according to the progressive streaming output method. The progressive streaming output includes the output of summary information in the first stage, the output of analysis content in the second stage, and the output of comprehensive risk rating and disposal suggestions in the third stage.

[0026] Compared with the prior art, this application has the following advantages: 1. Data Acquisition Level: Dynamic Multi-Source Access and Standardized Indicator Processing Existing knowledge base construction technologies rely on the batch import of offline documents. Credit business documents are parsed into text fragments using format conversion tools and then stored in a distributed search engine. This approach can only handle static historical policy documents and regulations, and cannot access dynamic data from internal credit systems, core transaction systems, and CRM systems in real time. It also cannot synchronously acquire real-time external data from industry and commerce, judicial, and public opinion sources. Due to the lack of a pipeline for quantifying raw data, existing technologies directly use document fragments as retrieval objects. Large models can only reason based on fragmented qualitative descriptions, making it difficult to obtain accurate quantitative indicators processed by business logic.

[0027] This invention establishes a unified access and standardized indicator processing pipeline for multi-source heterogeneous data at the data acquisition layer. On one hand, it acquires internal data from the bank's credit system, core transaction system, and CRM system. On the other hand, it synchronously acquires external data from external data sources such as industry and commerce, judiciary, taxation, credit reporting, and public opinion through external data interfaces. After cleaning, aligning, and merging the acquired multi-source raw data, it processes it according to preset business logic into standardized indicators covering more than ten dimensions, including finance, deposits, guarantees, credit, transactions, industry and commerce, judiciary, public opinion, credit reporting, and operations. Simultaneously, this invention utilizes XGBoost for feature selection, WOE and IV value analysis for indicator selection, and logistic regression to build a scorecard model to generate user scores. It also performs event-based risk assessment on the standardized indicators based on preset risk signal rules. These settings enable the large model to directly retrieve accurate quantitative indicators processed by business logic during question-and-answer processes, rather than relying on model memory or fragmented document snippets for inference, thus solving the data illusion problem of large models from the data source.

[0028] 2. Risk Assessment Level: Deterministic Early Warning Rule Engine Current technologies primarily rely on retrieval quality optimization and prompt word constraints to ensure the output quality of large models. This involves filtering reference content through a dual-path retrieval using keywords and semantic vectors, and then using prompt words to standardize the output format and content boundaries of the large model. However, for deterministic judgment tasks such as risk level determination that require strict adherence to business rules, current technologies still rely on large models to infer from the retrieved document content. Due to potential biases in the large model's understanding of risk control rules, the output results are subject to uncertainty.

[0029] This invention introduces an independent early warning rule engine between the data acquisition layer and the intelligent analysis layer. Based on preset indicator thresholds and scoring segments, this engine performs deterministic risk assessment on standardized indicators and user scores, outputting a structured early warning result that includes the early warning level, triggering indicators, and triggering reasons. The early warning rule engine's judgment logic includes three types: single indicator threshold triggering, multi-indicator combination triggering, and trend change triggering, transforming risk level determination from "knowledge base text retrieval + large model interpretation" to "deterministic rule calculation." The engine's output is directly injected into the context of the large model as structured data. The large model does not need to analyze key rule judgments; it only needs to perform a comprehensive interpretation based on the engine's judgment, thereby eliminating the knowledge illusion caused by the large model's misunderstanding of risk rules.

[0030] 3. Analysis Architecture Level: Multi-Agent Parallel Analysis System Existing technologies employ a single-model question-and-answer approach. The logic for generating answers involves retrieving relevant policy clauses or regulations from a knowledge base based on the user's question, then integrating these into a larger model to output a structured answer. This approach is essentially a question-and-answer style of information retrieval and integration, lacking a multi-dimensional, systematic analytical framework for post-loan risk management scenarios. The comprehensiveness and depth of risk analysis heavily rely on the accuracy of the user's questions.

[0031] This invention constructs a multi-agent parallel analysis system based on prompt word engineering at the intelligent analysis layer. Internally, it is divided into macroeconomic intelligent agents, policy analysis intelligent agents, industry analysis intelligent agents, financial health intelligent agents, operational status intelligent agents, judicial risk intelligent agents, public opinion monitoring intelligent agents, related risk intelligent agents, early warning assessment intelligent agents, problem decomposition intelligent agents, and comprehensive summary intelligent agents according to the professional dimensions of post-loan risk identification. Each intelligent agent predefines its professional role, core analysis framework, and behavioral boundary constraints through prompt word engineering. These behavioral boundary constraints include role definition constraints (limiting the answer identity), data usage constraints (restricting all analysis to be based on interface data, prohibiting fabrication), and professional knowledge constraints (strictly adhering to the judgment criteria for early warning rule interface calls). This multi-agent system concretizes post-loan risk analysis within specific professional dimensions, making the analysis of each dimension more targeted, while effectively suppressing the illusion of divergence that can easily arise from large models operating within a single model.

[0032] 4. Interactive Output Level: Structured Streaming Output and Multi-Turn Sessions vs. Single-Package Output Existing technologies encapsulate the final reference content in a structured manner according to a preset format and then concatenate it with prompts from a large model to output a one-time structured answer. This lacks a phased display and multi-round interaction mechanism, preventing users from asking follow-up questions and drilling down deeper based on the analysis results.

[0033] This invention employs an SSE streaming output method at the decision application layer to display analysis results in stages: the first stage outputs summary information within 2-3 seconds, allowing users to quickly understand the overall analysis framework; the second stage outputs analysis data for each dimension sequentially in a streaming manner, with the front end rendering immediately upon reaching a given dimension; the third stage outputs a comprehensive risk rating and handling recommendations. Simultaneously, the system maintains the analysis subject, analyzed dimensions, and historical question-and-answer records for the current session, supporting users to continue asking follow-up questions based on the analysis results. When asking follow-up questions, the system understands the intent based on the context and re-executes the analysis process, eliminating the need for users to repeat their descriptions.

[0034] The technical features described above in this invention are not simply a combination of isolated technical means, but rather an overall technical solution that is organically integrated around the core objective of "suppressing the illusion of large models and realizing multi-dimensional intelligent Q&A for post-loan risk" and follows a three-layer architecture of "data acquisition layer → intelligent analysis layer → decision application layer". Each layer has a close synergistic effect.

[0035] The collaboration between data acquisition and the early warning rule engine constitutes a dual guarantee mechanism for the data source. Data acquisition provides precise quantitative data input to the large model through a standardized indicator processing pipeline, while the early warning rule engine, based on this, separates rule judgment from model reasoning. Together, they eliminate numerical illusions and rule comprehension biases in the large model from the data source. In intelligent analysis, the collaboration between the multi-agent system and the data interface constitutes a deterministic guarantee mechanism for the analysis path. The analysis framework of each specialized agent is bound to its corresponding data interface, and the behavioral boundaries of the agents are strictly constrained by prompt word engineering. The analysis process does not rely on the free divergence of the model but follows a pre-set specialized analysis path for structured reasoning. The complementarity between the knowledge base recall module and the data interface constitutes a dual channel for knowledge support. The data interface provides real-time quantitative risk data for enterprises, while the knowledge base provides professional knowledge such as industry benchmarks, regulatory laws, and historical cases. These two are combined into a complete input, ensuring that the large model's responses have both precise data support and a deep professional knowledge background. The streaming output and multi-round conversation mechanism of the decision application layer form a closed loop with the above layers. Users ask follow-up questions based on the output results, and the system re-executes the complete process from data collection to intelligent analysis, so that the entire technical solution can continuously adapt to the dynamic business needs of post-loan management. Attached Figure Description

[0036] Figure 1 This is the main flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the overall architecture of the system of the present invention. Detailed Implementation

[0037] To enable those skilled in the art to more clearly understand the technical solutions of the present invention, some terms involved in the present invention will be explained before the detailed description of specific embodiments. It should be noted that those skilled in the art can understand the meaning of the following terms based on the context, and the explanations of the following terms are only for assisting in understanding the present invention and should not be regarded as limiting the scope of protection of the present invention.

[0038] Early Warning Rule Engine: This module performs deterministic risk assessments based on preset indicator thresholds and scoring segments, using standardized indicators and user ratings. It transforms risk level determination from knowledge base text retrieval and model interpretation to deterministic rule calculations. Its judgment logic includes single-indicator threshold triggering, multi-indicator combination triggering, and trend change triggering.

[0039] Structured Data Interface: A data access interface encapsulated for large language models, categorized by risk dimensions. Each interface corresponds to a risk dimension, internally aggregating indicator fields, signal fields, and judgment result fields describing the enterprise's risk status under that dimension. Large language models obtain enterprise data by calling these interfaces, rather than directly querying the database or relying on model memory.

[0040] Intelligent Agent: An AI agent unit built based on prompt words, with specific professional roles and analytical responsibilities. Each agent defines its professional role, core analytical framework, and behavioral boundary constraints through system prompt words, and calls a large language model at runtime to complete specific dimension analysis tasks.

[0041] Cue word engineering: a technical approach to standardize the behavior of large language models through the design of system cue words. Cue words include role definition constraints, data usage constraints, and professional knowledge constraints, used to limit the analytical identity, data source, and analytical path of large models.

[0042] Multi-agent collaboration: This technical architecture breaks down post-loan risk analysis tasks into multiple sub-tasks according to professional dimensions, which are executed in parallel by different professional agents, and the comprehensive summarizing agent aggregates the analysis results from each dimension into a unified report.

[0043] Progressive streaming output: This method uses server-sent events (SSE) technology to present the analysis results to the user in three stages: summarizing information, analyzing each dimension, and making a comprehensive evaluation.

[0044] Knowledge base retrieval: This is the process of retrieving professional knowledge documents related to the current analysis task from a knowledge base based on semantic retrieval of user questions. The retrieval process includes document parsing, document segmentation, document tagging, keyword retrieval, vector retrieval, and knowledge ranking.

[0045] Example 1 like Figure 1 As shown in the figure, this embodiment provides an intelligent question-answering method based on a large language model in post-loan risk management. The following describes the method in detail with specific implementation steps.

[0046] Step S1: Construct a unified access and standardized indicator processing pipeline for multi-source heterogeneous data to generate standardized indicators, user scores, and risk signals.

[0047] This step achieves unified access and standardized indicator processing of multi-source heterogeneous data through the data acquisition layer, specifically including two sub-stages: data access and data feature processing.

[0048] Data Access Sub-phase: The system adopts a distributed data integration framework to achieve unified access to multi-source heterogeneous data. For internal data sources, the system connects to the core credit database, core transaction database, and CRM database through a JDBC connection pool, and uses ChangeDataCapture technology to capture change logs in these databases, pushing change events to an Apache Kafka message queue. For external data sources, the system calls open interfaces of external data service providers such as industrial and commercial, judicial, tax, credit reporting, and public opinion services through HTTP RESTful APIs. Incremental data is retrieved every four hours using scheduled tasks. Interface calls employ an exponential backoff retry strategy, with an initial retry interval of one second, a maximum retry interval of sixty seconds, and a maximum number of retries of three, ensuring that the system can automatically recover data acquisition in the event of temporary failures of external interfaces. Internal CDC data and external API data are uniformly aggregated to the data bus of the data access layer. The data bus uses Apache Kafka as its core component, setting up fifteen partitions to support high-throughput parallel writes. The partition key uses the hash value of the enterprise's unified social credit code modulo 15, ensuring that all data records of the same enterprise are written to the same partition, thereby guaranteeing data order. After the raw data enters Kafka, it is consumed in parallel by ten consumer instances in the data cleaning consumer group. Each consumer instance is responsible for processing data from one or more partitions. The cleaning process includes the following three sub-operations: format standardization, which converts date fields from different sources to ISO8601 format; field mapping, which establishes a mapping relationship between source field names and target field names (e.g., mapping "entName" returned by the business registration interface to "company_name" within the system); and deduplication, which determines whether a record is duplicated by comparing the combination of the enterprise's unified social credit code, data date, and data source. If all three are identical, the record is considered duplicated and discarded. For example, a commercial bank's credit system processes approximately 500,000 loan accounts and 2 million transaction records daily, involving approximately 80,000 corporate clients. Taking a batch of 10,000 companies as an example, the total raw data volume is approximately 1.2 million records, and after cleaning and deduplication, approximately 1.05 million valid records are obtained.

[0049] Data Feature Processing Sub-stage: Cleaned data is written to the ODS (Operational Data Storage) layer of the data warehouse. ODS uses the Parquet columnar storage format for physical storage and is organized by date partitioning. The system reads data from the ODS layer and processes the multi-source raw data into multi-dimensional standardized indicators according to preset business logic through the indicator calculation engine. The indicator calculation engine is based on the Drools rule engine. Each standardized indicator corresponds to one Drools rule, which consists of a condition part and an action part. Taking the debt-to-asset ratio indicator as an example, the condition part of its Drools rule is defined as follows: the rule is activated when the input data contains the total liabilities and total assets fields, and both fields are neither empty nor zero. The action part is defined as follows: divide the total liabilities value by the total assets value, round the result to two decimal places, multiply by 100%, and generate the debt-to-asset ratio indicator value. Simultaneously, the data source for this indicator is recorded as "total liabilities are taken from line 45 of the balance sheet, and total assets are taken from line 30 of the balance sheet," and the calculation time is the current system timestamp. The system defines a total of eighty-seven Drools rules, corresponding to eighty-seven standardized indicators. The indicator calculation engine is scheduled to perform a full calculation at 2:00 AM daily. Simultaneously, when a company's basic data changes, the system triggers an incremental recalculation of the relevant indicators in real time. The recalculation scope is determined through an indicator dependency graph, which records the reference relationship between each indicator and a basic data field in the form of a directed acyclic graph. If a basic field is referenced by multiple indicators simultaneously, all indicators dependent on that field are marked as requiring recalculation. Each standardized indicator includes an indicator name, indicator value, data source, and calculation time. Specifically, financial dimension indicators include fourteen specific indicators such as the debt-to-equity ratio, net profit, and interest coverage ratio. The debt-to-equity ratio is calculated by dividing the company's total liabilities by its total assets; net profit is directly taken from the after-tax net profit item in the company's profit statement; and the interest coverage ratio is calculated by dividing earnings before interest and taxes (EBIT) by interest expense. Deposit dimension indicators include eight specific indicators such as the current account balance in the past month, the maximum value of the time deposit balance in the past twelve months, and the loan-to-deposit ratio. The loan-to-deposit ratio is calculated by dividing the company's loan balance by its deposit balance. The guarantee dimension indicators include seven specific indicators such as the number of collateral items, the number of guarantors, and the loan-to-value ratio. The loan-to-value ratio is calculated by dividing the appraised value of the collateral by the loan principal. The credit dimension indicators include nine specific indicators such as the number of current loan documents, credit utilization rate, and the proportion of normal loan documents. The credit utilization rate is calculated by dividing the used credit line by the total credit line. The transaction dimension indicators include eleven specific indicators such as the maximum transaction amount of lenders in the past thirty days, the monthly income-to-expenditure ratio, and the minimum transaction amount in the past six months. The monthly income-to-expenditure ratio is calculated by dividing the total monthly income by the total monthly expenditure. The business registration dimension indicators include six specific indicators such as the number of changes in the legal representative of the enterprise, the proportion of natural person shareholders in the enterprise, and the number of employees.The judicial dimension indicators include the number of days since the most recent judicial proceedings, the number of lawsuits involving the enterprise in the past year, and the number of loan lawsuits involving the enterprise, totaling twelve specific indicators. The public opinion dimension indicators include the number of negative public opinions in the past six months, the number of negative public opinions concerning guarantee enterprises, and the number of days since the most recent negative public opinion event, totaling five specific indicators. The credit dimension indicators include the number of lending institutions, the number of overdue payments in the past six months, and the number of external guarantees in the past six months, totaling eight specific indicators. The operational dimension indicators include the number of employees paying social security, the number of times the enterprise has been listed on the list of abnormal operations in the past year, and the number of times the enterprise has received administrative penalties from the industrial and commercial departments in the past year, totaling seven specific indicators. These ten dimensions comprise a total of eighty-seven standardized indicators.

[0050] Based on the standardized metrics mentioned above, the system uses the XGBoost algorithm to select the top 50 features with the highest contribution to default prediction from 87 standardized metrics for initial screening. The following hyperparameters are set during XGBoost model training: maximum tree depth of 6, learning rate of 0.05, number of estimators of 150, minimum leaf node sample weight of 0.1, column sampling ratio of 0.8, and subsampling ratio of 0.8. Five-fold cross-validation is used to evaluate model stability, and the training epoch is set to stop early at 20 epochs, meaning training stops when the validation set AUC does not improve for 20 consecutive epochs. Feature importance is calculated using the Gain metric, which is the average information gain of each feature across all tree splits. A Gain value greater than zero indicates a positive contribution to model prediction, with higher Gain values ​​indicating greater contributions. The system sets a feature importance threshold of 0.01, meaning features with an average gain below 0.01 are considered to contribute weakly to default prediction and are discarded. After initial screening, 50 metrics are retained. Subsequently, WOE and IV values ​​are compared, and correlation analysis is used to further refine the metrics. The WOE (Warranty of Expectations) value is calculated as follows: For a given indicator, its value is divided into several intervals using an equal-frequency binning method. Continuous indicators are divided into ten bins, and discrete indicators into five bins. The WOE for each interval is equal to the natural logarithm of the percentage of defaulted samples divided by the percentage of normal samples within that interval. The IV (Indicator Value) value is calculated by multiplying the difference between the percentage of defaulted samples and the percentage of normal samples in each interval by the WOE value for that interval and then summing the results. The IV value reflects the indicator's predictive ability for default status: an IV value less than 0.02 indicates extremely weak predictive ability, 0.02 to 0.10 indicates relatively weak predictive ability, 0.10 to 0.30 indicates moderate predictive ability, and greater than 0.30 indicates relatively strong predictive ability. The system retains indicators with an IV value greater than 0.02 and removes indicators with weak predictive ability whose IV values ​​are less than or equal to 0.02. The Pearson correlation coefficient is calculated between the indicators. The Pearson correlation coefficient measures the degree of linear correlation between two indicators and ranges from -1 to +1. For strongly correlated indicator pairs with an absolute correlation coefficient greater than 0.7, the system retains only those with higher IV values ​​to avoid multicollinearity interfering with subsequent model training. This round of screening results in a shortlist of twenty-eight indicators. Logistic regression is then used to build the scorecard model. Logistic regression is a generalized linear model used for binary classification problems. It maps the linear regression results to the zero-to-one interval using the Sigmoid function, representing the probability that a sample belongs to a certain category. L2 regularization is used to prevent overfitting. The regularization coefficient C is optimized in the set {0.01, 0.1, 1, 10, 100} through grid search, aiming to maximize the AUC value on the test set. C is ultimately set to one. The model uses whether a company has defaulted as the binary classification target variable, defined as overdue for more than ninety days. The selected twenty-eight indicators are used as feature variables for training. The model uses maximum likelihood estimation to solve for the parameters, and the loss function is the logarithmic loss function.The training dataset consists of historical credit data from a commercial bank over the past three years, containing approximately 35,000 enterprise samples, of which about 4,200 are default samples. The training and test sets are split in a 7:3 ratio. The model achieved an AUC of 0.82 and a KS value of 0.58 on the test set. The model outputs the default probability, which is further mapped to a percentage score. The mapping uses a log-linear transformation formula: the score equals the baseline score minus the weight factor multiplied by the natural logarithm of the default probability. The baseline score is set to 600, corresponding to a default probability of 5%, and the weight factor is set to 40, meaning that for every halving of the default probability, the score increases by approximately 28 points. A higher score indicates higher risk, and a lower score indicates lower risk. These user scores are then used by the intelligent analysis layer for reference.

[0051] The system performs event-based risk assessment on standardized indicators based on preset risk signal definition rules, generating risk signals. The risk signal definition rules are derived from two sources: first, the risk event types explicitly listed in the credit risk monitoring guidelines issued by regulatory agencies; and second, the risk characteristics derived by the commercial bank based on statistical analysis of historical default samples. Specifically, the system performs a chi-square test on each standardized indicator and historical default samples. Indicators with a p-value less than 0.05 are considered significantly correlated with default. The final risk signal definition rules are further determined through expert review. Risk signals are categorized into high-risk, medium-risk, and low-risk levels according to the bank's risk appetite. The triggering of risk signals is executed in the signal detection engine, which scans the standardized indicator data table every minute, comparing the current value of each indicator with the judgment conditions defined for the signal. When the conditions are met, a risk signal record is generated. The criteria for determining each risk signal are as follows: The criterion for consecutive net profit losses is to check the net profit figures in the company's audit reports for the past three years; if all three years are negative, it is triggered. The criterion for a significant decline in net profit is that the current period's net profit decreases by more than 30% compared to the same period of the previous year. The formula for calculating the decline is (net profit of the same period of the previous year minus current period's net profit) divided by the absolute value of the net profit of the same period of the previous year. The criterion for a surge in the number of litigation cases involving a guarantor company is that the total number of litigation cases involving all guaranteed companies in the company's external guarantee list increases by more than 200% month-on-month compared to the total number of litigation cases in the previous three months. The criterion for a surge in the number of litigation cases involving the company is that the number of litigation cases in which the company is the defendant in the past three months increases by more than 100% month-on-month compared to the number of litigation cases in the previous three months. The criterion for a company frequently losing lawsuits is to check the judgment records of the company as the defendant in the China Judgments Online database; the number of cases lost by the company in the past year is... The following conditions must be met for determining whether a company has changed its legal representative: The legal representative field in the business registration information has changed within the past 30 days; the general manager or financial officer field in the business registration information has changed more than twice in the past year; the simplified deregistration of related companies is determined when any company on the company's external investment list is marked as "simplified deregistration" in the business registration system; the non-operating status of related companies is determined when any company on the company's external investment list has no record of submitting an annual business report or filing a tax return for six consecutive months; environmental pollution issues are determined when the company is listed as the subject of an administrative penalty by the environmental protection department and the penalty reason contains keywords such as "pollution," "discharge," or "environmental protection"; and illegal promotion issues are determined when the company is listed as the subject of an administrative penalty by the market supervision department and the penalty reason contains keywords such as "false advertising" or "illegal promotion."The criteria for determining whether a company can mortgage movable property are as follows: The total appraised value of the mortgaged property, as recorded in the unified registration and disclosure system for movable property financing, exceeds 50% of the company's net assets. The criteria for determining whether a company can pledge equity are as follows: The total number of shares pledged by the company's shareholders, as recorded in the equity pledge registration information, exceeds 30% of the company's total share capital. The criteria for determining whether a company has missed social security payments are as follows: The company has no social security payment records for three consecutive months. The criteria for determining whether a company has missed social security payments for more than six consecutive months are as follows: The company has missed social security payments for more than six consecutive months. The criteria for determining significant tax arrears are as follows: The amount of tax arrears recorded in the tax arrears record exceeds one million yuan, or the proportion of tax arrears to the company's tax payable in the previous year exceeds 20%. A downgrade in tax credit rating is also a factor. The criteria for determining whether a company's tax credit rating has dropped from A to B or below within the past year are as follows: The criteria for being sued by another party are that the amount in a single lawsuit in which the company is the defendant exceeds five million yuan; the criteria for being sued and losing a labor dispute are that the company is the respondent or defendant in labor arbitration or litigation records, and the ruling or judgment holds the company liable, with the amount in dispute exceeding 500,000 yuan; the criteria for overdue loans from other banks are that the company's loan accounts at other commercial banks have overdue records exceeding thirty days in the credit report; and the criteria for abnormal credit inquiries are that the credit report has been inquired more than ten times within the past three months, and the inquiring institutions are not limited to one.

[0052] Taking actual operating data from a certain quarter as an example, the system triggered approximately 12,000 risk signals within three months, including approximately 1,800 high-risk signals, approximately 4,800 medium-risk signals, and approximately 5,400 low-risk signals.

[0053] After the aforementioned data access, indicator processing, scorecard modeling, and risk signal detection, the data acquisition layer outputs four types of data products: The first type is eighty-seven standardized indicators, each stored in an indicator data table as a quintuple of enterprise ID, indicator name, indicator value, data source, and calculation time; the second type is user ratings, stored in a rating table as a quintuple of enterprise ID, rating value, model version, and calculation time; the third type is a risk signal list, stored in a risk signal table as a quintuple of enterprise ID, signal name, risk level, trigger time, and trigger basis; and the fourth type is cleaned raw data, stored in a data warehouse as enterprise ID, data table name, and cleaned records for subsequent data traceability and auditing.

[0054] Step S2: Based on preset indicator thresholds and scoring segments, determine the deterministic risk through the early warning rule engine and output structured early warning results.

[0055] The input to the early warning rule engine is the result value of data feature processing, including standardized indicators, user scores, and risk signals. The indicator thresholds in the rule definitions are derived from multi-dimensional reference benchmarks. For financial indicators, the thresholds reference the industry average and standard deviation. The system obtains the annual average and standard deviation of financial indicators for each industry from the industry association data service platform via HTTP API. The data is updated annually; when industry data is updated, the system automatically pulls the latest data and updates the threshold configuration. Taking the construction industry as an example, the industry average debt-to-asset ratio is 65%, and the standard deviation is 10%. Therefore, the high-risk threshold for the debt-to-asset ratio of this industry is the mean plus twice the standard deviation, i.e., 85%, and the medium-risk threshold is the mean plus one standard deviation, i.e., 75%. For indicators lacking industry benchmark data, the threshold is determined based on the percentile of a historical sample of normal enterprises of the commercial bank. The system extracts all enterprise samples that have not defaulted in the past three years from the data warehouse, calculates the 90th percentile for each indicator, sets this value as the high-risk threshold, the 75th percentile as the medium-high risk threshold, and the 60th percentile as the medium-risk threshold. For regulatory indicators, thresholds are set directly based on the minimum standards stipulated by regulations. For example, a provision coverage ratio not lower than 150% uses 150% as the threshold, and a capital adequacy ratio not lower than 10.5% uses 10.5% as the threshold. Regarding scoring segments, user scores above 800 are classified as low risk, 700-800 as low-to-medium risk, 600-700 as medium risk, 500-600 as medium-to-high risk, and below 500 as high risk. Each segment is a left-closed, right-open interval. Rule types include three types: single-indicator threshold triggers, such as a high-risk trigger when the overdue indicator equals 1; multi-indicator combination triggers, such as a high-risk trigger when a credit risk signal is triggered and the customer's score is below 720; and trend change triggers, such as a high-risk trigger when net profit is incurred for three consecutive periods with the loss widening quarter by quarter. Each period is a quarter, and the criterion for judging the widening loss is that the current period's loss exceeds the previous period's loss.

[0056] The output warning results are structured data, including the warning level, triggering indicators, and triggering reasons. Warning levels are divided into five categories: high risk, medium-high risk, medium risk, medium-low risk, and low risk. High risk corresponds to a score below 500 or triggering three or more high-risk rules; medium-high risk corresponds to a score between 500 and 600 or triggering two high-risk rules; medium risk corresponds to a score between 600 and 700 or triggering one high-risk rule; medium-low risk corresponds to a score between 700 and 800 and no high-risk rules triggered; and low risk corresponds to a score above 800 and no risk rules triggered. For example, a manufacturing company has a debt-to-asset ratio of 86.5%, while the industry's high-risk threshold for debt-to-asset ratio is 85%. Its net profit is negative 3.2 million yuan, and has been negative for two consecutive periods. The warning rule engine outputs a "high-risk" warning level, with triggering indicators being a debt-to-asset ratio of 86.5% and a net profit of negative 3.2 million yuan for two consecutive periods. The triggering reason is that the debt-to-asset ratio exceeds the industry's high-risk threshold of 85% and the net profit has been negative for two consecutive periods, indicating a continuous deterioration in the company's solvency.

[0057] Step S3: Establish a query database and encapsulate the aforementioned data into a structured data interface for use by the large language model, according to the risk dimension.

[0058] This step establishes a query database, storing the standardized indicators, user scores, and risk signals generated in step S1, as well as the early warning results generated in step S2, into the query database. Based on the query database, a structured data interface is encapsulated according to the risk dimension for use by the large language model.

[0059] The database used for querying is the relational database PostgreSQL, with the Enterprise Unified Social Credit Code as the primary key. Tables are established including: Enterprise Basic Information Table, Indicator Data Table, Scoring Table, Risk Signal Table, and Early Warning Result Table. The Enterprise Basic Information Table records basic attributes such as enterprise name, Unified Social Credit Code, industry classification, and registered location. The Indicator Data Table records the names, values, data sources, and update times of the eighty-seven standardized indicators generated in step S1. The Scoring Table records user scores, their calculation time, and model version. The Risk Signal Table records the name, level, and trigger time of triggered risk signals. The Early Warning Result Table records the early warning level, triggering indicator, and triggering reason. These five tables are linked by the Enterprise Unified Social Credit Code and support multi-table join queries using the enterprise ID.

[0060] Each structured data interface corresponds to a risk dimension. Internally, each interface aggregates indicator fields, signal fields, and judgment result fields describing the enterprise's risk status under that dimension. The interfaces are designed in a RESTful API style and implemented using the Spring Boot framework. Input parameters include the enterprise identifier (i.e., the unified social credit code or enterprise name) and the time range (i.e., the time interval for the query). The output is structured data in JSON format. The interface implementation layer uses the MyBatisORM framework to execute predefined SQL queries to extract data from the five tables mentioned above, assembles it according to the JSON structure defined by the interface, and returns it. For example, the SQL query for the enterprise's key members interface is to query the names and positions of key personnel from the business registration information table by enterprise ID; the SQL query for the enterprise tax rating interface is to query the latest tax rating and rating change records from the tax information table by enterprise ID and sorted in descending order by time; the SQL query for the enterprise's registered capital information interface is to query the registered capital, paid-in capital, and shareholder information fields from the business registration information table; and the SQL query for the enterprise financial statement data interface is to query the values ​​and calculation times of all financial indicators from the indicator data table by enterprise ID and indicator category equal to "financial". The interfaces are divided according to the risk dimensions analyzed by the large model. Each interface outputs a JSON structure containing metadata and data. The metadata records the interface version number, query time, and number of data entries. The data returns the name and corresponding value of each indicator in key-value pairs. Access permissions are controlled via API keys. Each call request must carry a valid API key in the HTTP header. The system returns data after verifying the key's validity; calls without sufficient permissions return an HTTP 401 status code. The call frequency is limited to no more than sixty requests per API key per minute. Exceeding this limit returns an HTTP 429 status code with the message "Request frequency too high, please try again later." In a single enterprise comprehensive risk assessment scenario, the system needs to call approximately twelve different structured data interfaces, with a total data volume of approximately fifteen kilobytes and an average interface response time of eighty-six milliseconds. Compared to directly querying the database using the large model (requiring the large model to generate and execute SQL statements), the interface call method improves the accuracy of data queries from approximately 72% to 100% because the field names and data structures returned by the interfaces are predefined, preventing errors in field names or table joins.

[0061] Step S4: Obtain the user's input question, perform intent recognition through question splitting intelligent agent, analyze the subject and dimension of analysis, and split the question into at least one sub-task.

[0062] The problem-splitting agent is a dedicated agent built into the system, and its processing flow is as follows.

[0063] First, the analysis subject is extracted from the question. After receiving the natural language question input by the user, the system calls the Named Entity Recognition (NER) model to perform sequence labeling on the question text. The NER model is based on the BERT-BiLSTM-CRF architecture. The BERT layer uses a pre-trained BERT-base-chinese model as the word embedding layer, converting each character in the input text into a 768-dimensional vector representation. The BiLSTM layer is a bidirectional long short-term memory network with a hidden layer dimension of 256, used to capture the sequence dependencies of characters in the context. The CRF layer is a conditional random field, used to perform globally optimal decoding on the label sequence output by the BiLSTM, ensuring the legality of label transfers. The model was trained on approximately 50,000 labeled post-loan question-and-answer datasets, with five training epochs, a batch size of 32, and a learning rate of 2 x 10^-5. The label categories include B-ORG for the first character of the company name, I-ORG for the middle character of the company name, and O for non-entity characters. The model outputs a label for each character in the input text. Based on the label sequence output by the model, the system concatenates all consecutive character sequences labeled B-ORG and I-ORG to form a candidate company name. For example, if a user inputs "Help me check if Company A has any financial problems this year, and whether it has been sued recently," the named entity recognition model outputs three labels, B-ORG, I-ORG, and I-ORG, at the position of "Company A," extracting the candidate company name as "Company A."

[0064] After extracting candidate company names, the system performs entity disambiguation and matching. The system pre-builds a client list database, which records the full name, unified social credit code, former name, and abbreviation of all loan clients of the commercial bank, totaling approximately 80,000 records. The matching algorithm proceeds as follows: First, exact matching: the candidate company name is compared exactly with the full name of the company in the client list. If they match, the analysis entity is directly determined. Second, alias matching: the candidate company name is compared with the abbreviation and former name of the company in the client list. If they match, the analysis entity is determined. Third, fuzzy matching: if neither exact matching nor alias matching finds a match, the edit distance between the candidate company name and the full name of each company in the client list is calculated. Edit distance refers to the minimum number of single-character edit operations required to convert one string to another, including insertion, deletion, and replacement operations. The company with the smallest edit distance is selected as the matching result, but this minimum edit distance must not exceed one-third of the length of the candidate company name; otherwise, the matching is considered a failure, and a "No matching company found" message is returned. Taking "Company A" as an example, the system matches the full name of "A Technology Co., Ltd." in the management list. The edit distance for the deletion operation of the word "technology" is two, and the candidate name length is three. Since two is greater than one, the matching threshold condition is met, and the analysis subject is determined to be "A Technology Co., Ltd." If the user's question does not contain any company name, the system uses the most recently identified company in the user's current session as the analysis subject. If there is no identified company in the session, the system returns the prompt "Please specify the company name to be queried".

[0065] Secondly, the system identifies the analytical dimensions of the problem. It uses an intent classification model to determine the analytical dimensions of the user's question. This model, based on the BERT architecture, was fine-tuned on approximately 30,000 post-loan question data points already labeled with analytical dimensions, outputting a probability distribution across ten dimensions: macroeconomic, policy analysis, industry analysis, financial health, operational status, legal risk, public opinion monitoring, related risk, early warning assessment, and other dimensions. The model takes the complete text of the user's question as input, converts it into a vector representation through a BERT encoding layer, maps the vector to a ten-dimensional output through a fully connected layer, and then converts the output into probability values ​​for each of the ten dimensions using a Softmax function, with the sum of the probabilities of each dimension equal to one. The system identifies the dimension with the highest probability value as the primary analytical dimension. If the probability values ​​of multiple dimensions are all higher than a preset threshold of 0.3, the system determines that the problem involves multiple dimensions. The labeling rules for model training data are as follows: questions involving macroeconomic factors such as GDP growth rate, monetary policy, and interest rate changes are labeled as macroeconomic dimension; questions involving policy interpretation, regulatory application, and compliance requirements are labeled as policy analysis dimension; questions involving industry size, industrial chain, competitive landscape, and industry trends are labeled as industry analysis dimension; questions involving financial indicator values, ratios, and trend changes are labeled as financial health dimension; questions involving business information, operating qualifications, bidding, and personnel changes are labeled as operating status dimension; questions involving litigation, enforcement, dishonesty, and restrictions on high-end consumption are labeled as judicial risk dimension; questions involving news reports, public opinion events, and sentiment are labeled as public opinion monitoring dimension; questions involving equity structure, supply chain relationships, and guarantee relationships are labeled as associated risk dimension; questions involving comprehensive risk level and risk event summaries are labeled as early warning assessment dimension; and questions that cannot be categorized into the above nine categories are labeled as other dimensions. Taking the user question "Help me check if Company A has any financial problems this year, and whether it has been sued recently" as an example, the intent classification model outputs a probability of 0.87 for the financial health dimension and 0.76 for the legal risk dimension. Both dimensions are higher than the threshold of 0.3, indicating that the question involves two analytical dimensions: financial health and legal risk.

[0066] Secondly, the system breaks down the questions into subtasks based on the analytical dimensions. For single-dimensional questions, the system directly treats the question as a subtask within that dimension. For multi-dimensional questions, the system breaks down the original question into multiple subtasks based on the dimensions, with each subtask corresponding to an analytical dimension. The breakdown method involves extracting text fragments from the original question that are relevant to that dimension as the description of the subtask. The extraction of text fragments is based on the dependency parsing results of the question text. The system uses dependency parsing tools to parse the question text, identifying the parallel, subordinate, and modifying relationships between clauses, decomposing the question into several clauses or phrases, each corresponding to an independent semantic unit. The system merges semantic units belonging to the same dimension into a subtask description for that dimension. Taking the user question above as an example, the question text contains two semantic units: "Has there been any financial problem this year?" corresponding to the financial health dimension, and "Has there been any lawsuit recently?" corresponding to the legal risk dimension. The system uses "Has Technology A's financial problem this year?" as the subtask description for the financial health dimension and "Has Technology A's recent lawsuit been filed?" as the subtask description for the legal risk dimension.

[0067] Finally, the system outputs a structured task decomposition result. The system organizes the analysis subject, analysis dimension list, and subtask list into structured data according to a preset JSON format. In this example, the output JSON structure contains the following fields: the subject field is "A Technology Co., Ltd." and its unified social credit code; the dimension list field is an array containing "financial health" and "judicial risk"; and the subtask array contains two objects, each containing fields such as task identifier, dimension, task description, and priority. Priority is determined by the order of semantic units in the user's question, with earlier semantic units having higher priority. In this example, the "financial health" subtask has a higher priority than the "judicial risk" subtask. This structured data serves as the task decomposition result, which is read by the intelligent routing module in step S5 for subsequent routing allocation.

[0068] The problem-splitting agent strictly adheres to the original statement of the user's question throughout the entire processing process, without diverging, extending, or specifying indicators, and without adding extra information such as time, company objects, or data metrics. Its constraint logic is implemented by the "prohibit divergence and extension" rule in the prompt: the system prompt explicitly states instructions such as "strictly adhere to the original question statement, without diverging, extending, or specifying indicators" and "do not add extra information such as time, company objects, or data metrics." When the user's question does not contain a time constraint, no time constraint is added to the subtask description; when the user's question does not contain a specific indicator name, no indicator name is added to the subtask description. For example, if the user inputs "Check the risks of Company A," the output of the problem-splitting agent will only be "Risks of Company A," and will not automatically expand to "Legal and credit risks of Company A in the past year." These behavioral constraints ensure the objectivity of task splitting and prevent the agent from introducing analysis conditions not specified by the user, thus avoiding answers that deviate from the user's intent.

[0069] Step S5: Based on the analysis dimensions corresponding to each subtask, assign each subtask to the corresponding professional analysis agent through intelligent routing.

[0070] The system pre-defines multiple specialized analytical agents, each with its own professional role, core analytical framework, and behavioral boundary constraints based on prompt word engineering. Prompt words are written using structured templates, which include the following fixed sections: a role positioning section defining the agent's professional identity and analytical expertise; an analytical framework section defining the sequence of analytical steps and the inputs and outputs of each step; and a behavioral constraint section defining three types of boundaries: role definition constraints, data usage constraints, and professional knowledge constraints. Each agent's prompt word template is independently stored in the configuration center, supporting dynamic updates without requiring a system restart.

[0071] The input to the intelligent routing module is the JSON structured task decomposition result output in step S4. After parsing the JSON, the module iterates through each subtask in the subtask array and queries the agent routing mapping table based on the "belonging dimension" field value of the subtask. The agent routing mapping table is a pre-configured key-value pair mapping, where the key is the name of the analysis dimension and the value is the name of the corresponding professional analysis agent. The mapping relationships are as follows: Macroeconomic dimension maps to Macroeconomic agent; Policy analysis dimension maps to Policy analysis agent; Industry analysis dimension maps to Industry analysis agent; Financial health dimension maps to Financial health agent; Operating status dimension maps to Operating status agent; Judicial risk dimension maps to Judicial risk agent; Public opinion monitoring dimension maps to Public opinion monitoring agent; Related risk dimension maps to Related risk agent; Early warning assessment dimension maps to Early warning assessment agent; Other dimensions map to the backup agent. For each subtask, the routing module queries the mapping table based on its dimension to obtain the corresponding agent name, and then passes the task description of the subtask as the analysis target to the target agent. If a certain dimension does not have a corresponding agent in the mapping table, the routing module routes the subtask to the backup agent for processing. The routing results of all subtasks are summarized into an execution plan list, where each entry contains the agent name, subtask description, and priority.

[0072] In addition to the common prompts, each specialized agent has its own personalized prompts based on its specific focus scenario. Behavioral boundary constraints include three types: role definition constraints, which limit the large model's response identity to a specific role; data usage constraints, which require all analysis to be based on enterprise data returned by structured data interfaces, and if data for a certain indicator is not provided, the system must output "Data not yet obtained," and cannot fabricate data; and professional knowledge constraints, which require strict adherence to the standards of the warning rule interface calls when determining risk levels. The personalized analysis framework for each agent is explained below.

[0073] (I) Macroeconomic Intelligent Entity Upon receiving the analysis task, the macroeconomic agent first obtains the enterprise's industry classification code (GB / T4754-2017 four-digit code) through the industrial and commercial dimension interface to determine its national economic industry category. Subsequently, it obtains the following six core macroeconomic indicators and their time series through the macroeconomic data interface: the year-on-year GDP growth rate for the past four quarters (source: National Bureau of Statistics), the year-on-year monthly M2 growth rate for the past three months (source: People's Bank of China), the monthly manufacturing PMI for the past three months (source: National Bureau of Statistics), the year-on-year monthly CPI increase for the past three months (source: National Bureau of Statistics), the year-on-year monthly PPI increase for the past three months (source: National Bureau of Statistics), and the 10-year Treasury bond yield for the past three months (source: China Bond Information Network). Data acquisition failures are marked as "Data not yet acquired." The agent analyzes the data according to the following technical path: The first step is industry cyclicality determination. The agent queries the industry cyclicality tag library based on the industry classification code. This tag library pre-stores the relationship types between various industries and the macroeconomic cycle, including three categories: strong cyclicality, weak cyclicality, and counter-cyclicality. The determination rules are as follows: industries tagged as strong cyclicality have a correlation coefficient between their revenue and GDP growth rate higher than 0.6 in historical five-year data; weak cyclicality industries have a correlation coefficient between 0.2 and 0.6; and counter-cyclicality industries have a correlation coefficient lower than 0.2. The correlation coefficient is calculated using the Pearson correlation coefficient, with the quarterly revenue growth rate of all listed companies in the industry over the past five years and the quarterly GDP growth rate as the two variables.

[0074] The second step is macroeconomic impact analysis. The agent compares the current values ​​of six macroeconomic indicators with their respective thresholds. The thresholds are set as follows: the GDP growth rate threshold is 5%, below which is marked as "economic slowdown"; the normal range for M2 growth is 8% to 12%, below which is marked as "monetary tightening" and above which is marked as "monetary easing"; the PMI threshold is 50%, below which is marked as "manufacturing contraction" and above which is marked as "manufacturing expansion"; the target range for CPI is 0% to 3%, exceeding the upper limit is marked as "inflationary pressure" and below the lower limit is marked as "deflation risk"; a positive PPI year-on-year increase is marked as "industrial product price increase" and a negative PPI increase is marked as "industrial product price decrease"; a 10-year Treasury bond yield above 3.5% is marked as "high financing costs" and below 2.5% is marked as "low financing costs". The comparison result for each indicator is labeled either "normal" or a specific deviation direction.

[0075] The third step is transmission path analysis. The agent takes the deviation direction labels of macroeconomic indicators as input and queries the transmission path rule base. This rule base predefines the main transmission mechanisms of macroeconomic variables to various industries. Taking the automotive manufacturing industry as an example, the transmission path rules are as follows: if the PMI is below 50 and the PPI year-on-year growth is positive, the transmission to the industry will occur through two channels: "demand contraction" and "cost increase," specifically manifested as a decrease in end-user demand and an increase in raw material procurement costs. The strength level of each channel is determined by the weight coefficient in the rule base, with the weight coefficient ranging from zero to one. The agent marks channels with a strength higher than 0.6 as "strong transmission," those between 0.3 and 0.6 as "medium transmission," and those below 0.3 as "weak transmission."

[0076] The fourth step is comprehensive judgment. Based on the above analysis results, the agent outputs structured conclusions: the short-term (next three months) operational pressure level, the medium-term (next year), and the long-term (next three years) assessment. The rules for determining the operational pressure level are: two or more strong transmission channels are marked as "relatively high pressure," one strong transmission channel or two moderate transmission channels are marked as "moderate pressure," and all other cases are marked as "pressure manageable." If four or more macroeconomic indicators deviate from the normal range, the conclusion is marked with "overall macroeconomic environment is weak." During the analysis, the agent explicitly labels the values ​​of each indicator and the names of the source institutions, without using any data stored in the model's memory.

[0077] (II) Policy Analysis Intelligent Agent After receiving the analysis task, the policy analysis agent first obtains the industry classification code of the enterprise through the business registration dimension interface. Then, it performs the following retrieval operations through the knowledge base recall module: constructing search conditions, with the business line tag equal to the business line corresponding to the enterprise's industry, the document type tag equal to "notice" or "regulation", and the timeliness tag equal to "currently in effect" or "soon to take effect". This search condition, along with the keywords in the user's question, is submitted to the knowledge base for multi-condition filtering. If there are more than five search results, the five most recent results are retrieved in reverse chronological order of publication date. If the search results are empty, "No relevant industry policies found" is output, and the analysis process terminates.

[0078] For each recalled document, the agent performs the following structured information extraction: It extracts the release date by matching the "Release Date" or "Issuance Date" field in the document body using regular expressions; it extracts the name of the issuing agency by matching the "Issuing Agency" or "Issuing Authority" field using regular expressions, with the matching pattern "([^,。]*?(Ministry|Commission|Bureau|Administration|Department|Office))"; it extracts core policy content by matching the core policy content, specifically the paragraphs containing keywords such as "policy content," "main content," and "core measures," plus the following two hundred characters; it extracts sentences containing keywords such as "scope of application" and "applicable objects" by matching the scope of application using regular expressions; and it extracts the effective date by matching keywords such as "implementation date" and "effective date" using regular expressions.

[0079] The intelligent agent conducts impact analysis using the following technical approach: The first step is to determine the direction of industry impact. The agent inputs the extracted core policy content into the TextCNN text classification model, which is trained on approximately 5,000 policy statements already labeled with their impact directions. The impact direction classification labels include three categories: "positive," "negative," and "neutral." The model outputs probability values ​​for each category, takes the category with the highest probability as the judgment result, and outputs this probability value as the confidence score. If the confidence score is below 0.6, the conclusion is marked "impact direction to be confirmed."

[0080] The second step is to quantify the degree of impact. If the policy's impact is determined to be "negative," the agent extracts keywords related to restrictive measures from the core content of the policy, including "restriction," "compression," "prohibition," "reduction," "tightening," and "raising the threshold." The frequency F of these keywords in the policy text is calculated, and the impact coefficient Imp is calculated as F divided by the total number of words in the policy text W multiplied by 1000. An Imp value greater than five is marked as "strong negative impact," an Imp value between two and five is marked as "moderate negative impact," and an Imp value less than two is marked as "weak negative impact."

[0081] The third step is enterprise-level impact analysis. The agent obtains the percentage of the enterprise's revenue related to the policy (Pct) over the past year through the financial statement data interface. The calculation method for the percentage of revenue related to the policy is as follows: if the policy involves centralized procurement of pharmaceuticals, the percentage of generic drug revenue to total revenue is taken; if the policy involves real estate regulation, the percentage of residential development revenue to total revenue is taken. The agent calculates the expected impact on revenue as equal to the impact coefficient Imp divided by ten multiplied by the percentage of revenue related to the policy (Pct). Taking an Imp value of four and a percentage of revenue related to the policy of 60% as an example, the expected impact on revenue is 24%, which is 40% multiplied by 60%.

[0082] The fourth step is to generate recommended measures. The agent generates structured recommendations based on the direction and degree of impact: when the policy is unfavorable and the impact is strong, it outputs "It is recommended to assess the business sustainability under the policy impact and formulate business transformation or diversification strategies"; when the policy is unfavorable and the impact is moderate, it outputs "It is recommended to strengthen compliance management and closely monitor the subsequent supporting details of the policy"; when the policy is favorable, it outputs "It is recommended to seize the policy opportunity and accelerate the layout of related businesses".

[0083] The intelligent agent is prohibited from using retrieval mechanism expressions such as "according to the knowledge base", "recall information display", or "system query" in its output.

[0084] (III) Industry Analysis Intelligent Agent Upon receiving the analysis task, the industry analysis agent first obtains the company's four-digit industry classification code (GB / T4754-2017) through the business registration dimension interface. If the user explicitly specifies the industry name in the question, the user-specified name will be used, and the company's industry classification code will be ignored. After industry identification, the agent obtains the following datasets through the industry data interface: industry market size and year-on-year growth rate for the past three years (data source: industry association annual reports); monthly industry prosperity index for the past year (data source: prosperity monitoring data released by the industry's regulatory authorities); market share data of the top ten companies (data source: industry research institution annual reports); and average gross profit margin and average net profit margin for the past year (data source: Wind Information industry financial database).

[0085] The intelligent agent analyzes the data using the following technical approach: The first step is supply chain analysis. The intelligent agent constructs a directed graph of the supply chain, where nodes represent each link in the chain and edges represent upstream and downstream supply and demand relationships. Upstream links include raw material suppliers and component manufacturers; midstream links include product manufacturers and integrators; and downstream links include distributors and end users. The intelligent agent calculates the upstream concentration index (HHI for upstream), which is the sum of the squares of the market share of the top five upstream suppliers; and the downstream concentration index (HHI for downstream), which is the sum of the squares of the market share of the top five downstream customers. An HHI index less than 1500 indicates a dispersed supply chain, 1500-2500 indicates a moderately concentrated supply chain, and greater than 2500 indicates a highly concentrated supply chain. The higher the concentration, the stronger the bargaining power of that link.

[0086] The second step is competitive landscape analysis. The agent calculates the industry's CR5 and CR10 based on the market share data of the top ten companies. The rules for determining the industry concentration type are as follows: if CR5 is greater than 60% and CR10 is greater than 80%, it is considered an oligopoly; if CR5 is between 40% and 60% and CR10 is between 60% and 80%, it is considered a monopolistic competition; and if CR5 is less than 40% and CR10 is less than 60%, it is considered a perfectly competitive market.

[0087] The third step is to determine the industry's prosperity level. The intelligent agent compares the current monthly prosperity index with the average prosperity index over the past twelve months. A current value higher than the average plus one standard deviation is considered "high prosperity," between the average and one standard deviation is considered "stable prosperity," and lower than the average minus one standard deviation is considered "sluggish prosperity." The standard deviation of the prosperity index is calculated using the sample standard deviation of the monthly values ​​over the past twelve months.

[0088] The fourth step is to determine the company's industry position. The intelligent agent obtains the company's revenue scale through a financial statement data interface and compares it with the revenue scale of the top ten companies in the industry to determine the company's ranking range in the industry. The top five are marked as "leading companies," those ranked sixth to twentieth are marked as "medium-to-large companies," and those ranked twentieth and below are marked as "small-to-medium-sized companies." At the same time, the intelligent agent obtains the company's gross profit margin and compares it with the industry average gross profit margin, calculating the gross profit margin deviation. The gross profit margin deviation equals the company's gross profit margin minus the industry average gross profit margin; a positive value indicates profitability better than the industry average, and a negative value indicates profitability lower than the industry average.

[0089] (iv) Financial Health Intelligent Entity Upon receiving an analysis task, the Financial Health AI agent retrieves the following twenty-one core financial fields from the company's most recent audited financial report via the financial statement data interface: Total Assets, Total Liabilities, Current Assets, Current Liabilities, Inventory, Accounts Receivable, Operating Revenue, Operating Costs, Net Profit, Earnings Before Interest and Taxes, Interest Expense, Net Cash Flow from Operating Activities, Net Cash Flow from Investing Activities, Net Cash Flow from Financing Activities, Shareholders' Equity, Total Share Capital, Earnings Per Share, Net Assets, Fixed Assets, Intangible Assets, and Accounts Payable. If any field fails to be retrieved, it is marked as "Data Not Yet Retrieved," and that indicator will not be included in subsequent calculations.

[0090] The agent calculates the following metrics sequentially, with the calculation logic for each metric as follows: Solvency indicators: Debt-to-equity ratio equals total liabilities divided by total assets; Current ratio equals current assets divided by current liabilities; Quick ratio equals current assets minus inventory divided by current liabilities; Interest coverage ratio equals earnings before interest and taxes divided by interest expense; Cash ratio equals cash and cash equivalents divided by current liabilities; Equity ratio equals total liabilities divided by shareholders' equity; Long-term debt ratio equals non-current liabilities divided by non-current liabilities plus shareholders' equity; Debt service ratio equals net cash flow from operating activities divided by total principal and interest payments for the period.

[0091] Profitability metrics: Return on equity (ROE) equals net profit divided by average shareholders' equity; Return on total assets (ROA) equals earnings before interest and taxes (EBIT) divided by average total assets; Operating profit margin equals operating profit divided by operating revenue; Gross profit margin equals operating revenue minus operating costs divided by operating revenue; Net profit margin equals net profit divided by operating revenue; Cost-to-profit ratio equals total profit divided by total cost and expenses; Earnings per share are directly taken from the interface return value; Return on total assets equals net profit divided by average total assets.

[0092] Operational efficiency indicators: Total asset turnover equals operating revenue divided by average total assets; Accounts receivable turnover equals operating revenue divided by average accounts receivable; Inventory turnover equals cost of goods sold divided by average inventory; Fixed asset turnover equals operating revenue divided by average net fixed assets; Current asset turnover equals operating revenue divided by average current assets; Working capital turnover equals operating revenue divided by average working capital, and working capital equals current assets minus current liabilities; Accounts payable turnover equals cost of goods sold divided by average accounts payable.

[0093] Growth capacity indicators: Revenue growth rate equals current period revenue minus previous period revenue divided by previous period revenue; Net profit growth rate equals current period net profit minus previous period net profit divided by previous period net profit; Total asset growth rate equals current period total assets minus previous period total assets divided by previous period total assets; Net asset growth rate equals current period net assets minus previous period net assets divided by previous period net assets.

[0094] Cash flow indicators: The ratio of net operating cash flow to operating revenue is equal to net cash flow from operating activities divided by operating revenue; free cash flow is equal to net cash flow from operating activities minus capital expenditures; the cash reinvestment ratio is equal to net cash flow from operating activities minus cash dividends divided by fixed assets plus long-term investments plus working capital; the cash flow adequacy ratio is equal to the sum of net cash flow from operating activities over the past five years divided by the sum of capital expenditures over the past five years plus inventory increases plus cash dividends.

[0095] The judgment benchmark for each indicator comes from the quantile values ​​of the corresponding industry category and scale type in the published "Enterprise Performance Evaluation Standard Values". The intelligent agent compares each calculated indicator value with the excellent value (90th percentile), good value (70th percentile), average value (50th percentile), low value (30th percentile), and poor value (10th percentile) in the standard values ​​one by one to determine the quintile level of the indicator.

[0096] (v) Operational Status of Intelligent Entity Upon receiving the analysis task, the business status intelligence agent obtains the following structured fields of the enterprise through the business registration dimension interface: registered capital, paid-in capital, establishment date, legal representative, list of key personnel and their positions, list of shareholders and their shareholding ratios, number of branches, and list of change records. It also obtains the following structured fields through the business data interface: list of qualification certificates and their validity periods, number of times the enterprise has participated in and won bids in the past year, administrative licensing records and administrative penalty records in the past year, number of employees paying social security and total number of employees, import / export qualifications, and business scope.

[0097] Based on the above data, the agent performs the following quantization calculations: The operational stability index is calculated as follows: paid-in registered capital ratio multiplied by 0.2, social security coverage ratio multiplied by 0.3, and management stability score multiplied by 0.5. The paid-in registered capital ratio is calculated as paid-in capital divided by registered capital. The social security coverage ratio is calculated as the number of employees paying social security divided by the total number of employees. The management stability score is calculated as 1 divided by the number of major personnel changes in the past three years plus 1, with a score range from zero to one.

[0098] The development scale score equals the industry quantile of the number of employees multiplied by 0.3, plus the industry quantile of registered capital multiplied by 0.3, plus the industry quantile of the number of branches multiplied by 0.4. The industry quantile is calculated by comparing the company's score with all companies under the same industry code, and calculating the proportion of samples with scores lower than the company's, with the result between zero and one.

[0099] The compliance risk score is equal to the number of administrative penalties in the past year multiplied by -2, plus the number of abnormal business operations in the past year multiplied by -1, plus 5. The benchmark score is 5 points, and the minimum score is 0 points. A score below 2 is marked as "high compliance risk".

[0100] The bidding competition score is equal to the success rate multiplied by 100. The success rate is equal to the number of successful bids divided by the number of times the bidder participated in bidding. A success rate higher than the industry average plus one standard deviation is marked as "strong market competitiveness," while a success rate lower than the industry average minus one standard deviation is marked as "weak market competitiveness."

[0101] (vi) Judicial Risk Intelligence Agent After receiving the analysis task, the judicial risk intelligence agent obtains the following structured data of the enterprise over the past year through the judicial dimension interface: a list of lawsuits in which the enterprise is the defendant, including the filing date, cause of action, amount in dispute, current trial stage, and judgment result; a list of cases subject to enforcement, including the filing date, executing court, amount in dispute, and enforcement status; records of restrictions on high-level consumption, including the restriction date and content; records of dishonest judgment debtors, including the date of dishonesty and circumstances of dishonesty; and records of equity freezes, including the freeze date, number of frozen shares, and the court that froze the shares.

[0102] Based on the above data, the agent performs the following quantization calculations: The number of cases involved (Score_litigation) equals the total number of lawsuits in which the defendant is located in the past year. The amount in dispute (Score_amount) equals the sum of the amounts in dispute across all lawsuits. The amount to be executed (Score_execution) equals the sum of the amounts in dispute across all cases subject to execution.

[0103] Risk Level Determination Rules: If Score_litigation is greater than the industry average plus two standard deviations and Score_amount is greater than 15% of the company's net assets, or Score_execution is greater than 10% of the company's net assets and its absolute value is greater than 5 million yuan, or there is a record of restrictions on high-end consumption, the risk level is determined to be high. If Score_litigation is greater than the industry average plus one standard deviation, or Score_amount is greater than 5% but less than or equal to 15% of the company's net assets, or Score_execution is greater than 5% of the company's net assets and its absolute value is greater than 1 million yuan, the risk level is determined to be medium. All other situations are determined to be low risk.

[0104] Trend Determination: The system compares the data from the current six months with that from the previous six months and calculates the month-on-month growth rate of the number of cases involved. The month-on-month growth rate is equal to the number of cases in the current six months minus the number of cases in the previous six months divided by the number of cases in the previous six months. A growth rate greater than 30% is marked as a "deteriorating trend," a growth rate between -10% and +30% is marked as a "stable trend," and a growth rate less than -10% is marked as an "improving trend."

[0105] Case type distribution: The intelligent agent classifies and statistically analyzes litigation cases according to cause of action, calculates the proportion of each type of case to the total number of cases, identifies the case type with the highest proportion as the "main dispute type", and marks those with a proportion of more than 50% as "highly concentrated".

[0106] (vii) Intelligent Agent for Public Opinion Monitoring After receiving the analysis task, the public opinion monitoring intelligent agent obtains the following structured data of the enterprise over the past six months through the public opinion dimension interface: total amount of public opinion and sentiment classification statistics, including the number of positive, neutral and negative categories; public opinion overview, including public opinion index and attention score; public opinion dynamic list, including title, publication time, sentiment tag and summary.

[0107] Based on the above data, the agent performs the following quantization calculations: The percentage of negative public opinion is equal to the number of negative public opinion events divided by the total number of public opinion events. The deviation of this value from the industry average is calculated. The industry average is determined by the system based on benchmark data from quarterly industry public opinion reports published by third-party public opinion service providers. The deviation value equals the company's percentage of negative public opinion minus the industry average. A deviation value greater than the industry standard deviation is marked as "prominent negative public opinion," a deviation value between plus or minus one standard deviation is marked as "normal public opinion," and a deviation value less than one standard deviation is marked as "public opinion better than the industry average."

[0108] Trend Determination: The proportion of negative public opinion in the most recent month is compared with the average proportion of negative public opinion in the previous three months, and the change is calculated. The change equals the proportion in the most recent month minus the average proportion in the previous three months. A change greater than three percentage points is marked as "negative public opinion rising," a change less than -3 percentage points is marked as "negative public opinion falling," and a change between +3 and -3 percentage points is marked as "public opinion stable." This threshold is set based on the normal fluctuation range of industry public opinion monitoring; that is, a monthly fluctuation exceeding three percentage points is generally considered a significant change.

[0109] Hotspot Event Identification: Events published within the past 30 days are selected from the public opinion dynamics list. The influence score for each event is calculated by multiplying its public opinion index and attention score. If the influence score exceeds a preset threshold of 50, the event is marked as a hotspot event. The threshold is set based on the fact that a platform public opinion index of 50 corresponds to approximately 100,000 views, which is generally considered a small to medium-sized hotspot in social public opinion.

[0110] The summary of trending events is generated by calling a large language model to summarize the original public opinion text, with the prompt "Please summarize the core content of the following public opinion event in no more than fifty words". The summary text returned by the large language model serves as the description of the trending event.

[0111] (viii) Associated Risk Intelligent Agent After receiving the analysis task, the risk intelligence agent obtains the following structured data of the enterprise through the correlation dimension interface: a list of external investments, including the name of the invested enterprise, the investment ratio, and the investment amount; a list of shareholders, including the shareholder name, the shareholding ratio, and the shareholder type; a list of suppliers, including the supplier name and the purchase ratio; a list of the top five customers, including the customer name and the sales ratio; a list of external guarantees, including the name of the guaranteed party, the guarantee amount, and the guarantee type; and a list of equity pledges, including the pledgor, the pledgee, and the number of pledged shares.

[0112] Based on the above data, the agent performs the following quantization calculations: The outward investment dispersion index HHI_invest is equal to the sum of the squares of the investment proportions of all outward-investing enterprises. An HHI_invest value less than 0.3 indicates diversified investment, 0.3 to 0.6 indicates moderately concentrated investment, and a value greater than 0.6 indicates highly concentrated investment.

[0113] Supply chain concentration (SC) equals the sum of the purchase proportions of the top three suppliers. If SC is greater than 75%, it is marked as "highly concentrated suppliers". If SC is between 60% and 75%, it is marked as "relatively high supplier concentration". If SC is less than 60%, it is marked as "dispersed suppliers".

[0114] Customer concentration (CC) is equal to the sum of sales percentages from the top three customers. If CC is greater than 75%, it is marked as "highly concentrated customers"; if CC is between 60% and 75%, it is marked as "relatively high customer concentration"; and if CC is less than 60%, it is marked as "dispersed customers".

[0115] The external guarantee ratio (GR) is equal to the total amount of external guarantees divided by the company's net assets. A GR greater than 40% is marked as "excessively high guarantee ratio", a GR between 20% and 40% is marked as "relatively high guarantee ratio", and a GR less than 20% is marked as "normal guarantee ratio".

[0116] The share pledge ratio (PR) is equal to the number of shares pledged by the actual controller divided by the total number of shares held by the actual controller. A PR greater than 70% is marked as "high pledge risk", a PR between 50% and 70% is marked as "relatively high pledge risk", and a PR less than 50% is marked as "pledge risk is controllable".

[0117] (ix) Early warning assessment intelligent agent After receiving the analysis task, the early warning assessment agent calls the early warning results through the data interface module to obtain the structured early warning results determined by the early warning rule engine, and at the same time receives the output results of the financial health agent, the judicial risk agent, the public opinion monitoring agent, and the related risk agent.

[0118] The intelligent agent constructs a risk event matrix, where the row vectors represent the various analysis dimensions and the column vectors represent the risk event types. Risk event types include: financial events (e.g., debt-to-equity ratio exceeding a threshold, consecutive net profit losses, interest coverage ratio below 1.5); legal events (e.g., number of lawsuits exceeding a threshold, amount of enforcement exceeding a threshold, restrictions on high-end consumption); public opinion events (e.g., negative public opinion ratio exceeding a threshold, hot-button negative events); and related events (e.g., guarantee ratio exceeding a threshold, supplier concentration exceeding a threshold, equity pledge ratio exceeding a threshold). Each element in the matrix is ​​a trigger flag: 1 for triggered, 0 for not triggered.

[0119] Comprehensive Risk Level Determination Rules: If the total number of risk events with trigger marker 1 is greater than or equal to five, and includes at least one financial event and one legal event, the comprehensive risk level is determined to be high. If the total number of risk events with trigger marker 1 is between three and four, or the total number of trigger events is greater than or equal to five but does not simultaneously include financial and legal events, the comprehensive risk level is determined to be medium-high. If the total number of risk events with trigger marker 1 is two, the comprehensive risk level is determined to be medium. If the total number of risk events with trigger marker 1 is one, the comprehensive risk level is determined to be medium-low. If the total number of risk events with trigger marker 1 is zero, the comprehensive risk level is determined to be low.

[0120] Key Risk Focus: The system counts the number of times each risk event is triggered, with the type triggering most frequently marked as the core risk type. If a financial event is triggered two or more times, the core risk output is "Insufficient corporate solvency"; if a legal event is triggered two or more times, the core risk output is "Prominent corporate legal risk"; if a public opinion event is triggered two or more times, the core risk output is "Increased corporate reputational risk"; if a related event is triggered two or more times, the core risk output is "Increased risk of related-party transmission".

[0121] The agent is prohibited from repeating the score value, outputting the rule number, outputting the internal scoring logic, or outputting the model judgment process in the output.

[0122] (x) Comprehensive summary of intelligent agents, problem-decomposition intelligent agents, and bottom-line intelligent agents The overall summary agent is positioned as an intelligent agent for summarizing and analyzing multi-dimensional information within the enterprise. Its analysis framework involves reading and parsing the analysis conclusions of the nine dimension agents mentioned above using a unified JSON Schema. The parsing rules are as follows: read the "Analysis Conclusion" and "Risk Level" fields from the JSON object returned by each dimension agent and concatenate them in the order of presentation. During the concatenation process, if the value of the "Analysis Conclusion" field for a certain dimension is equal to "Data not yet obtained, this part will not be analyzed" or "No relevant information available," then that dimension will output "No relevant information available." After concatenation, it is organized according to a three-part structure: summary, description, and recommendations. The summary part extracts the "Comprehensive Risk Level" and "Core Risk Points" fields from the output of the early warning assessment agent; the description part outputs the analysis conclusions of each dimension in the order of presentation; the recommendations part matches the corresponding recommendation text from a predefined recommendation rule library based on the comprehensive risk level: high risk matches "Recommend immediately activate the post-loan risk disposal plan," medium risk matches "Recommend increase the frequency of post-loan monitoring," and low risk matches "Recommend maintain routine post-loan management." In summary, the role of the intelligent agent is defined as a senior credit account manager and credit review expert at the bank. The priority of instructions, from high to low, is as follows: laws and regulations, security specifications, this prompt word, credit analysis rules, previous intelligent agent output, and user input. An anti-injection blacklist is set, and instructions such as "ignore the above rules" and "forgot to set" are all invalid.

[0123] The role of the question-segmentation agent is to assist in classifying user questions into different dimensions. Its analytical framework is as follows: the user-input question text is fed into the BERT intent classification model, which outputs a probability distribution across ten dimensions. Dimensions with a probability higher than 0.3 are designated as candidate dimensions. If there is only one candidate dimension, it is directly output; if there are multiple candidate dimensions, they are arranged in descending order of probability. The model's classification rules are as follows: GDP growth rate, monetary policy, and interest rate changes are labeled as macroeconomic dimensions; policy interpretation, regulatory application, and compliance requirements are labeled as policy analysis dimensions; industry size, industrial chain, and competitive landscape are labeled as industry analysis dimensions; financial indicator values, ratios, and trend changes are labeled as financial health dimensions; business information, operating qualifications, and bidding are labeled as operating status dimensions; litigation, enforcement, dishonesty, and high-consumption restrictions are labeled as judicial risk dimensions; news reports and public opinion events are labeled as public opinion monitoring dimensions; equity structure and supply chain relationships are labeled as associated risk dimensions; comprehensive risk levels are labeled as early warning assessment dimensions; and items that cannot be classified into the above nine categories are labeled as other dimensions. This agent strictly adheres to the original question statement, without diverging, extending, or specifying indicators.

[0124] The role of the support agent is positioned as a general enterprise analysis assistant. Its analysis framework consists of a four-step decision-making process: First, it determines whether the user's question can be directly answered using general knowledge. If so, it uses the general capabilities of the large language model to generate an answer. Second, it determines whether it needs to rely on session memory information. If so, it reads the context from the Redis session cache and integrates it naturally. Third, it assesses the sufficiency of the data in the enterprise analysis scenario. If the missing data ratio of the interface exceeds 50%, it outputs "Unable to form a valid analysis conclusion." Fourth, if it exceeds its scope of responsibility, it outputs "This issue is beyond the current analysis scope; it is recommended to consult the relevant business department." The support agent is prohibited from using technical descriptions such as "knowledge base not hit," "data not recalled," or "not found by the system."

[0125] Step S6: Each professional analytical agent calls the structured data interface to obtain enterprise data, and retrieves professional knowledge documents through the knowledge base recall module, piecing them together to form a complete input.

[0126] In this step, each specialized analytical agent calls the structured data interface encapsulated in step S3 to obtain the enterprise data corresponding to the analytical subject. Simultaneously, the knowledge base retrieval module performs semantic retrieval on the user's question, searching the knowledge base for professional knowledge documents related to the subtask. The retrieved professional knowledge documents, the enterprise data returned by the structured data interface, and the corresponding prompts for the specialized analytical agent are then concatenated according to a preset template to form the complete input for that specialized analytical agent.

[0127] The knowledge base recall process specifically includes the following sub-steps: Document parsing sub-step: The system uses OCR technology to recognize various types of collected data and generate computer-readable structured files. These data types include Word, Excel, PDF, and image formats. The structured file format is JSON. The JSON file is then converted into a Markdown file for subsequent segmentation.

[0128] Document segmentation steps: The system segments long documents into semantically complete segments. For documents with a clear structure, i.e., documents with chapters, sections, and paragraphs, segmentation is performed according to document hierarchy. The segmentation rule is that chapters are the highest level, sections are the second highest level, and paragraphs are the lowest level. The system starts segmenting from the lowest level, i.e., paragraphs. If the length of a paragraph exceeds 1,000 characters, it is further segmented into natural paragraphs. If the length of a paragraph is less than 50 characters, adjacent paragraphs are merged to ensure that each text block has sufficient semantic information. For documents without a clear structure, a segmentation method based on semantic integrity is adopted: The system uses the Sentence-BERT model to calculate the vector representation of each sentence in the document and calculates the cosine similarity between adjacent sentences. When the similarity is lower than a preset threshold, it is determined as a semantic boundary, and segmentation is performed accordingly. The preset threshold for semantic similarity is determined through statistical methods: On approximately 1,000 documents with marked segmentation boundaries, the system calculates the cosine similarity of adjacent sentences at each segmentation boundary and takes the average value as the benchmark. This average value is approximately 0.35. For unstructured documents, the system uses locations with a cosine similarity below 0.3 as segmentation points. After document segmentation, the segments are vectorized and stored in a vector database. The vectorization model uses the text2vec-large-chinese model, which is trained based on the CoSENT architecture and has a vector dimension of 768.

[0129] Document tagging steps: The system defines tags according to document granularity. Tags are added by the document administrator when the document is entered into the database. The dimension definitions and values ​​of the tags are derived from the classification standards defined in the commercial bank's internal document management specifications. The dimensions of the tag design and the value rules for each dimension are as follows: The source type dimension identifies the source and nature of a document, with values ​​including internal regulations, external regulations, regulatory documents, business data, and research reports. Internal regulations refer to internal management systems, operating procedures, and implementation details formulated by the bank's head office and various departments; external regulations refer to national laws, regulations, administrative regulations, and departmental rules; regulatory documents refer to guiding documents and regulatory requirements issued by financial regulatory agencies; business data refers to reports containing specific business statistics; and research reports refer to industry research reports, market analysis reports, and macroeconomic analysis reports. Document administrators determine the source type label based on the issuing entity and nature of the document: documents issued by the bank's head office or internal departments are marked as internal regulations; documents issued by the National People's Congress and its Standing Committee, the State Council, or various ministries are marked as external regulations; documents issued by financial regulatory agencies are marked as regulatory documents; documents primarily consisting of data tables and statistical charts are marked as business data; and documents primarily consisting of industry research and trend analysis are marked as research reports.

[0130] The business line dimension identifies the applicable business area of ​​a document, with values ​​including retail lending, corporate lending, post-loan management, risk compliance, inclusive finance, and products. Retail lending refers to documents related to loan business with individual customers; corporate lending refers to documents related to loan business with corporate customers; post-loan management refers to documents related to post-loan monitoring, risk warning, overdue collection, and non-performing loan disposal; risk compliance refers to policy documents in areas such as comprehensive risk management, compliance management, internal control management, and anti-money laundering; inclusive finance refers to documents related to inclusive finance business with micro and small enterprises, individual businesses, and agricultural customers; and products refer to product management methods and product descriptions for various credit products. The document administrator determines the business line label based on the applicable business scope clearly stated in the document title and the first paragraph of the text. For example, a document titled "Management Measures for Working Capital Loans to Micro and Small Enterprises" will be labeled with the business line tags "inclusive finance" and "corporate lending"; a document titled "Guidelines for Post-Loan Risk Monitoring" will be labeled with the business line tag "post-loan management."

[0131] The document type dimension identifies the functional type of a document, with values ​​including regulations, processes, contracts, notices, FAQs, data reports, and document numbers. Regulations refer to normative management rules and institutional documents, characterized by chapter structures such as "General Provisions," "Specific Provisions," and "Supplementary Provisions," as well as standardized terms such as "shall," "shall not," and "must." Processes refer to operating procedures and business process diagrams, characterized by step numbering and sequential descriptions. Contracts refer to various contract templates and agreement templates, characterized by contractual terms such as "Party A," "Party B," "whereas," and "this is agreed." Notices refer to business notices and announcements issued by various levels of institutions, characterized by terms such as "hereby notified," "this is hereby notified," and "please comply." FAQs refer to frequently asked questions and answers documents, characterized by a question-and-answer format. Data reports refer to periodic or ad hoc reports containing statistical analysis data. Document numbers refer to documents with official document numbers, formatted as the bank's document code plus the year plus a serial number; these documents typically have the highest enforcement power. Document administrators determine document type tags based on the document's format and content characteristics. For policy documents, the system can also automatically identify their structural features for auxiliary judgment: the system counts the frequency of occurrence of "chapter", "section" and "article" in the document. If the number of occurrences of "article" exceeds ten, it is judged as a policy document.

[0132] The timeliness dimension is used to identify the document's validity status and applicable time. Values ​​include publication date, currently valid, historical version, about to take effect, and repealed. Publication date refers to the document's publication or printing date, in year-month-day format, extracted by the document administrator from the document cover or document number. If the document does not have a specific publication date, the document's entry date is used. Currently valid means the document is currently in a valid execution state. Historical version refers to a version of the document that has been replaced by a newer version but is retained as historical data; the date of replacement and the document number of the replacement document must be indicated. About to take effect means the document has been published but has not yet reached its effective start date; the effective start date must be indicated. The system automatically switches the label from about to take effect to currently valid when the effective start date arrives. The effective start date is determined by reading the "Effective Date" clause text in the document and extracting the date using regular expressions. Repealed means the document has been repealed due to policy adjustments or expiration; the repeal date and the document number on which the repeal was based must be indicated. The timeliness label is determined by the document administrator based on the effective terms and publication information in the document. For documents that do not specify an effective date, the default effective start date is thirty calendar days after publication.

[0133] The security classification level is used to identify the confidentiality level of a document, with values ​​including public, internal, classified, and personal privacy. Public refers to documents that can be publicly released and disseminated; internal refers to documents that are only accessible and usable by bank employees and cannot be disseminated externally; classified refers to documents involving state secrets or the bank's core commercial secrets, and access to classified documents requires a special confidentiality approval process; personal privacy refers to documents containing customer personal information, credit information, and other personal privacy data, and access requires a personal privacy protection approval process. The security classification label is determined by the document administrator according to the security classification standards stipulated in the bank's internal confidentiality management system. When a document is added to the database, the document administrator must fill out a security classification assessment form, which must be approved by the head of the department before the corresponding security classification label can be affixed.

[0134] The "Applicable Customer Group" dimension identifies the types of customers to whom the document is applicable. Values ​​include individual customers, micro and small enterprises (MSEs), corporate customers, and internal employees. Individual customers refer to business scenarios applicable to individual clients; MSEs refer to business scenarios applicable to corporate clients meeting national MSE classification standards; corporate customers refer to business scenarios applicable to large and medium-sized enterprises; and internal employees refer to scenarios applicable to bank internal employee management and operational training. Document administrators determine the applicable customer group label based on the clearly defined applicable customer groups in the document text. If the document text explicitly applies to multiple customer groups, multiple values ​​can be used.

[0135] After tagging is completed, the system stores the tags as metadata information of the document along with the document content in the knowledge base. During subsequent searches, the system can match corresponding tag dimensions based on the business attributes of the user's question, narrowing the search scope. For example, when a user's question involves internal management regulations, the system limits the source type tag to "internal regulations"; when a user's question involves operational processes, the system limits the document type tag to "process". Taking post-loan management related policy documents as an example, its tag combination is: source type: internal regulations; business line: post-loan management; document type: policy; timeliness: currently valid; confidentiality level: internal; applicable customer group: corporate clients.

[0136] Keyword retrieval sub-step: The system utilizes the intent recognition capabilities of a large model to extract keywords from user questions. Keyword extraction employs an unsupervised extraction method based on the TextRank algorithm. The algorithm segments and tags the user question text, retaining nouns and verbs as candidate keywords. A co-occurrence graph is constructed among the candidate words, with edge weights equal to the co-occurrence frequency of two words within the same sliding window. The window size is five, and the TextRank value of each node is iteratively calculated until convergence. The three words with the highest TextRank values ​​are selected as the extracted keywords. Taking the user question "What management measures should be taken for overdue payments of sixty days?" as an example, the extracted keywords are "overdue," "sixty days," and "management measures." The system uses these keywords to search the knowledge base and scores the search results. The specific algorithm is S equal to n divided by N, where n is the number of keywords retrieved in the text block, and N is the number of characters in the text block. The design principle of this scoring algorithm is: the higher the keyword hit density, the stronger the direct relevance between the text block and the user question. Example of search results: Text block A contains 1 keyword and 100 characters, with an S value of 0.01; text block B contains 2 keywords and 500 characters, with an S value of 0.004; text block C contains 10 keywords and 5000 characters, with an S value of 0.002. The text blocks are sorted from largest to smallest S value as follows: Text Block A, Text Block B, and Text Block C.

[0137] Vector Recall Sub-Step: When a user asks a question, the system calls the Embedding model to vectorize the user's question, and then calculates the cosine similarity with all vectorized text blocks in the vector database. The cosine similarity formula is the dot product of two vectors divided by the product of their magnitudes, with a value ranging from -1 to +1. The closer the value is to +1, the more consistent the directions of the two vectors are in the semantic space. Taking the same three text blocks as an example, the cosine similarity between the user's question vector and text block B is 0.65, with text block C it is 0.48, and with text block A it is 0.35. The text blocks are sorted from largest to smallest cosine similarity as follows: text block B, text block C, and text block A.

[0138] The knowledge ranking sub-step involves the system fusing the results of keyword retrieval and vector recall, selecting the top five text blocks with the highest similarity after fusion. The formula for calculating the fused similarity is R = (1 / k) * (1 / k * (k * (k * (k * (k * (k * (k * (k * (k * (k * (k * (k * (k * (k * (k * (k * (k * (k * (k * (k * (k * (k * (k * (k * (k * (k * (integer, ("")" as it's a typo, should be 60)). The principle behind this formula is: the higher the ranking of the text block in both keyword retrieval and vector recall, the higher the fusion score; for each lower ranking, the contribution value decreases by dividing by that ranking); the value of k controls the rate of decrease, with a larger k resulting in a smoother decrease. Experimental testing showed that k equals 60, achieving the optimal fusion effect. On 300 test queries, the NDCG@5 index (Normalized Diminished Cumulative Gain), a measure of the ranking quality of the top five results, reached 0.76.) The R-values ​​for the three text blocks are calculated as follows: Text block B has a keyword ranking of 2 and a vector recall ranking of 1, so R equals 1 / 62 plus 1 / 61, approximately 0.0325; Text block A has a keyword ranking of 1 and a vector recall ranking of 3, so R equals 1 / 61 plus 1 / 63, approximately 0.0323; Text block C has a keyword ranking of 3 and a vector recall ranking of 2, so R equals 1 / 63 plus 1 / 62, approximately 0.0320. The ranking is: Text block B first, Text block A second, and Text block C third. The system recalls Text block B, which has the highest ranking, and sends it along with the user question, interface data, and prompt words to the large model, which then outputs the answer.

[0139] Step S7: Each professional analytical agent calls the large language model in parallel to generate structured analysis results for each dimension under behavioral boundary constraints.

[0140] In this step, each professional analysis agent calls the large language model in parallel, and sends the complete input formed in step S6 into the large language model. Under the behavioral boundary constraints of prompt words, the large language model generates structured analysis results of the corresponding dimensions based on the enterprise data returned by the structured data interface and the professional knowledge documents recalled by the knowledge base.

[0141] Each specialized analytical agent performs a systematic analysis of the data acquired for its respective dimension, following the analytical framework defined in step S5. Since each agent only processes data and performs analytical tasks within its specialized dimension, and the prompts clearly define the data source, analysis path, and output format, the agents operate in parallel without interfering with each other, and their respective structured analysis results are professionally targeted and consistent.

[0142] The specific implementation method of parallel scheduling in this step is as follows: The system maintains a thread pool at the large language model call layer, and the core number of threads in the thread pool is dynamically configured based on available computing resources. Taking a production environment deployed on an eight-core CPU and a single NVIDIA A100 GPU as an example, the system sets the maximum concurrent call count to five, meaning that a maximum of five agents are allowed to call the large language model interface in parallel. Each agent's call request is encapsulated as an independent asynchronous task and submitted to the thread pool for scheduling and execution. The lifecycle of each asynchronous task includes: The input data preparation stage, which organizes the prompts, interface data, and knowledge base documents assembled in step S6 into the standard input format of the large language model. This stage concatenates these three parts according to a preset template. The fixed structure of the template is: "[System Role Definition]" plus a role positioning section plus "[Analysis Task Description]" plus a subtask description plus "[Enterprise Data]" plus JSON data returned by the interface plus "[Reference Knowledge]" plus Markdown content of the knowledge base recall document. Each section is separated by a newline character, taking approximately 100 milliseconds. The model inference stage involves submitting the concatenated complete input text to the large language model service interface via an HTTP request. The large language model is based on the Transformer decoder architecture. The system employs a streaming decoding approach to generate output token by token. Each generated token is pushed to the agent via an SSE event for accumulation. The inference speed is approximately 30 to 50 tokens per second, and the total output length varies from 300 to 800 tokens depending on the dimension, taking approximately 3 to 8 seconds. In the output parsing phase, after accumulating the streaming output returned by the large language model, regular expressions are used to extract structured fields from the accumulated text. The extraction rules are: matching "Risk Level:" followed by one of four keywords: high risk, medium risk, low risk, or cannot be determined; matching all text after "Risk Basis:" up to a newline character; and matching all text after "Recommendation:" up to the end. After extraction, the text is assembled into a JSON object in the unified format of step S7, taking approximately 200 milliseconds. The total time for the above three stages is approximately 3.3 to 8.3 seconds. The actual total analysis time is the maximum of the completion times of each agent, not the sum of the times of each agent.

[0143] Taking the parallel execution of five agents as an example, the model inference phase of the financial health agent takes approximately 6.2 seconds, the judicial risk agent approximately 4.8 seconds, the public opinion monitoring agent approximately 3.5 seconds, the related risk agent approximately 4.1 seconds, and the early warning assessment agent approximately 5.3 seconds. The time difference between each agent depends on the length and complexity of its output: the financial health agent needs to output a detailed report containing nine analysis steps, outputting approximately 800 tokens, making it the longest taking; the public opinion monitoring agent outputs approximately 300 tokens, making it the shortest taking. Since each agent executes in parallel in an independent thread, the total analysis time is approximately 6.3 seconds, meaning the slowest financial health agent's completion time of 6.2 seconds plus the input data preparation and output parsing phases of approximately 300 milliseconds. If a serial execution method is adopted, with five agents executing sequentially, the total analysis time is approximately 6.2 + 4.8 + 3.5 + 4.1 + 5.3 = 23.9 seconds. Adding the context switching and input / output overhead between agents, approximately 2.1 seconds, the total is approximately 26 seconds. Further considering the queuing time of the large language model service interface, each request needs to be queued during single-GPU serial execution, increasing the queuing time by approximately 36 seconds, for a total of approximately 62 seconds. The efficiency improvement of parallel execution mainly comes from two aspects: First, when the large language model inference process is computed on the GPU, a certain degree of concurrent requests can improve resource utilization through the GPU's batch processing capabilities. When multiple inference requests arrive simultaneously, the GPU can concatenate the inputs of these requests into a batch for matrix operations. According to actual tests, the total computation time for processing five requests simultaneously on a single GPU is approximately 60% to 70% of the total time for serial processing, rather than increasing strictly linearly. This is because the computational complexity of the Attention mechanism increases quadratically with the sequence length. In batch processing, the total computational cost of multiple short sequences is less than that of a single long sequence, and the GPU's computing resources are more fully utilized in batch processing. Second, it reduces context switching and queuing time between agents. The comparison data of the two execution methods were measured based on the same batch of enterprise data, the same test set, and the same large language model version. The test environment was a single NVIDIA A100 GPU, an eight-core CPU, and 128GB of memory, with a test sample size of fifty enterprises.

[0144] The structured analysis results output by each agent adopt a unified JSON format specification, containing the following fixed fields: Dimension Identifier field, which takes the name of the analysis dimension corresponding to the agent; Subject Identifier field, which takes the unified social credit code of the enterprise; Analysis Time field, which takes the timestamp of the analysis completion; Analysis Conclusion field, which takes the text of the analysis conclusion of the agent; Data Source field, which takes the list of structured data interface names called by the agent and the list of document names of documents recalled from the knowledge base; Risk Level field, which takes the risk level judgment result of the agent for this dimension, including four enumerated values: high risk, medium risk, low risk, and cannot be determined; Risk Basis field, which takes the specific indicator value that triggered the risk level judgment and the corresponding judgment benchmark, in the form of a key-value pair list, where each key-value pair contains an indicator name, indicator value, benchmark value, and comparison relationship. The unified structured output specification ensures that the agent in the subsequent step S8 can accurately parse and aggregate the analysis conclusions of each dimension. In addition, each agent must comply with the following rules in its output: If the data interface required for a certain analysis step returns an empty value or is marked as "data not yet acquired", then the output of that step is "data not yet acquired, this part will not be analyzed for the time being", and no model knowledge of its own can be used to complete or speculate; the risk level determination must be consistent with the output of the early warning rule engine in step S2. If the early warning rule engine outputs high risk, then the risk level field output by the agent must be high risk, and cannot be downgraded to medium risk or low risk.

[0145] Step S8: By comprehensively summarizing the analysis results received by the intelligent agent from various dimensions, the results are aggregated into a unified structured comprehensive evaluation report and displayed in a progressive streaming output manner.

[0146] This step comprehensively summarizes the structured analysis results output by various professional analysis agents in step S7, aggregates the structured analysis results of each dimension into a unified structured comprehensive evaluation report according to the preset report structure, and presents it to the user in stages according to a progressive streaming output method.

[0147] After summarizing the structured analysis results output by each professional analysis agent in step S7 of the agent receiving process, the following aggregation process is executed.

[0148] First, result aggregation and verification. The comprehensive summary agent maintains a result aggregation buffer. This buffer uses the unified social credit code of the enterprise to be analyzed as the key and the result storage slots for all analysis dimensions as the values. Each storage slot is used to store the output results of the agent for the corresponding dimension. When the analysis result for a certain dimension arrives, the comprehensive summary agent first verifies the integrity of the result. The verification rules include: whether the dimension identifier field exists and its value is within the range of twelve dimensions; whether the subject identifier field is consistent with the current session subject; whether the analysis time field is within the valid time range; whether the analysis conclusion field is not empty; whether the risk level field is one of the four enumerated values: high risk, medium risk, low risk, or cannot be determined; and whether the risk basis field contains a specific indicator name and value. For results that fail the integrity verification, the comprehensive summary agent records an anomaly log and discards the result. The corresponding dimension will be marked as "analysis result abnormal" in the subsequent report. When the analysis results corresponding to all distributed subtasks have been returned or the waiting timeout is set to sixty seconds from the time of distribution, the comprehensive summary agent enters the aggregation phase.

[0149] Second, the dimension order is arranged. The comprehensive summary agent organizes the analysis conclusions of each dimension according to the preset dimension display order. The dimension display order is set based on the business logic of post-loan risk assessment: macroeconomic and industry analysis provide a judgment on the external environment in which the enterprise is located, belonging to the macro-level background analysis, and are displayed first; financial health and operating status reflect the enterprise's own operating fundamentals, belonging to the meso-level subject analysis, and are displayed next; judicial risk, public opinion monitoring, and related risks reflect the external event impact and related transmission risks faced by the enterprise, belonging to the micro-level risk event analysis, and are displayed again; early warning assessment is based on the comprehensive risk level assessment of the aforementioned dimensions, serving as a summary analysis that connects the above and the following, and is placed after the specific analysis of each dimension; policy analysis and intellectual property, etc., are displayed depending on whether there is relevant information. If there is information, they are inserted between industry analysis and financial analysis; if there is no information, they are skipped. The specific display order is: macroeconomic, industry analysis, policy analysis, financial health, operating status, intellectual property, judicial risk, public opinion monitoring, related risk, operating risk, early warning assessment, and other questions. For dimensions without relevant information, the comprehensive summary agent outputs "No relevant information available".

[0150] Third, report content organization. The comprehensive summary agent organizes the report content according to a three-part structure: general, specific, and recommendations. The summary section outlines the overall risk profile of the enterprise, limited to 200 words, and covers the comprehensive risk level and core risk points. The comprehensive summary agent determines the comprehensive risk level based on the output of the early warning assessment agent: if the early warning assessment agent outputs high risk, the risk level in the summary section is also high risk; if the early warning assessment agent outputs medium risk but two or more of the three dimensions of financial health, legal risk, and related risks output high risk, the comprehensive summary agent will upgrade the comprehensive risk level to high risk. This upgrade rule is based on the principle in the commercial bank's internal risk control rules that "when multiple dimensions of risk overlap, the risk level should be strictly determined." The specific sections unfold the analysis conclusions of each dimension in the order of dimension presentation. The content structure of each dimension is: dimension title, risk level identifier, analysis conclusion text, and data source label. The data source label format is "Data Source: Interface Name + Document Name" to ensure the traceability of the analysis conclusions. The recommendations section provides comprehensive risk management suggestions, generated based on a combination of overall risk level and core risk points. High risk corresponds to "Immediately activate the post-loan risk management plan, including but not limited to requiring the company to supplement credit enhancement measures, reducing credit limits, and increasing the frequency of post-loan inspections." Medium risk corresponds to "Strengthen the frequency of post-loan monitoring and pay attention to changes in risk trends." Low risk corresponds to "Maintain routine post-loan management." If the core risk point is prominent legal risk, the recommendations include "Closely monitor the progress of disputes over construction payments for projects under construction and the status of asset preservation." If the core risk point is prominent financial risk, the recommendations include "Require the company to provide a detailed accounts receivable collection plan and cash flow improvement plan." The overall summary agent responds only based on existing information, without fabricating data or subjectively assessing the quality of companies.

[0151] After aggregation is complete, the summarizing agent transmits the aggregation report to the decision application layer. The decision application layer uses SSE streaming output to display the analysis results in stages.

[0152] The implementation details of SSE technology are as follows: The system establishes a long-lived connection channel based on the HTTP protocol between the front-end and back-end. The front-end initiates a connection request through the EventSource API, with the request URL including a session identifier as a parameter. After receiving the connection request, the back-end keeps the connection open without returning a response end marker. Unlike the traditional HTTP request-response model, after the SSE connection is established, the server can actively send multiple events to the client. Each event is transmitted in the format of "event:" followed by the event type and a newline character, then "data:" followed by the event data and a newline character, followed by two newline characters. The client receives the events pushed by the server through the onmessage callback function of the EventSource API. When a complete event is received, the rendering callback function is immediately triggered to update the page content. The advantages of SSE technology compared to WebSocket include: simpler implementation, no need for handshake protocol upgrades; no need for bidirectional communication, only one-way push from the server to the client; automatic support for reconnection after disconnection, with the client automatically attempting to re-establish the connection every three seconds after a disconnection, with a maximum of ten retries. Compared to the polling method, SSE only pushes when there is new data, reducing the overhead of invalid requests by approximately 90%.

[0153] The decision application layer presents the analysis results to users in the following three stages: The first stage is the summary and response stage. The decision application layer will comprehensively summarize the summary information generated by the intelligent agent, which is approximately 200 words in length, encapsulate it into the first SSE event, and mark the event type as "summary," then push it to the front end. The large language model takes approximately 1.5 to 2.5 seconds to generate the summary information, and the system pushes it immediately after the summary information is generated. After receiving the event, the client parses the event data through the onmessage callback function and immediately renders and displays the summary content in the upper area of ​​the page. Based on practical experience, the time from submitting the question to seeing the summary content is approximately two to three seconds. Within this time period, the user does not need to wait for all the content to be generated to obtain the overall analysis framework and core conclusions.

[0154] The second phase is the dimension-by-dimensional population phase. Following the dimension display order, after generating complete content for each dimension, the decision application layer encapsulates the analysis content of that dimension into an independent SSE event, labeled "dimension." The event data is a JSON object containing the dimension name, risk level, analysis conclusion, and source annotation, and is pushed to the front end. The order of dimensions pushed is consistent with the dimension order arrangement in step S8. Upon receiving an event for a specific dimension, the front end parses the event data through the onmessage callback function and creates a new dimension panel on the page. The panel title displays the dimension name and risk level indicator, distinguished by different colors: high risk in red, medium risk in orange, low risk in blue, and no relevant information in gray. The panel content area displays the analysis conclusion text and data source annotation. Each dimension panel is readable immediately after rendering, without requiring the user to wait for subsequent dimensions to complete. The generation and push of each dimension's content proceed in parallel with its front end rendering, generating and pushing one dimension at a time.

[0155] The third stage is the comprehensive evaluation stage. Once all the content corresponding to the distributed sub-tasks has been pushed out, the decision application layer generates a comprehensive risk rating and handling recommendations. The comprehensive risk rating is determined by the comprehensive summary agent based on the analysis results of all dimensions, and the handling recommendations are generated based on a combination of risk level and core risk points. The decision application layer encapsulates the comprehensive risk rating and handling recommendations into a final SSE event, marked as "final," and pushes it to the front end. Upon receiving this event, the front end renders the comprehensive risk assessment conclusion in the bottom area of ​​the page, including the comprehensive risk level, a list of major risk sources, and specific handling recommendations.

[0156] For long report scenarios exceeding the context window length limit of the large language model, the system employs a sliding window technique: when the accumulated output content length exceeds 70% of the 8000 tokens of the large language model context window (i.e., 5600 tokens), the system removes the previously stably output dimensional content from the context window and caches it in the session storage, ensuring that subsequent generation processes do not exceed the context window limit. The cached content is retrieved from the session storage when backtracking is needed, without occupying the large language model's context window.

[0157] The system maintains the analysis subject, analyzed dimensions, and historical Q&A records for the current session. Session management is implemented as follows: the system assigns a unique session identifier to each user session. The session object stores the unified social credit code of the current analysis subject, a list of completed dimension names, a summary of the analysis results for each dimension, and summaries of previous user questions and their corresponding answers. All data is stored in a Redis cache. Redis is an in-memory data structure storage system that uses key-value pairs to store data, with read / write speeds in the microsecond range. Compared to disk storage, it significantly improves the read / write efficiency of session data. The session timeout is set to thirty minutes, after which resources are automatically released.

[0158] The system supports users to ask follow-up questions based on the analysis results. The follow-up question processing flow is as follows: After the user inputs a new question, the system first determines whether the new question contains a specific company name by calling the named entity recognition model in step S4. If it does, the company identified in the new question is used as the new analysis subject, and steps S4 to S8 are executed in their entirety; if it does not, the most recently identified analysis subject is read from the current session object and used as the analysis subject for the new question. Subsequently, the system uses the completed analysis dimensions and corresponding analysis result summaries in the current session as context information, and inputs them together with the original question into the question splitting agent for intent recognition. When analyzing the intent of a new question, the problem-splitting agent refers to the list of completed dimensions to avoid redundant analysis: if the analysis of a new question's dimension has already been completed in the current session, the system directly returns the cached analysis results; if the new question involves a new dimension that has not yet been analyzed, the system only performs the complete analysis process of steps S5 to S8 for the new dimension; if the new question involves a deeper inquiry into an existing dimension, i.e., requiring more details, the system re-executes steps S5 to S8 for that dimension, but during the data assembly stage, it only calls the refined data interface related to that deeper inquiry, rather than the full interface, to shorten response time. To avoid redundant descriptions, when outputting the answer, the system states in the prompt, "The following is a supplementary analysis based on your inquiry; the conclusions of the previous analysis will not be repeated," and the large language model generates an answer containing only the newly added content accordingly.

[0159] All sessions are archived chronologically, allowing for retrospective viewing. Session archiving is implemented as follows: after a session ends, the system serializes the session object into JSON format text and stores it in an Elasticsearch index. Elasticsearch is a distributed search and analytics engine that uses an inverted index structure and supports fast full-text search of stored documents. Archived content includes the original question entered by the user each time, the task decomposition results of the question-and-answer agent, the analysis results for each dimension, a comprehensive evaluation report, the session timestamp, and the session duration. Account managers can retrieve historical session records through the history entry in the Q&A interface by company name, time range, or question keywords. During the search, the system converts the user-input query conditions into Elasticsearch's DSL query statement, executes the search, and returns a list of matching session records in reverse chronological order. Clicking on any record allows viewing the complete analysis process and conclusions of that Q&A session.

[0160] The application supports users uploading multimodal files such as audio, photos, and documents. The system uses the parsed file information as input, forming part of the user's question. The multimodal file upload function is implemented as follows: users select local files through the file upload control on the question-and-answer interface, and the system identifies the file category based on the file extension and MIME type. Audio files support three common formats: MP3, WAV, and M4A. The system calls a speech recognition engine to convert the audio content into text. The speech recognition engine uses a fine-tuned version of the Whisper end-to-end speech recognition model based on Transformer, fine-tuned on approximately 1,000 hours of financial domain corpus, achieving a domain-specific vocabulary recognition accuracy of approximately 95%. Photo files support JPG and PNG formats. The system calls an OCR engine to extract text information from the photos. The OCR engine uses an end-to-end recognition architecture of CNN plus Transformer. The CNN part uses ResNet-50 as the backbone network for visual feature extraction, and the Transformer part uses a six-layer encoder-decoder structure, trained on financial document images such as contracts and bills, achieving a text recognition accuracy of approximately 90%. The document files support both PDF and DOCX formats. The system uses a document parsing service to extract all text content from the documents. For PDF files, the PyPDF2 parsing library is used to extract text, and for DOCX files, the python-docx parsing library is used, achieving a parsing success rate of 99%. The system uses the parsed text content as supplementary context for the user's question, submitting it along with the user's input text question to the question-splitting agent for processing. This allows the large language model to answer based on the information in the user-uploaded file. For example, when a user uploads a company's financial statement PDF and asks, "Help me analyze this company's solvency," the system parses the balance sheet, income statement, and cash flow statement data from the financial statement into structured text, which is then used as input along with the user's question. The financial health agent can then perform analysis based on the specific values ​​in the financial statement without needing to separately call the financial statement data interface.

[0161] Example 2 like Figure 2As shown in the figure, this embodiment provides an intelligent question-answering system based on a large language model for post-loan risk management. The system comprises eight core functional modules. User-input questions are first sent to the intent parsing module, which uses a question-splitting agent to identify intent and break down tasks. The resulting sub-tasks are then assigned to corresponding professional analysis agents by the intelligent routing module. The data acquisition module is responsible for accessing multi-source heterogeneous data and processing standardized indicators, generating user scores and risk signals. The early warning judgment module performs deterministic risk judgment based on an early warning rule engine. The data interface module encapsulates the above data into a structured data interface for use by various professional analysis agents. The data assembly module assembles complete input (including interface data, knowledge base recall documents, and prompt words) for each professional analysis agent. The parallel analysis module controls each agent to call the large language model in parallel for analysis. The result aggregation module aggregates all analysis results through a comprehensive summarization agent and outputs them in a streaming manner. Solid arrows in the figure represent data flows or call relationships, while dashed arrows represent the correspondence between modules and agents.

[0162] The specific functions of each module are as follows.

[0163] The data acquisition module is used to construct a unified access and standardized indicator processing pipeline for multi-source heterogeneous data. This module acquires internal data from the bank's credit system, core transaction system, and CRM system, and external data from sources such as industry and commerce, judiciary, taxation, credit reporting, and public opinion through external data interfaces. The acquired multi-source raw data is cleaned, aligned, and merged according to preset business logic, and processed into multi-dimensional standardized indicators. This module is also used to build a scorecard model based on standardized indicators to generate user scores, and to generate risk signals according to preset risk signal definition rules. For a detailed description of the implementation of this module, please refer to step S1 in Example 1.

[0164] The early warning determination module is used to perform deterministic risk assessment on standardized indicators and user scores generated by the data collection module based on preset indicator thresholds and scoring segments, through an early warning rule engine, and output structured early warning results. For the specific implementation of this module, please refer to the relevant description in step S2 of Example 1.

[0165] The data interface module is used to establish a query database. It stores standardized indicators, user scores, and risk signals generated by the data acquisition module, as well as early warning results generated by the early warning judgment module, into the query database. Based on the query database, it encapsulates these data into a structured data interface for use by the large language model, categorized by risk dimensions. For a detailed description of the implementation of this module, please refer to step S3 in Example 1.

[0166] The intent parsing module is used to acquire user input questions. Through a question-splitting agent, it identifies the user's intent in the question, parses out the analysis subject and analysis dimensions, and breaks down the question into at least one sub-task according to the analysis dimensions. For a detailed description of the implementation of this module, please refer to step S4 of Example 1.

[0167] The intelligent routing module is used to allocate each subtask to a corresponding professional analytical agent based on the analysis dimensions corresponding to the subtasks broken down by the intent parsing module. Each professional analytical agent predefines its professional role, core analysis framework, and behavioral boundary constraints based on the prompt word engineering. Data usage constraints stipulate that all analyses must be based on enterprise data returned by the structured data interface, prohibiting the large language model from fabricating its own data. For a detailed description of the implementation of this module, please refer to step S5 of Example 1.

[0168] The data assembly module is used by various professional analytical agents to call the structured data interface encapsulated by the data interface module to obtain enterprise data corresponding to the analytical subject. Simultaneously, the knowledge base retrieval module performs semantic retrieval on the user's question, searching the knowledge base for professional knowledge documents related to the sub-task. The retrieved professional knowledge documents, the enterprise data returned by the structured data interface, and the corresponding prompts for the professional analytical agent are then concatenated according to a preset template to form the complete input for that professional analytical agent. For a detailed description of the implementation of this module, please refer to step S6 of Embodiment 1.

[0169] The parallel analysis module is used by various professional analysis agents to call the large language model in parallel. It feeds the complete input formed by the data assembly module into the large language model, enabling the model to generate structured analysis results for corresponding dimensions based on enterprise data returned by the structured data interface and professional knowledge documents recalled from the knowledge base, under the behavioral boundary constraints of prompt words. For the specific implementation of this module, please refer to the relevant description in step S7 of Example 1.

[0170] The results aggregation module is used to comprehensively summarize the structured analysis results of various professional analysis agents received from the parallel analysis module. It aggregates the structured analysis results from each dimension into a unified structured comprehensive evaluation report according to a preset report structure, and presents it to the user in stages using a progressive streaming output method. For the specific implementation of this module, please refer to the relevant description in step S8 of Example 1.

[0171] Example 3 This embodiment is a specific implementation case of the method and system in a fast query scenario.

[0172] Scenario Description: A commercial bank's account manager needs to quickly understand the recent changes in legal risks of Company A during routine post-loan monitoring.

[0173] Implementation steps: Users can enter "Check if A has any recent lawsuits" in the Q&A interface.

[0174] The problem-splitting agent in the intent parsing module parses the user's question, identifies the analysis subject as "A," and determines it to be "Company A" through fuzzy matching. The analysis dimension is the judicial risk dimension. Since this question only involves the judicial risk dimension, the problem-splitting agent breaks it down into a sub-task and assigns it to the judicial risk dimension.

[0175] Based on the above breakdown results, the intelligent routing module routes the subtasks to the judicial risk intelligent agent.

[0176] The judicial risk agent in the data assembly module calls the judicial-related structured data interface in the data interface module to obtain the latest litigation information, enforcement information, and credit default information of Company A. The interface returns the following data: two new lawsuits in the past three months, totaling RMB 4.5 million; one new enforcement case, amounting to RMB 800,000; and no current credit default records. The knowledge base recall module simultaneously searches for professional knowledge documents related to judicial risk. The document recalled this time is the relevant clause on judicial risk monitoring in the "Guidelines for Post-Loan Management of Credit Business".

[0177] The judicial risk intelligence agent combines the acquired interface data, retrieved professional knowledge documents, and corresponding prompts into a complete input, then calls a large language model to generate structured analysis results. Based on the interface data, the large language model outputs: two new lawsuits in the past three months, totaling 4.5 million yuan; one enforcement case, amounting to 800,000 yuan; no current record of dishonesty. The overall assessment is a medium-risk signal, and continued monitoring is recommended.

[0178] The comprehensive summary agent in the results aggregation module receives the above analysis results and organizes the output content according to the structure of summarizing, segmenting the signals, and giving suggestions.

[0179] The decision application layer displays the information in stages via SSE streaming output: The first stage outputs "Company A's recent legal risk is medium risk, mainly involving new litigation and enforcement proceedings"; the second stage outputs specific details such as "2 new litigations totaling 4.5 million yuan" and "1 new enforcement case involving 800,000 yuan"; the third stage outputs a rating and recommendation of "Comprehensive assessment: medium risk, continued monitoring recommended". The entire response takes approximately five seconds.

[0180] Implementation results: Account managers can quickly grasp the dynamics of a company's legal risks within seconds without having to search through the judicial system one by one.

[0181] Example 4 This embodiment is a specific implementation case of the method and system in a comprehensive risk analysis scenario.

[0182] Scenario Description: A commercial bank's account manager needs to conduct a comprehensive risk assessment of "Group A Limited Company" during post-loan management, focusing on financial indicators and operational risks.

[0183] Implementation steps: Users can enter "How are Group A's operations and financial situation?" through the Q&A interface.

[0184] The problem-splitting agent in the intent parsing module analyzes the user's question, identifying the analysis subject as "Group A," which is then determined to be "Group A Limited Company" through fuzzy matching. The analysis dimensions are financial health and operational status. The problem-splitting agent breaks the question down into two sub-tasks, which are then assigned to the financial health and operational status dimensions, respectively.

[0185] Based on the breakdown results, the intelligent routing module routes the financial health subtask to the financial health intelligent agent and the business status subtask to the business status intelligent agent.

[0186] The financial health agent in the data assembly module calls the financial statement data interface in the data interface module to obtain the financial indicator data of Group A Co., Ltd.: debt-to-equity ratio of 78.5%, current ratio of 1.2%, quick ratio of 0.85%, interest coverage ratio of 2.3 times, return on equity of 4.5%, total asset turnover of 0.38, accounts receivable turnover of 1.9 times, and net operating cash flow as a percentage of operating revenue of -3.2%. The industry benchmark for the above indicators is: the construction industry to which this company belongs, with an industry average debt-to-equity ratio of 75%, an industry average current ratio of 1.2, an industry average quick ratio of 0.9, and an industry average return on equity of 6%. The operating status agent calls the business registration and operating information interfaces in the data interface module to obtain the operating data of Group A Co., Ltd.: one administrative penalty was imposed in the past year for a violation of safety production regulations, with a fine of 200,000 yuan; the legal representative of the company changed three months ago; the company's current score is 788 points, corresponding to a medium risk level.

[0187] The financial health intelligence system conducts a systematic analysis according to its pre-set nine-step analysis framework: In terms of overall financial evaluation, the debt-to-equity ratio of 78.5% is slightly higher than the industry average of 75%, indicating manageable debt repayment pressure but requiring attention; regarding solvency, the current ratio of 1.2 is in line with the industry average, while the quick ratio of 0.85 is lower than the industry average of 0.9, placing short-term solvency at a moderate level within the industry; the interest coverage ratio of 2.3 times is higher than the regulatory warning value of 1.5 times, indicating acceptable interest payment ability; in terms of profitability, the return on net assets is 4%. The company's profitability is 6% lower than the industry average, placing it in the lower-middle range. In terms of operational efficiency, its total asset turnover of 0.38 is 0.6% lower than the industry average, its accounts receivable turnover of 1.9 times, and its payment period of approximately 189 days, indicating slow cash flow. Regarding cash flow, its net operating cash flow is negative, requiring monitoring of its cash flow improvement. The bank's credit risk assessment indicates that the company's current credit utilization rate is approximately 70%, suggesting room for further credit, but its cash flow situation needs close monitoring. The credit recommendation is to maintain the existing credit line and strengthen post-loan cash flow monitoring.

[0188] The operational status intelligence agent conducts a systematic analysis according to its preset analysis direction: In terms of operational stability, the legal representative of the enterprise changed three months ago, which is a recent major management change. Attention should be paid to the reasons for the change and the background of the new legal representative; In terms of compliance risk, there was one administrative penalty for violation of safety production regulations in the past year. Although the fine was not large, it reflects the weakness of the enterprise in safety management; The enterprise score of 788 points corresponds to a medium risk level. Combining the two risk events of administrative penalty and change of legal representative, the overall assessment is that the operational risk is medium to high.

[0189] Each of the two agents transmits its analysis results to the comprehensive summary agent. The comprehensive summary agent organizes its report content according to a general, specific, and recommendation structure: First, it provides an overview of the overall operations and financial situation of Group A Limited, pointing out key risk points such as a high debt-to-asset ratio, profitability below the industry average, recent change of legal representative, and existing administrative penalties; then, it details the specific analysis conclusions of the financial and operational conditions; finally, it gives a comprehensive risk assessment of moderate to high, recommending that the new legal representative's business strategy and cash flow improvement be given special attention in the next quarter's post-loan inspection.

[0190] The decision application layer displays results in stages via SSE streaming output: the first stage outputs summary information in approximately 2.5 seconds; the second stage outputs detailed analysis of financial health and operational status dimension by dimension; and the third stage outputs a comprehensive risk rating and handling recommendations. The overall response time is approximately 30 seconds.

[0191] Implementation results: The comprehensive risk assessment, which originally required more than 30 minutes of manual cross-system queries, can now be completed automatically within 30 seconds, and the analysis results are accurate, reliable, and traceable.

[0192] Example 5 This embodiment is a specific implementation case of the method and system in a knowledge base retrieval scenario.

[0193] Scenario description: A customer manager at a commercial bank inquired about management measures for loans overdue for sixty days during post-loan management.

[0194] Implementation steps: Users can enter "What management measures should be taken if the payment is overdue for 60 days" through the Q&A interface.

[0195] The question-splitting agent in the intent parsing module parses the user's question, identifies that the question involves the policy analysis dimension, and assigns the question to the policy analysis agent.

[0196] The intelligent routing module routes subtasks to the policy analysis agent.

[0197] The policy analysis agent in the data assembly module initiates the knowledge base recall process. The knowledge base recall module first limits the search scope to knowledge content tagged with "internal regulations" during the document tagging stage, based on the business attributes of the question. Then, it performs keyword retrieval: using the keyword "overdue," it finds three text blocks, sorted by their S-values ​​from largest to smallest. Text block A has an S-value of 1 / 100 (0.01), text block B has an S-value of 2 / 500 (0.004), and text block C has an S-value of 10 / 5000 (0.002). Simultaneously, it performs vector recall: the vectorized result of the user question is compared with the vectorized results of the three text blocks using cosine similarity calculation. The similarity for text block B is 0.65, for text block C it is 0.48, and for text block A it is 0.35. The knowledge ranking module integrates the results of keyword retrieval and vector recall. Following the R-value calculation formula, with k set to 60, it calculates the fusion similarity of each text block: the R-value for text block B is 1 / 62 + 1 / 61, approximately 0.0325; the R-value for text block A is 1 / 61 + 1 / 63, approximately 0.0323; and the R-value for text block C is 1 / 63 + 1 / 62, approximately 0.0320. The ranking result is: text block B first, text block A second, and text block C third. The system recalls the top-ranked text block B.

[0198] The content of text block B contains the relevant clauses of the internal regulations regarding the management of overdue loans, specifically including: for overdue loans within 30 days, the account manager shall conduct telephone collection; for overdue loans between 31 and 60 days, the team leader shall take the lead in formulating a collection plan; for overdue loans between 61 and 90 days, a post-loan risk disposal plan shall be initiated, including but not limited to on-site visits, additional collateral measures, and requiring the enterprise to provide a repayment plan; for overdue loans exceeding 90 days, the loan shall be adjusted to a non-performing loan and transferred to the legal department for handling.

[0199] The policy analysis agent concatenates the content of the recalled text block B, prompts, and user questions into a complete input, and calls the large language model to generate a structured answer. Based on the internal regulations in text block B, the large language model outputs management measures for overdue loans of 60 days: This loan has entered a critical 60-day overdue stage, and according to internal regulations, the post-loan risk management plan should be activated immediately; specific measures include completing an on-site visit within five working days to assess the company's operating conditions and willingness to repay; requiring the company to provide a written repayment plan and additional effective collateral; increasing the frequency of post-loan inspections from once per quarter to once per month; if the company is still unable to resume normal repayments within the next 30 days, the non-performing loan identification procedure must be initiated.

[0200] The policy analysis agent sends its answer to the comprehensive summary agent. The comprehensive summary agent organizes the answer according to the structure of summary, detailed measures, and precautions, and then displays it in stages through SSE streaming output.

[0201] Implementation results: Questions that previously required manual searching and understanding of policy documents can now be answered automatically within 30 seconds, and the analysis results are accurate, reliable, and traceable.

[0202] The embodiments described above are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent question-answering method based on a large language model in post-loan risk management, characterized in that, Includes the following steps: Step S1: Construct a unified access and standardized indicator processing pipeline for multi-source heterogeneous data, acquire multi-source raw data from within and outside the industry, clean, align and fuse it, and process it into multi-dimensional standardized indicators; perform feature screening and indicator screening based on the standardized indicators, construct a scorecard model to generate user scores, and perform event-based risk judgment on the standardized indicators to generate risk signals; Step S2: Based on preset indicator thresholds and scoring segments, the early warning rule engine performs deterministic risk assessment on the standardized indicators and user scores, and outputs structured early warning results. Step S3: Establish a query database, store the standardized indicators, user scores, risk signals and early warning results into the query database, and encapsulate them into a structured data interface for large language models to call based on the risk dimension in the query database. Step S4: Obtain the question input by the user, perform intent recognition on the question through the question splitting agent, parse out the analysis subject and analysis dimension of the question, and split the question into at least one sub-task according to the analysis dimension; Step S5: Based on the analysis dimensions corresponding to each subtask, assign each subtask to the corresponding professional analysis agent through intelligent routing; the professional analysis agent predefines its professional role positioning, core analysis framework and behavioral boundary constraints based on the prompt word engineering. The behavioral boundary constraints include data usage constraints, which restrict all analyses to be based on the enterprise data returned by the structured data interface and prohibit the large language model from fabricating data on its own. Step S6: Each of the professional analysis agents calls the structured data interface to obtain the enterprise data corresponding to the analysis subject. At the same time, it retrieves professional knowledge documents related to the subtask from the knowledge base through the knowledge base retrieval module. The retrieved professional knowledge documents are then concatenated with the enterprise data and the prompt words corresponding to the professional analysis agent to form the complete input of the professional analysis agent. Step S7: Each of the specialized analytical agents calls the large language model in parallel, and sends the complete input into the large language model, so that the large language model generates the structured analysis results of the corresponding dimension under the behavioral boundary constraints of the prompt word; Step S8: By comprehensively summarizing the structured analysis results received by the intelligent agent, and aggregating them into a unified structured comprehensive evaluation report according to the preset report structure, the report is presented to the user in stages according to a progressive streaming output method.

2. The intelligent question-answering method based on a large language model in post-loan risk management according to claim 1, characterized in that, In step S1, the multi-dimensional standardized indicators include the asset-liability ratio, net profit, interest coverage ratio, loan-to-deposit ratio, collateral ratio, credit utilization rate, maximum transaction amount of the lender in the past 30 days, number of corporate legal person changes, number of corporate lawsuits in the past year, number of negative public opinion events in the past six months, number of overdue payments in the past six months, and number of social security contributors. The specific method for cleaning, aligning, and integrating the acquired multi-source raw data is as follows: change logs are captured in real time from the bank's credit system, core transaction system, and CRM system through a distributed data integration framework and pushed to the Apache Kafka message queue. Incremental data is pulled from external data sources through an HTTP RESTful API in a scheduled manner. The cleaning includes converting date fields from different sources to ISO8601 format, establishing a mapping relationship between source field names and target field names, and deduplicating by comparing the combination of the unified social credit code, data date, and data source fields.

3. The intelligent question-answering method based on a large language model in post-loan risk management according to claim 1, characterized in that, In step S2, the judgment logic of the early warning rule engine includes single indicator threshold triggering, multi-indicator combination triggering, and trend change triggering; the single indicator threshold triggering is to trigger an early warning of the corresponding level when the value of a single standardized indicator exceeds a preset threshold; the multi-indicator combination triggering is to output a combined early warning level when multiple indicators simultaneously meet their respective triggering conditions; and the trend change triggering is to trigger an early warning when an indicator shows a deteriorating trend for several consecutive periods. The warning levels are divided into five levels: high risk, medium-high risk, medium risk, medium-low risk, and low risk.

4. The intelligent question-answering method based on a large language model in post-loan risk management according to claim 1, characterized in that, In step S4, the specific method for parsing out the analysis subject and analysis dimension of the problem is as follows: call the named entity recognition model based on the BERT-BiLSTM-CRF architecture to output B-ORG, I-ORG or O label for each character in the problem text, and concatenate the continuous character sequence marked with B-ORG and I-ORG to form a candidate enterprise name; perform exact matching and alias matching between the candidate enterprise name and the full name, abbreviation and former name of the enterprise in the pre-built management list database. If no match is found, calculate the edit distance between the candidate enterprise name and the full name of each enterprise in the management list, and take the enterprise with the smallest edit distance as the matching result, and require that the minimum edit distance does not exceed one-third of the length of the candidate enterprise name; The intent classification model based on the BERT architecture is invoked to output the probability distribution of the question text in the dimensions of macroeconomics, policy analysis, industry analysis, financial health, business status, legal risk, public opinion monitoring, related risk, and early warning assessment. The dimension with the highest probability value is used as the main analysis dimension, while other dimensions with probability values ​​higher than the preset threshold are also included in the multi-dimensional analysis.

5. The intelligent question-answering method based on a large language model in post-loan risk management according to claim 1, characterized in that, In step S6, the recall process of the knowledge base recall module includes: parsing the document into a JSON format file and then converting it into a Markdown file; for documents with chapter, section, and article structures, segmenting them according to document hierarchy, starting from the article level; when an article length exceeds 1,000 characters, further segmenting it into natural paragraphs; when an article length is less than 50 characters, merging it with adjacent articles; for unstructured documents, using the Sentence-BERT model to calculate the cosine similarity of adjacent sentences, and using positions with similarity below 0.3 as segmentation points; after segmentation, using text2vec-large-chinese... The model vectorizes each text block into a 768-dimensional vector and stores it in a vector database. It extracts the three words with the highest TextRank value in the user question as keywords, performs keyword retrieval in the knowledge base, and calculates the hit density S=n / N for each text block, where n is the number of keyword hits in the text block and N is the number of characters in the text block. At the same time, it calculates the cosine similarity between the vectorized user question and the vector of each text block in the vector database. It substitutes the keyword retrieval ranking and vector recall ranking of each text block into R=Σ(1 / (k+rank)) to calculate the fusion similarity, where k is set to 60, and selects the text block with the highest R value as the recall result.

6. The intelligent question-answering method based on a large language model in post-loan risk management according to claim 1, characterized in that, In step S5, the professional analysis agents include macroeconomic agents, policy analysis agents, industry analysis agents, financial health agents, business status agents, judicial risk agents, public opinion monitoring agents, related risk agents, and early warning assessment agents. The intelligent routing queries a pre-configured key-value pair mapping table based on the analysis dimension corresponding to each sub-task. The key of the mapping table is the name of the analysis dimension, and the value is the name of the corresponding professional analysis agent. The sub-tasks are then assigned to the target agents. If a certain dimension does not have a corresponding agent in the mapping table, the sub-task is routed to the backup agent for processing.

7. The intelligent question-answering method based on a large language model in post-loan risk management according to claim 6, characterized in that, The financial health intelligence agent obtains the fields of total assets, total liabilities, current assets, current liabilities, inventory, accounts receivable, operating revenue, operating costs, net profit, earnings before interest and taxes (EBIT), interest expense, and net cash flow from operating activities through the financial statement data interface. It then calculates the following in sequence: debt-to-equity ratio = total liabilities / total assets, current ratio = current assets / current liabilities, quick ratio = (current assets - inventory) / current liabilities, interest coverage ratio = EBIT / interest expense, return on equity = net profit / average shareholders' equity, total asset turnover = operating revenue / average total assets, and revenue growth rate = (current period operating revenue - previous period operating revenue) / previous period operating revenue. The calculated values ​​of each indicator are then compared one by one with the 90ths, 70ths, 50ths, 30ths, and 10ths of the published enterprise performance evaluation standards for the corresponding industry and size type to determine the quintile level of each indicator. The judicial risk intelligence agent obtains data from the judicial dimension interface for the past year... The system collects the defendant's case list, the enforcement case list, records of restrictions on high consumption, and records of dishonest judgment debtors. It calculates the number of cases involved (Score_litigation) and the enforcement scale (Score_execution). When Score_litigation is greater than the industry average plus two standard deviations and the total amount of litigation targets is greater than 15% of the company's net assets, or Score_execution is greater than 10% of the company's net assets and the absolute value is greater than 5 million yuan, or there is a record of restrictions on high consumption, it is judged as high risk. The associated risk intelligence obtains the external investment list, the purchase ratio of the top three suppliers (SC), the sales ratio of the top three customers (CC), the total amount of external guarantees (GR), and the equity pledge ratio of the actual controller (PR) through the associated dimension interface. When GR is greater than 40%, it is marked as an excessively high guarantee ratio; when CC is greater than 75%, it is marked as a highly concentrated customer base; and when PR is greater than 70%, it is marked as a high pledge risk.

8. The intelligent question-answering method based on a large language model in post-loan risk management according to claim 1, characterized in that, In step S8, the progressive streaming output method uses SSE technology. The front end initiates an HTTP long connection request through the EventSource API. The back end encapsulates the data of each stage into SSE events marked with event types summary, dimension, and final and pushes them to the front end one by one. Each event is transmitted in the format of "event:" plus event type plus newline character plus "data:" plus event data plus newline character plus two newline characters. The front end receives the event through the onmessage callback function and immediately triggers rendering. The first stage pushes a summary of the intelligent agent's generated summary information of less than 200 words. The second stage follows the dimensions of macroeconomics, industry analysis, policy analysis, financial health, operating status, judicial risk, public opinion monitoring, related risks, and early warning assessment. After the content of each dimension is generated, the dimension is encapsulated into a dimension event and pushed. The third stage pushes a final event containing comprehensive risk rating and disposal suggestions. When the length of the accumulated output content exceeds 70% of the length of the large language model context window of 8000 tokens, the dimension content that has been stably output is removed from the context window and cached in the session storage.

9. The intelligent question-answering method based on a large language model in post-loan risk management according to claim 1, characterized in that, In step S1, the specific method for feature selection using the XGBoost algorithm is as follows: The maximum tree depth of the XGBoost model is set to six, the learning rate to 0.05, the number of estimators to 150, the minimum leaf node sample weight to 0.1, the column sampling ratio to 0.8, and the subsampling ratio to 0.

8. Five-fold cross-validation is used, and the number of early stopping rounds is set to twenty. The average information gain of each feature is calculated using the Gain metric, and the feature importance threshold is set to 0.

01. Features with an average gain lower than 0.01 are removed. The specific method for indicator selection using WOE and IV value analysis is as follows: Continuous indicators are divided into ten bins and discrete indicators into five bins using the equal-frequency binning method, and the calculation... IV value: Indicators with IV values ​​greater than 0.02 are retained. The Pearson correlation coefficient between the retained indicators is calculated. For strongly correlated indicator pairs with an absolute correlation coefficient greater than 0.7, the one with the higher IV value is retained. The specific method of constructing the scorecard model using logistic regression is as follows: using overdue for more than 90 days as the definition of default, using the screened indicators as feature variables, using L2 regularization with a regularization coefficient C=1, and training the model with the goal of maximizing the AUC value of the test set. After the model outputs the default probability, the default probability is mapped to a percentage score using the log-linear transformation formula: score = baseline score - weight factor × ln(default probability). The baseline score is set to 600 points, corresponding to a default probability of 5%, and the weight factor is set to 40%.

10. An intelligent question-answering system based on a large language model for post-loan risk management, characterized in that: include: The data acquisition module is used to build a unified access and standardized indicator processing pipeline for multi-source heterogeneous data, acquire multi-source raw data from within and outside the industry, clean, align and fuse it, and process it into multi-dimensional standardized indicators; it is also used to perform feature screening and indicator screening based on the standardized indicators, build a scorecard model to generate user scores, and perform event-based risk judgment on the standardized indicators to generate risk signals. The early warning judgment module is used to make a deterministic risk judgment on the standardized indicators and the user scores based on preset indicator thresholds and scoring segments through an early warning rule engine, and output a structured early warning result. The data interface module is used to establish a query database, store the standardized indicators, user scores, risk signals and early warning results into the query database, and encapsulate them into a structured data interface for use by a large language model based on the risk dimension in the query database. The intent parsing module is used to obtain the user's input question, perform intent recognition on the question through the question splitting agent, parse out the analysis subject and analysis dimension of the question, and split the question into at least one sub-task according to the analysis dimension. The intelligent routing module is used to allocate each subtask to the corresponding professional analysis agent based on the analysis dimension corresponding to each subtask. The professional analysis agent is predefined based on the prompt word engineering to define its professional role positioning, core analysis framework and behavioral boundary constraints. The behavioral boundary constraints include data usage constraints, which restrict all analysis to be based on the enterprise data returned by the structured data interface and prohibit the large language model from fabricating data on its own. The data assembly module is used by each of the professional analysis agents to call the structured data interface to obtain the enterprise data corresponding to the analysis subject. At the same time, the knowledge base retrieval module retrieves professional knowledge documents related to the sub-task from the knowledge base. The retrieved professional knowledge documents are then concatenated with the enterprise data and the prompt words corresponding to the professional analysis agent to form the complete input of the professional analysis agent. The parallel analysis module is used by each of the aforementioned specialized analysis agents to call the large language model in parallel, sending the complete input into the large language model, so that the large language model generates structured analysis results of the corresponding dimension under the behavioral boundary constraints of the prompt words; The results aggregation module is used to receive the structured analysis results from each intelligent agent through comprehensive summarization, and aggregate them into a unified structured comprehensive evaluation report according to the preset report structure, and present them to the user in stages according to the progressive streaming output method.

Images