Method and system for large-model illusion governance and operation and maintenance through cooperation of multiple models in financial field

Abstract

The invention discloses a financial field multi-model collaborative large-model illusion governance and operation and maintenance method and system, and the method comprises the steps: firstly carrying out the type recognition and feature extraction of a received user query, obtaining a type feature, and obtaining a corresponding routing tag and constraint configuration based on the type feature; then selecting a processing path according to the routing label; then, according to a routing label, a retrieval result and constraint configuration, a prompt template is constructed, an evidence fragment is embedded, and format requirements and mathematical strong constraints are referenced; then, in combination with RAG retrieval evidence, a calculation strategy result and a constructed prompt template, a preset large model is called for reasoning; evaluating the output reasoning result by adopting a multi-layer evaluation strategy; and finally carrying out risk grading and strategy execution according to an evaluation result. According to the invention, the illusion of the large model can be detected and evaluated in real time.

Description

Methods and systems for governance and operation of multi-model collaborative large-scale model illusion in the financial field Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing technology, and more specifically, to a method and system for the governance and operation of a large-scale collaborative multi-model illusion in the financial field. Background Technology

[0002] Operations and maintenance in the financial securities industry is a core element in ensuring the stable, secure, and efficient operation of trading systems. Faced with stringent requirements such as high-frequency trading, strong regulation, and 24 / 7 uninterrupted service, the operations and maintenance system has gradually evolved from traditional manual on-duty to a modern model that is automated, intelligent, and fully visualized.

[0003] Through extensive research and practice, the inventors of this application have found that the current AI operation and maintenance in the financial securities field mainly adopts a combination of traditional rule engines and manual review, which has the following main technical problems: (1) The hallucination detection is lagging behind, and there is a lack of real-time hallucination assessment mechanism. Errors are often only discovered after they have caused an impact. This is because existing methods or systems rely on post-event manual review and customer complaints to trigger checks. They cannot assess the hallucination risk in real time before LLM output and lack real-time assessment mechanisms such as consistency scoring, coverage checks, and feature detection. As a result, the hallucination output is only discovered after it has reached the customer.

[0004] (2) Insufficient retrieval coverage: Traditional vector retrieval lacks knowledge graph enhancement and has a low evidence recall rate, causing LLM to still force output even when the evidence is insufficient. Pure vector retrieval cannot handle complex relationships between entities and lacks the multi-hop traversal capability of knowledge graphs, resulting in a low evidence recall rate. As a result, it cannot achieve a complete mapping of entity relationships, and LLM generates incorrect conclusions based on incomplete evidence.

[0005] (3) The operation and maintenance response is passive, lacking an intelligent anomaly linkage mechanism. Fault response relies on manual investigation, resulting in a long average repair time. In existing methods, alarms rely on fixed threshold rules and lack intelligent anomaly detection based on GNN+time series generation models. For example, there is a lack of learning system behavior patterns and early warning, as well as a lack of natural language alarm analysis, such as automatically generating root causes of faults, scope of impact, and suggested solutions. In addition, there is a lack of anomaly linkage mechanisms, such as automatically triggering knowledge base governance and model rollback. Therefore, the average repair time is long and the scope of fault impact is expanded. Summary of the Invention

[0006] To address the technical problems existing in the prior art, this invention proposes a method and system for the governance and operation of multi-model collaborative large-scale model illusion in the financial field.

[0007] Specifically, the first aspect of this invention provides a method for governance and operation of multi-model collaborative large-scale model illusion in the financial field, comprising: receiving input user queries; performing type identification and feature extraction on the received user queries to obtain type features, and obtaining corresponding routing labels and constraint configurations based on the type features; selecting a processing path according to the routing labels, wherein the processing path includes a computational strategy path and a RAG retrieval enhancement path, the computational strategy path being used to call a pre-built computational strategy engine to generate computational strategy results, and the RAG retrieval enhancement path being used to recall relevant evidence fragments from the knowledge base, calculate retrieval coverage, and select whether to add the recalled evidence fragments to the context based on the retrieval coverage; constructing a prompt template based on the routing labels, retrieval results, and constraint configurations, embedding evidence fragments, citation format requirements, and strong mathematical constraints; combining RAG retrieval evidence, computational strategy results, and the constructed prompt template, calling a preset large-scale model for inference; evaluating the output inference results using a multi-layer evaluation strategy; and performing risk classification and strategy execution based on the evaluation results.

[0008] In one implementation, the received user query is subjected to type identification and feature extraction to obtain type features, including: identifying the type of the received user query to obtain the question intent category, including fact query, creation, mathematical calculation, logical reasoning and document writing; performing preset keyword detection on the user query to obtain risk markers; and statistically analyzing the frequency of numbers and the density of numerical values ​​in the user query to obtain frequency analysis results.

[0009] In one implementation, the method further includes: identifying numerically intensive problems based on frequency analysis results.

[0010] In one implementation, the RAG retrieval enhancement path is specifically used for: loading documents from the knowledge base and cutting the documents into slices of a preset length; semantically embedding each slice using an embedding model to construct a vector database; converting user queries into vectors using an embedding model; performing vector retrieval based on the similarity between the user query vector and the vectors in the vector database; constructing a knowledge graph from the knowledge base; embedding the user query; performing nearest neighbor search or graph traversal to identify nodes related to the query embedding from the knowledge graph; and obtaining evidence fragments based on the results of vector retrieval and graph retrieval.

[0011] In one implementation, the system combines RAG retrieval evidence, computational strategy results, and constructed prompt templates to invoke a pre-defined large model for reasoning. This includes: for user queries of mathematical calculation type, the pre-defined large model is responsible for interpreting the computation steps and using the computational strategy results generated by the computational strategy path as the numerical results obtained from the reasoning; for user queries of other types besides mathematical calculation type, the pre-defined large model uses a thought chain or thought tree to expand the logic, and the reasoning results include answer text, evidence citations, and reasoning chains.

[0012] In one implementation, a multi-layered evaluation strategy is used to evaluate the output reasoning results, including: performing a reasonableness analysis on the characteristic statements and numerical ranges in the reasoning results; analyzing the semantic consistency between the reasoning results and the retrieved evidence, and calculating a consistency score; tracing the evidence chain to verify the integrity of the reasoning chain, and calculating the citation rate to represent the proportion of evidence cited in the answer; and using the consistency score, retrieval coverage, and citation rate as a comprehensive evaluation result.

[0013] In one implementation, risk classification and strategy execution are performed based on the assessment results, including: determining whether the risk meets the first preset condition as a high risk based on the risk marker; if so, manual review or verification is performed; otherwise, the data is released and marked as referenced; comparing the calculation strategy result generated by the calculation strategy path with the output of the preset large model; if the values ​​are inconsistent, interception and recalculation are triggered; determining whether the retrieval coverage is less than the coverage threshold; if so, downgrading is performed and retry is performed or an insufficient evidence prompt is returned.

[0014] In one embodiment, the method further includes collecting and monitoring data during the input phase, reasoning phase, evaluation phase, and decision-making phase.

[0015] Based on the same inventive concept, a second aspect of this invention provides a system for the governance and operation of multi-model collaborative large-scale model illusions in the financial field, comprising: a user query receiving module for receiving input user queries; a problem analysis and type identification module for performing type identification and feature extraction on the received user queries to obtain type features, and obtaining corresponding routing labels and constraint configurations based on the type features; and a processing path selection module for selecting a processing path based on the routing labels, wherein the processing path includes a computational strategy path and a RAG retrieval enhancement path, the computational strategy path being used to call a pre-built computational strategy engine to generate computational strategy results, and the RAG... The retrieval enhancement path is used to retrieve relevant evidence fragments from the knowledge base, calculate retrieval coverage, and select whether to add the retrieved evidence fragments to the context based on the retrieval coverage; the prompt template construction module is used to construct prompt templates based on routing tags, retrieval results, and constraint configurations, embedding evidence fragments, citation format requirements, and strong mathematical constraints; the reasoning module is used to combine RAG retrieval evidence, calculation strategy results, and constructed prompt templates to call a preset large model for reasoning; the evaluation and decision-making module is used to evaluate the output reasoning results using a multi-layer evaluation strategy; and the strategy execution and risk classification module is used to classify risks and execute strategies based on the evaluation results.

[0016] Based on the same inventive concept, a third aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method for governance and operation of multi-model collaborative large-scale model illusion in the financial field as described in the first aspect.

[0017] Compared to existing technologies, the advantages and beneficial technical effects of this invention are as follows: This invention discloses a method for governance and operation of illusions in a large-scale collaborative multi-model system in the financial field. First, it identifies the type and extracts features from received user queries to obtain type features, and then obtains corresponding routing labels and constraint configurations based on these features. Next, it selects a processing path based on the routing labels, search results, and constraint configurations, and constructs a prompt template, embedding evidence fragments, citation format requirements, and strong mathematical constraints. Then, it combines RAG retrieval evidence, calculation strategy results, and the constructed prompt template to call a preset large-scale model for inference. Next, it uses a multi-level evaluation strategy to evaluate the output inference results. Finally, it performs risk classification and strategy execution based on the evaluation results. Because this invention selects whether to add recalled evidence fragments to the context based on the calculated search coverage when triggering the RAG retrieval enhancement path, it ensures that the output is only performed when the search coverage meets the conditions, thus ensuring sufficient evidence. Furthermore, it performs real-time evaluation of the output inference results based on a multi-level evaluation strategy, thereby enabling real-time detection and evaluation of illusions in the large-scale model.

[0018] Furthermore, the RAG retrieval enhancement path combines vector retrieval with graph-enhanced retrieval to reduce the illusion rate of the model.

[0019] Furthermore, the computational strategy results generated by the computational strategy path are compared with the output of the preset large model. If the values ​​are inconsistent, an interception and recalculation are triggered to ensure the accuracy of the numerical calculation.

[0020] Furthermore, data from the input, reasoning, evaluation, and decision-making stages are collected and monitored to form a complete observable operation and maintenance and emergency response system. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 is a flowchart of the method for governance and operation of multi-model collaborative large-scale model illusion in the financial field in an embodiment of the present invention; Figure 2 is a detailed technical roadmap of the method for governance and operation of multi-model collaborative large-scale model illusion in the financial field in an embodiment of the present invention; Figure 3 is a flowchart of inference result evaluation and strategy execution in an embodiment of the present invention; Figure 4 is a flowchart of intelligent anomaly detection based on GNN+ temporal generation model in an embodiment of the present invention; Figure 5 is a system architecture diagram of governance and operation of multi-model collaborative large-scale model illusion in the financial field in an embodiment of the present invention. Detailed Implementation

[0023] This embodiment provides a method for governance and operation of a large-scale collaborative multi-model illusion in the financial field. Please refer to Figure 1. It includes: S1: Receiving input user queries; specifically, S1 is the input data receiving stage.

[0024] S2: Perform type identification and feature extraction on the received user queries to obtain type features, and obtain the corresponding routing labels and constraint configurations based on the type features.

[0025] Specifically, S2 is the problem type analysis. After a user query enters through the entry gateway, the problem type is identified, and type characteristics are obtained, including fact query, content creation, mathematical calculation, logical reasoning, and document writing. Then, based on the type characteristics, the corresponding route tags and constraint configurations are obtained.

[0026] Specifically, S2 performs type identification and feature extraction on the received user queries to obtain type features, including: identifying the type of the received user queries to obtain the question intent category, including fact query, creation, mathematical calculation, logical reasoning and document writing; performing preset keyword detection on the user queries to obtain risk markers; and statistically analyzing the frequency of numbers and the density of numerical values ​​in the user queries to obtain frequency analysis results.

[0027] The method also includes identifying numerically intensive problems based on frequency analysis results.

[0028] Specifically, type features are used to characterize the fine-grained attributes of the query itself and its context, serving as input evidence for routing decisions and constraint configuration. Type features include at least: question intent category, specific keyword detection results (e.g., triggering a high-risk flag when high-risk words such as "prediction" or "estimation" are included), frequency and density of statistical data (used to identify numerically intensive questions and trigger "strong constraints on computational strategies"), and the risk level derived from these.

[0029] When the numerical values ​​appear frequently in the problem (such as containing multiple percentages, amounts, dates, and other numerical entities), it is determined to be a numerically intensive problem. At this time, strong constraints on the calculation strategy will be triggered, specifically: (1) When processing path selection, the pre-built calculation strategy engine is forced to generate deterministic values; (2) During reasoning, LLM is only responsible for language interpretation and is prohibited from generating values ​​on its own; (3) During evaluation, the output of the calculation strategy is compared with the output of LLM. If the values ​​are inconsistent, interception and recalculation will be triggered.

[0030] Specifically, the aforementioned type features are first generated through type recognition and feature extraction. Then, based on these type features, corresponding routing labels and constraint configurations are inferred (e.g., different question types are bound to different search coverage thresholds and prompt template constraints). High-risk markers and other type features are then passed to the subsequent Prompt factory and illusion evaluation layer to select appropriate prompt templates, tighten SLO thresholds, and determine whether to proceed to the human review process. Thus, type features are the foundation for generating routing labels and constraint strategies, while routing labels represent high-level decision-making results formed based on these type features.

[0031] Routing labels are used to abstractly identify queries in terms of business scenarios and technical processing paths. They are the core control signals for the system to select retrieval paths, computational strategy branches, multi-model candidate pools, and configure evaluation and guard strategies. Based on intent recognition results, the question type analysis module categorizes queries into types such as fact queries, content creation, mathematical computation, logical reasoning, or long-form writing, and generates corresponding routing labels accordingly. For example, for the "Operations + Business Integration Diagnosis" scenario, it generates the routing label `fact_query_ops_biz`. The routing labels directly determine whether the second layer follows the RAG retrieval path or a pre-built computational strategy path, which retrieval coverage threshold is used, and whether to prioritize compliant models, mathematically specialized models, or general-purpose generative models.

[0032] In the specific implementation process, the output route label and type characteristics play a key constraining role in the subsequent steps, specifically including: (1) The route label determines the selection of the processing path of S2. For example, the mathematical calculation type will trigger the calculation strategy path and force the call of the deterministic calculation engine; the fact query type will trigger the RAG retrieval enhancement path, which adopts a strict retrieval coverage threshold, such as a retrieval coverage threshold of 0.65; the creation type will trigger the RAG retrieval enhancement path, which adopts a more relaxed coverage threshold, such as a retrieval coverage threshold of 0.5.

[0033] (2) Type characteristics affect the template selection of the Prompt factory. For mathematical problems, the constraint of calculation before expression is forced to be embedded. For factual problems, the citation format is forced to be embedded.

[0034] (3) Risk markers will be used for subsequent risk assessments. For example, high-risk markers will be subject to stricter SLO thresholds at the assessment level and may be subject to mandatory human review processes.

[0035] S3: Select the processing path based on the routing label. The processing path includes the computation strategy path and the RAG retrieval enhancement path. The computation strategy path is used to call the pre-built computation strategy engine to generate computation strategy results. The RAG retrieval enhancement path is used to retrieve relevant evidence fragments from the knowledge base, calculate the retrieval coverage, and select whether to add the retrieved evidence fragments to the context based on the retrieval coverage.

[0036] Specifically, S3 is path selection, where the problem type analysis results are split into two parallel paths: the computation strategy path and the RAG retrieval enhancement path.

[0037] Specifically, the RAG retrieval enhancement path is used for: S3.1: Loading documents from the knowledge base and cutting them into slices of a preset length; S3.2: Using an embedding model to perform semantic embedding on each slice, constructing a vector database, converting user queries into vectors using the embedding model, and performing vector retrieval based on the similarity between the user query vector and the vectors in the vector database; S3.3: Constructing the knowledge base into a knowledge graph, embedding user queries, and performing nearest neighbor search or graph traversal to identify nodes related to the query embedding from the knowledge graph; S3.4: Obtaining evidence fragments based on the results of vector retrieval and graph retrieval.

[0038] Specifically, the RAG retrieval path is generally divided into document loading, text segmentation, embedding vectorization, and retrieval. This process combines vector retrieval with graph retrieval, specifically GraphRAG graph-enhanced retrieval. This retrieval process includes entity linking, 2-3 hop multi-hop traversal of the knowledge graph, temporal filtering, and graph statistical evidence fusion. When obtaining evidence fragments based on the results of vector and graph retrieval, a hybrid sorting method is used, for example, setting the weight of graph retrieval to 0.6 and the weight of vector retrieval to 0.4. Compared to pure vector retrieval, this hybrid retrieval method significantly reduces the illusion rate and substantially decreases token consumption, ultimately returning coverage scores and evidence fragments.

[0039] The knowledge graph is built offline before system deployment and continuously updated through the knowledge base governance module. It includes the following: (1) Financial entity nodes, including listed companies, regulatory agencies, financial products, key figures, etc.; (2) Relationship edges, shareholding structure, related transactions, regulatory chains, industry affiliation, etc.; (3) Time attributes, each node and relationship is marked with an expiration date, and temporal filtering is supported. Main function: It stores complex relationships between entities through a graph structure, supports multi-hop reasoning and related queries, and makes up for the shortcomings of pure vector retrieval in relational reasoning.

[0040] Multi-hop traversal refers to starting from the query entity and performing 2-3 hops along the relationship edges of the knowledge graph to retrieve relevant evidence. For example, to query "Does Company A have related-party transactions with Company B?", the process starts from Company A → (1 hop) finds Company A's subsidiary C → (2 hops) discovers that C has an equity relationship with Company B → (3 hops) confirms the related-party transaction path. It should be noted that 2-3 hops is an empirical setting, which can capture indirect relationships while avoiding over-diffuse noise.

[0041] Domain graphs are financial specialization subsets of knowledge graphs. They are built offline based on authoritative materials such as regulatory documents, financial statements, and research reports before system deployment and are updated regularly through the knowledge base governance module. Specific functions include: (1) providing standardized definitions and associations for financial professional terms; (2) supporting entity disambiguation—distinguishing similar company names and stock codes; and (3) supporting temporal reasoning—verifying the timeliness of evidence and filtering outdated data.

[0042] Regarding the computational strategy path, mathematical problems directly call the deterministic computation engine to generate recalculated numerical results, avoiding the LLM's reliance on intuition to output numbers.

[0043] S4: Construct a prompt template based on the routing label, search results and constraint configuration, and embed evidence fragments, citation format requirements and strong mathematical constraints.

[0044] Specifically, S4 is the prompt template construction. The Prompt factory constructs a prompt template based on the question type and search results, embedding evidence fragments, citation format requirements (such as source, page number or paragraph), and strong mathematical constraints (calculate before stating, prohibit certain conclusions without evidence, etc.).

[0045] S5: Combining RAG retrieval evidence, computational strategy results, and constructed prompt templates, the system invokes a pre-defined large model for inference.

[0046] Specifically, S5 stands for model inference.

[0047] S5 specifically includes: S5.1: For user queries of mathematical calculation type, the preset large model is responsible for interpreting the calculation steps and using the calculation strategy results generated by the calculation strategy path as the numerical results obtained from the reasoning; S5.2: For user queries of other types besides mathematical calculation type, the preset large model uses the thinking chain or thinking tree to expand the logic, and the reasoning results include answer text, evidence references and reasoning chain.

[0048] In the specific implementation process, the results output by LLM after performing inference include: (1) Answer text: the answer content generated for the user query; (2) Evidence citation: the source of evidence retrieved by RAG (e.g., document, page number, or paragraph); (3) Inference chain: the logical step sequence of CoT / ToT expansion; (4) Numerical results: if it is a mathematical problem, it includes the deterministic numerical value output by the calculation strategy engine and the language explanation of LLM. These results output after inference are passed as a whole to step S6 for quality inspection.

[0049] S6: Employ a multi-level evaluation strategy to evaluate the output reasoning results.

[0050] Specifically, S6 is the hallucination assessment.

[0051] S6 specifically includes: S6.1: Analyzing the rationality of characteristic statements and numerical ranges in the reasoning results; S6.2: Analyzing the semantic consistency between the reasoning results and the retrieved evidence, and calculating the consistency score; S6.3: Tracing the evidence chain, verifying the integrity of the reasoning chain, and calculating the citation rate to represent the proportion of evidence cited in the answer; S6.4: Using the consistency score, retrieval coverage, and citation rate as a comprehensive evaluation result.

[0052] In the specific implementation process, the LLM output enters the automatic evaluator, which adopts a four-layer evaluation pipeline, as shown in Figure 3.

[0053] The first layer is fast rule filtering, which detects the reasonableness of feature words and numerical ranges; the second layer is semantic consistency, which calculates the coverage and factual consistency of retrieved evidence; the third layer is CoT reasoning verification, which traces evidence step by step to verify the integrity of the reasoning chain; the fourth layer is machine learning model (ML model) scoring, where the ensemble classifier outputs a comprehensive score that includes consistency score, coverage, and citation rate.

[0054] S7: Based on the assessment results, classify risks and implement strategies.

[0055] Specifically, S7 refers to risk classification and strategy execution.

[0056] S7 specifically includes: S7.1: Determine whether it is a high-risk condition that meets the first preset condition based on the risk marker. If so, conduct manual review or verification. Otherwise, allow it and mark it as a reference. S7.2: Compare the calculation strategy result generated by the calculation strategy path with the output of the preset large model. If the values ​​are inconsistent, trigger interception and recalculation. S7.3: Determine whether the search coverage is less than the coverage threshold. If it is less, perform downgrade processing and retry or return an insufficient evidence prompt.

[0057] In the specific implementation process, the overall score is checked for SLO (Service Level Objective) thresholds. For example, if the consistency score, retrieval coverage, and citation rate all meet the conditions, then it is verified whether the current prompt model version is in the grayscale whitelist.

[0058] Then, a hybrid detection combining rules and machine learning models (ML) is performed to conduct risk grading and fine-grained illusion type identification. The specific workflow is as follows: (1) Rule detection layer: First, predefined rules are applied for rapid screening, including feature detection, numerical range rationality verification, and timestamp consistency check; (2) ML detection layer: For the output passing through the rule layer, a trained ensemble classifier is used for deep detection to identify subtle illusion patterns that are difficult to capture by rules; (3) Risk grading method: Combining consistency score, coverage, citation rate, illusion type and business scenario, the output is divided into four risk levels - low risk, insufficient evidence, numerical inconsistency, and high risk; (4) Fine-grained illusion type identification: The HalluFin framework is applied to identify four types of illusions: time drift, projection overgeneralization, numerical inconsistency and entity confusion, providing accurate guidance for subsequent traffic splitting.

[0059] Specifically, please refer to Figure 3. The specific decisions for the four branches of risk level are as follows: Insufficient evidence branch: If the search coverage is less than the threshold, downgrade, retry, or expand the recall; Inconsistent values ​​branch: If the calculation strategy result is inconsistent with the LLM output value, the calculation strategy will be forced to recalculate, and the system will be blocked and prompted "There is a deviation in the numerical calculation. The result of the calculation strategy shall prevail".

[0060] For low-risk branches: if consistency and coverage meet the standards and references are in compliance with regulations, the branches are allowed to proceed and the reference sources are marked. Then, auditing and record-keeping are performed, and logs are retained throughout the entire process.

[0061] High-risk branches: those with insufficient consistency, high-risk characteristic words, or sensitive areas will enter the human review and compliance guard.

[0062] In one embodiment, the method further includes collecting and monitoring data during the input phase, reasoning phase, evaluation phase, and decision-making phase.

[0063] In the specific implementation process, the collected data fields include: prompt_id: a unique identifier for the prompt word, used to trace the Prompt version and historical changes; retrieval_k: the number of top-k evidence fragments returned by RAG retrieval; coverage_ratio: retrieval coverage, measuring the degree to which the RAG returned fragments hit the question; consistency_score: fact consistency score, measuring the semantic consistency between the LLM output and the retrieved evidence; citation_rate: citation rate, measuring the proportion and normativity of evidence citations in the answer; hallucination_score: hallucination score, comprehensively assessing the hallucination risk level of the output; temporal_drift_count: temporal drift count, recording the number of times outdated data is cited in the output; graph_hop_depth: graph traversal hop count, recording the multi-hop depth during GraphRAG retrieval. This implementation method constructs an operation and maintenance monitoring system, as shown in Figure 5. The right side shows the operation and maintenance module group, which includes five layers of modules from top to bottom: observable data collection, SLO threshold monitoring, monitoring alarms and auditing, cost and performance governance, and scaling and disaster recovery, forming a complete observable operation and maintenance and emergency response system.

[0064] The operation and maintenance monitoring system consists of the following components from top to bottom: observable data collection, SLO threshold monitoring, monitoring alarms and auditing, cost and performance management, and scaling and disaster recovery.

[0065] The method of this invention supports three types of models, all of which are obtained by retraining and optimizing existing large language models for different specific domains. The system dynamically selects the appropriate model based on the needs of the scenario: Compliance Model: A model fine-tuned for the compliance domain, suitable for scenarios such as interpreting regulatory policies and conducting compliance reviews, emphasizing the compliance and conservatism of the output; Mathematical Expertise Model: A model that excels in mathematical reasoning, suitable for scenarios such as calculating and interpreting risk indicators and analyzing financial data, and is used in conjunction with a computational strategy engine; General Generative Model: A general large language model, suitable for generative scenarios such as research report generation and customer service.

[0066] The model is dynamically selected based on the following factors: (1) Problem type: The routing label identified in step S2 determines the candidate model pool. For example, the "compliance and supervision" type prioritizes the compliance model, and the "mathematical calculation" type prioritizes the mathematical expertise model; (2) Latency budget: The model is selected based on the latency requirements of the business scenario. For example, in real-time transaction scenarios, a lightweight model with faster inference speed is selected if a response is required in a short time. In non-real-time scenarios, a larger and more accurate model can be selected; (3) Risk level: In high-risk scenarios, the compliance model is forced to be used and the human review process is triggered.

[0067] The method of the present invention will be described in detail below through a specific example.

[0068] A securities firm's compliance department discovered that this morning, the investment advisory service system triggered a large number of downgrade strategies for investment recommendations from VIP clients. The compliance manager, in conjunction with the operations team, inquired: "Why are VIP client investment recommendations frequently triggering downgrades? Is it a system malfunction or normal blocking? Please provide evidence and assess the business impact." Step 1: Input Question and Type Identification. The compliance manager's query entered the system through the entry gateway. Through the question type analysis module, the keywords "investment recommendations," "downgrade," "normal blocking," "VIP clients," and "business impact" were identified. It was determined to be a fact query type, and the output routing label was: fact_query_ops_biz.

[0069] Step 2: Specific keyword detection and digital frequency statistics. The keywords "investment advice" and "VIP customer" were detected and marked as high risk; "downgrade" and "normal interception" triggered the operation and maintenance scenario marking, which was determined to be an operation and maintenance and business integration scenario. At this time, it is necessary to retrieve operation and maintenance data and business data at the same time.

[0070] Step 3: RAG Retrieval and Coverage Check. Initiate the multi-source RAG retrieval process: Operations dimension retrieval: The monitoring system retrieves evaluation records, degradation trigger logs, and SLO indicator changes for the large model system illusion detection during the period from 09:30 to 11:00 today; Business dimension retrieval: Based on the service knowledge base, obtain VIP customer classification standards, investment advisory service SLA requirements, and investment advice compliance red lines; Strategy dimension retrieval: Based on the strategy configuration library, obtain the retrieval coverage threshold, consistency threshold, and degradation rule temporal filtering for investment advice scenarios, retain only the data for the preset time period (e.g., 09:30 to 11:00 today), and return the retrieval coverage. Add evidence fragments based on the relationship between retrieval coverage and thresholds to continue.

[0071] The evidence retrieved includes: Operational evidence: The downgrade rate in the investment advisory scenario increased between 09:45 and 10:30, triggered by insufficient search coverage; Operational evidence: The knowledge base update event at 09:40—the expiration of the securities research report database TTL led to the offline of research report data; Business evidence: VIP customer service SLA requirements—investment advice must be accompanied by research report evidence, and affirmative conclusions are prohibited without evidence; Strategy evidence: The search coverage threshold for the investment advisory scenario is set to 0.80 (which is higher than the default value of 0.65), triggering downgrade.

[0072] Step 4: The mathematical problem determination system determines that the problem is not a mathematical problem, does not call the calculation strategy engine, and uses the CoT (Thinking Chain) method for multi-dimensional analysis and reasoning.

[0073] Step 5: Prompt Construction and LLM Inference. The Prompt factory is constructed based on the operation and maintenance + business integration template: embedding multi-dimensional RAG evidence and mandatory constraints: "Analyze from the operation and maintenance perspective and the business perspective respectively. All conclusions must be based on the retrieved evidence and clearly distinguish between system failures and normal interception."

[0074] The LLM inference service enables CoT inference, and the specific process includes: Step 1: Conduct operational analysis to confirm the cause of the downgrade, specifically "retrieval coverage < coverage threshold 0.80", rather than a system failure; Step 2: Perform root cause analysis, the research report database TTL (Time To Live) expired at 09:40, causing the research report data to be offline, and the retrieval evidence was insufficient; Step 3: Conduct business impact assessment, the VIP customer service SLA (Service Level Agreement) requires "no conclusion without evidence", and the downgrade strategy complies with compliance requirements; Step 4: Conclusion judgment: This is a normal downgrade behavior of the large model system, used to protect customers from compliance risks of unsupported recommendations; Step 5: Handling recommendation: Immediately update the research report database TTL configuration to restore retrieval coverage.

[0075] Step 6: The automated evaluator performs a four-layer evaluation pipeline: Layer 1: Fast rule filtering; if no feature words are detected, the conclusion is based on multi-source evidence; Layer 2: Semantic consistency check; the LLM output is compared with operation and maintenance logs and business rules, and each rule is verified; consistency_score=0.93; Layer 3: CoT inference verification; the integrity of the causal chain is verified, for example, TTL expiration → insufficient evidence → downgrade trigger; Layer 4: ML model scoring; the comprehensive score output evaluation result is obtained: coverage_ratio=0.72, consistency_score=0.93, citation_rate=0.96 (full-dimensional evidence citation), hallucination_score=0.02.

[0076] Step 7: Fact consistency check. Determine if the fact consistency score is greater than or equal to the fact consistency score threshold for the financial scenario, i.e., consistency_score=0.93>=0.90 (fact consistency score threshold for the financial scenario). If it is greater than or equal to, continue execution without triggering a downgrade or blocking branch.

[0077] Step 8: SLO Threshold Check. SLO threshold verification: First, factual consistency is determined. The factual consistency score is checked against the financial scenario factual consistency score threshold, which meets the requirement. Next, the median search coverage is checked against the search coverage threshold. In this implementation, the median search coverage is 0.72 < 0.80, which does not meet the requirement because the research report database TTL has expired, resulting in insufficient coverage. Finally, the citation rate is checked against the citation rate threshold. In this implementation, the citation rate is 0.96 ≥ 0.95, which meets the requirement. The number of temporal violations is 0, which meets the requirement.

[0078] It should be noted that the above is an operation and maintenance fault analysis query. The insufficient retrieval coverage is due to the problem being analyzed, not a quality issue with the current query, so the process should continue.

[0079] Step 9: Verify the current configuration of the version and grayscale control system: Prompt version = compliance-ops-v2.1.0 (compliance + operation and maintenance integrated template), model version = compliance-v1.3 (compliance-specific model), current time = 11:05 (within the trading session), therefore, use the stable version configuration.

[0080] Step 10: High-risk assessment and strategy execution risk classification: This query involves a compliance-sensitive area, but this query is an analysis of the fault rather than providing investment advice, and the evidence is sufficient, so it is judged as "low risk" and the release strategy is executed.

[0081] Policy execution: Allow output and automatically label the source of reference, such as operation and maintenance log timestamps, business SLA documents, and policy configuration versions.

[0082] Step 11: Audit Traceability The entire decision-making process is written into the blockchain audit log. The log is retained for 7-10 years to support compliance audit and operation and maintenance review. The specific content of the blockchain audit log includes: (1) compliance manager query original text and employee number; (2) multi-dimensional RAG retrieval evidence, including operation and maintenance log, business SLA, and strategy configuration; (3) LLM model output: including evaluation score and strategy decision; (4) timestamp: YYYY-MM-DD 11:05:18.892, end-to-end latency: 1,250ms; Step 12: Anomaly Detection and Change Freeze The anomaly detection module monitors in real time: if there is no anomaly in this query, the normal process continues. At the same time, the system detects the TTL expiration problem of the research report library. It triggers the strategy engine through Webhook (an event notification system implemented through HTTP callback mechanism, which can send data to a specified URL when a specific event occurs) and links the knowledge base governance module to generate an emergency update work order. The change is currently frozen during trading hours. However, this type of knowledge base TTL configuration update is an emergency repair exception and will be executed immediately after approval by the on-duty operations and maintenance personnel.

[0083] The final output system returns a report on the integration of operations and business.

[0084] Example 2, based on the same inventive concept, discloses a system for the governance and operation of a multi-model collaborative large-scale model illusion in the financial field. It includes: a user query receiving module for receiving input user queries; a problem analysis and type identification module for performing type identification and feature extraction on the received user queries to obtain type features, and obtaining corresponding routing labels and constraint configurations based on the type features; and a processing path selection module for selecting a processing path based on the routing labels. The processing path includes a computational strategy path and a RAG retrieval enhancement path. The computational strategy path is used to call a pre-built computational strategy engine to generate computational strategy results, and the RAG retrieval enhancement path... The enhanced path module retrieves relevant evidence fragments from the knowledge base, calculates retrieval coverage, and selects whether to add the retrieved evidence fragments to the context based on the retrieval coverage. The prompt template construction module constructs prompt templates based on routing tags, retrieval results, and constraint configurations, embedding evidence fragments, citation format requirements, and strong mathematical constraints. The reasoning module combines RAG retrieval evidence, calculated strategy results, and constructed prompt templates to call a preset large model for reasoning. The evaluation and decision-making module evaluates the output reasoning results using a multi-layer evaluation strategy. The strategy execution and risk grading module performs risk grading and strategy execution based on the evaluation results.

[0085] Please refer to Figure 2, which is a detailed technical roadmap of the method for governance and operation of multi-model collaborative large model illusion in the financial field in this embodiment of the invention. The complete process from input question to final decision includes the following steps: 1. Input question reception: The user query enters the system through the entry gateway, and the system begins to analyze the question.

[0086] 2. Type Identification Stage: The problem analysis and type identification module identifies which category the query belongs to: fact query, content generation, mathematical calculation, logical reasoning, or long article writing. At the same time, it performs specific keyword detection and statistical frequency analysis, and finally outputs routing labels and type characteristics to provide guidance for subsequent processing.

[0087] 3. RAG Retrieval and Coverage Check: The system initiates the RAG retrieval process, retrieves relevant evidence fragments from the knowledge base, and calculates the retrieval coverage. If the retrieval coverage reaches a threshold, the evidence fragment is added to the context; if the coverage is insufficient, the system triggers a degradation strategy, including rewriting the query, expanding the recall scope, or switching to a stronger retrieval model to ensure the sufficiency of the evidence.

[0088] 4. Prompt Construction Phase: The system selects the corresponding prompt template based on the question type, embeds the evidence fragments obtained from RAG retrieval into the prompt, and adds citation format requirements. For mathematical problems, the prompt is required to "calculate before stating," and guessing values ​​without evidence is prohibited.

[0089] 5. LLM Reasoning and Computation Strategy Branch: The system sends the constructed hints to the LLM for reasoning. If the problem is identified as a mathematical problem, the system forcibly calls the computation strategy engine to generate deterministic values, while the LLM is only responsible for verbal interpretation to avoid numerical illusions; if the problem is a logical reasoning problem, a thought chain or thought tree is used to proceed with step-by-step reasoning.

[0090] 6. Evaluation and Scoring Process: Immediately after the LLM output, the system enters the automatic evaluator, executing a four-layer evaluation pipeline. The first layer uses rapid rule filtering to detect the reasonableness of feature words and numerical values; the second layer performs semantic consistency evaluation by verifying sentence embeddings and SPO triples (relation, entity relation, entity attribute triples) to calculate the consistency score; the third layer uses CoT inference to verify the completeness of the evidence chain through step-by-step tracing; and the fourth layer uses the ML model to score and output a comprehensive evaluation score. The system calculates three core indicators: consistency score, coverage, and citation rate.

[0091] 7. SLO Threshold Check Decision Point: The system compares the evaluation score with the SLO threshold. For example, in a financial scenario, the requirements are: factual consistency ≥ factual consistency threshold, median search coverage ≥ coverage threshold, and citation rate ≥ citation rate threshold. If any metric fails to meet the standard, the system triggers a downgrade or branches to be blocked; if all criteria are met, the subsequent process continues.

[0092] 8. Version and Gray-Scale Control Verification: The system verifies whether the currently used prompt template version and LLM model version are within the gray-scale whitelist and checks whether it is in a change freeze period (such as trading hours 09:30-15:00). If it is in a freeze period or the version is not on the whitelist, the new strategy execution is rejected, and a stable version configuration is used.

[0093] 9. Strategy Execution and Risk Classification: The strategy engine makes decisions based on the assessment score and risk level. If a scenario is determined to be high-risk (sensitive areas such as trading instructions, valuation opinions, compliance issues, or insufficient consistency score), it is forced into a human review or verification process, where experts conduct manual approval. If it is low-risk and the score meets the standard, the output is allowed and the cited source is indicated. If the evidence is insufficient or the data is unreliable, a downgrade strategy is implemented or the process is retried. If a definite error is detected, the output is intercepted and a compliance warning text is returned.

[0094] 10. Audit Log Recording: The system writes the complete decision-making chain (including the original user query, RAG retrieval of evidence sources, LLM model output, evaluation score, strategy decision, operator, and timestamp) into the blockchain audit log system, ensuring end-to-end traceability and immutability, meeting regulatory compliance requirements. The log retention period is 7-10 years. Audit fields adopt standardized GenAI semantic conventions, including observable fields such as coverage_ratio, consistency_score, citation_rate, hallucination_score, temporal_drift_count, and decision.

[0095] 11. Change Freeze and Kill Switch Emergency Mechanism: The system monitors anomaly detection indicators in real time, including abnormal increases in hallucination rate, number of temporal violations, and SLO exceeding thresholds. Once an anomaly is detected, the policy engine is triggered via Webhook to execute an emergency response, including downgrading (switching to a more conservative strategy), blocking (pausing LLM generation and only returning retrieved evidence), and rolling back (restoring to the previous stable version). In extreme cases, the Kill Switch emergency switch is activated, forcibly downgrading to the "evidence retrieval + human review" mode, completely stopping automatic LLM generation to ensure system security. The system in this application also includes underlying knowledge base governance and anomaly linkage closed loop. Specifically, the monitoring, alarm, and auditing modules monitor hallucination rate, coverage, and accuracy indicators in real time. Anomaly samples trigger an anomaly linkage mechanism via Webhook. This mechanism adopts an intelligent anomaly detection method based on GNN and temporal generation models. It replaces fixed thresholds by learning system behavior patterns, then performs natural language alarm analysis, and finally automatically stores the resulting samples in the database.

[0096] Please refer to Figure 4, which is a flowchart of the implementation of the intelligent anomaly detection method based on GNN and temporal generative models. This method combines existing GNN (Graph Neural Network) and temporal generative model technologies and applies them to anomaly detection scenarios in large model services. GNN is used to learn the topological relationships and dependency patterns between system components, while the temporal generative model is used to learn the time series patterns of indicators. The combination of the two can identify potential problems in advance before anomalies show obvious symptoms, replacing the traditional fixed threshold alarm method. The GNN module can learn the topological dependencies and fault propagation paths between system components, and can automatically locate the source of anomalies and trace the impact links. This enables the system to identify anomaly trends in advance and achieve proactive early warning. Intervention can be carried out before the fault appears, effectively reducing the scope of impact and preventing the fault from spreading; it automatically generates natural language alarms with fault root causes, impact links, and solution suggestions, which significantly shortens the mean time to recovery (MTTR), greatly improves the efficiency of operation and maintenance response, and does not rely on personal experience level.

[0097] Please refer to Figure 5, which is an architecture diagram of the system for governance and operation of multi-model collaborative large model illusion in the financial field in an embodiment of the present invention.

[0098] Key decision points: (1) Search coverage check: If the detection coverage is insufficient, rewrite the query, expand the recall scope or switch to a strong search model to ensure sufficient evidence.

[0099] (2) Mathematical problem branch: Force the use of computational strategies to generate numerical values, and LLM is only responsible for interpretation, thus avoiding numerical illusion.

[0100] (3) Fact consistency check: A four-layer evaluation pipeline (rule filtering → semantic consistency → CoT verification → ML scoring) is adopted. If the standard is not met, it will be downgraded or blocked.

[0101] (4) High-risk determination: When sensitive areas (such as transaction instructions, compliance issues, etc.) or inconsistencies are involved, the process will be forced to enter the manual review process.

[0102] (5) Abnormal linkage: When the illusion rate is abnormal, the temporal violation occurs, or the SLO exceeds the threshold, the Webhook triggers the strategy engine to perform downgrade, interception, or rollback.

[0103] Hallucination Detection and Reward Feedback Mechanism The hallucination detection and processing process includes four stages: evaluation, classification, triage, and reward feedback. I. Evaluation Stage 1. Model Output Reception: The answers generated by LLM enter the automatic evaluator for quality detection.

[0104] 2. Four-Layer Evaluation Pipeline Execution: The automated evaluator executes four layers of detection according to the principle of speed from fast to slow and complexity from simple to complex. The first layer is rule filtering (quickly detecting feature language (uncertain expressions such as "as far as I know," "maybe," "estimate," "probably," "perhaps," etc.) and the reasonableness of numerical ranges, identifying obvious low-quality outputs); the second layer is semantic consistency evaluation, which calculates the semantic consistency score and coverage between the LLM output and the retrieved evidence through sentence embedding vector comparison and SPO triple verification; the third layer is CoT inference verification, which step-by-step traces the inference chain to verify whether each inference step is supported by evidence, ensuring the integrity of the logical chain; the fourth layer is ML model comprehensive scoring, which uses an ensemble classifier to output a final comprehensive score of consistency score, coverage, and citation rate.

[0105] 3. Observable data collection: The system collects complete distributed tracing data and audit fields, including hallucination_score, coverage_ratio, consistency_score, citation_rate, and temporal_drift_count. All fields conform to the GenAI semantic convention standard, which facilitates subsequent fault location and root cause analysis.

[0106] 4. SLO Threshold Check Decision: The system compares the evaluation score with the preset SLO threshold. For financial securities scenarios, strict requirements are placed on factual consistency ≥ 0.90, median retrieval coverage ≥ 0.80, citation rate ≥ 0.95, and tense violations = 0 (zero tolerance). If any indicator fails to meet the standard, it is marked as a potential quality issue and enters the subsequent risk classification process.

[0107] 5. Version and Gray-scale Control Verification: The system verifies whether the current evaluation strategy version and threshold configuration are within the gray-scale whitelist to ensure the consistency and controllability of the evaluation standards.

[0108] 6. In-depth analysis of hallucination detectors: A rule-based + ML hybrid detection method is adopted, and the HalluFin framework is applied to identify fine-grained hallucination types, including four major categories of hallucination problems: time drift, projection overgeneralization, numerical inconsistency, and entity confusion.

[0109] II. Risk Level Classification: Based on the comprehensive assessment score, hallucination type, and business scenario, the output is divided into four levels: low risk, insufficient evidence, inconsistent values, and high risk.

[0110] III. After risk classification and triage, the system proceeds to the corresponding processing branch based on different risk levels: Branch 1: Insufficient Evidence Handling: If insufficient evidence is determined, the system initiates a downgrade, retry, or expanded recall strategy. Specific measures include rewriting the user query, expanding the recall scope, switching to a stronger retrieval model, and re-executing the RAG retrieval process to obtain more sufficient evidence support. If sufficient evidence cannot be obtained after a preset number of retries (e.g., 3), a compliance prompt of "Insufficient evidence, unable to provide a definitive answer" is output, along with a link to authoritative materials for the user to consult.

[0111] Branch 2: Handling Numerical Inconsistencies: If a numerical inconsistency is detected (the value generated by LLM does not match the result of the calculation strategy engine, or the value is obviously unreasonable), the system will force the calculation strategy engine to recalculate, and the deterministic calculation result will prevail. The system intercepts the raw output of LLM and returns a message to the user: "There is a deviation in the numerical calculation. It has been recalculated. The following is the accurate result," along with the complete calculation process, formulas, parameter settings, and data sources, ensuring that the value is reproducible and verifiable. This sample is marked as a negative sample (the reward value is set to a negative value: -2 or -4) for subsequent reward model optimization.

[0112] It should be noted that "this sample" in branch 2 refers to a complete record of a numerical inconsistency interaction. Specifically, a user submits a query involving numerical computation, and the numerical value generated by the large model is inconsistent with the result of the computational strategy engine (or the value is clearly unreasonable). After detecting this issue, the system intercepts the original output of the LLM and replaces it with the correct result from the computational strategy engine. This entire interaction process is recorded, forming a negative sample (with a reward value of -2 or -4), which is used in the subsequent optimization training of the reward model—the negative sample input training process—to allow the model to learn the decision boundary that "numerical computation must be accurate."

[0113] Branch 3 Low-Risk Release: If determined to be low-risk (sufficient evidence, high consistency, and citation compliance), the system releases the output and automatically marks the citation source. The output format includes the answer text and citation information (source document, page number, paragraph, and timestamp) to ensure traceability. The system records the complete decision chain (query + evidence + output + scoring + strategy) in the audit log, preserving the evidence chain and meeting compliance requirements. This sample is marked as a positive sample (reward value set to positive: 0.5 or 1) for positive reinforcement learning guidance.

[0114] It should be noted that "this sample" in branch 3 refers to the complete question-and-answer record that has been assessed as low-risk (sufficient evidence, high consistency, and citation standard) and allowed to be output by the system. Specifically, it includes: the user's original query, the evidence documents returned by RAG retrieval, the answer text generated by LLM and automatically labeled citation information (e.g., source documents, page numbers, paragraphs, timestamps), the complete decision chain, all recorded in the audit log, the scores of each of the four levels of evaluation, and the reward value label. The reward value label is set to a positive value (0.5 or 1), indicating that this is a correct / high-quality sample, used to guide the model's learning.

[0115] Branch 4 High-Risk Human Review: If a problem is determined to be high-risk (definitely incorrect, uncertain output in sensitive areas, extremely low consistency), the system will force it into a human review or verification process. The issue is assigned to a domain expert for manual review, and the expert can review the complete audit log for a comprehensive judgment. The expert's approval result is linked to the audit log, forming a traceable record of the human review. High-risk intercepted samples are marked as important negative samples (reward value = -4), and their high weight is used for reward model optimization and strategy threshold adjustment.

[0116] It should be noted that the "high-risk intercepted samples" in branch 4 refer to samples that are judged as high-risk by the system and forcibly intercepted to enter the manual review process. The triggering conditions are: the LLM output contains a definite error, the output involves uncertainties in sensitive areas, or the factual consistency is extremely low. Specifically, it includes: the user's original query, the evidence documents returned by the RAG retrieval, the original answer generated by the LLM (the intercepted sample), the scores of each item in the four-level evaluation, the risk level and illusion type determined by the system, the complete audit log, the expert approval results, and the reward value label. The reward value label is set to -4, indicating that this is an important negative sample, and it is used with high weight for reward model optimization and policy threshold adjustment.

[0117] IV. The reward feedback evaluation results are mapped to a 5-level reward value system for offline RLHF (Reinforcement Learning Based on Human Feedback) and online adaptive learning: Reward value = 1 (affirmative correct answer): sufficient evidence (coverage ≥ coverage threshold 0.80), high factual consistency (≥ financial scenario factual consistency score threshold 0.90), citation standard (citation rate ≥ citation rate 0.95), no feature words, no temporal violations, and complete logical chain. These samples serve as benchmark cases for model learning, guiding the model to learn the decision-making pattern of "sufficient evidence → high-quality output".

[0118] Reward value = 0.5 (fuzzy correct answer): The evidence is partially supportive (coverage between 0.65 and 0.80), with moderate consistency (factual consistency score between 0.75 and 0.90). The output is basically correct but has slight uncertainty or improper citation. This type of sample serves as a neutral-biased case, suggesting that the model should be cautious when the evidence is insufficient, but can still provide a biased opinion.

[0119] Reward value = 0 (no relevant knowledge found): Evidence is missing or coverage is extremely low (e.g., coverage less than 0.50), but the model correctly outputs "Insufficient evidence, unable to provide a definitive answer" instead of forcibly fabricating an answer. These samples reinforce the model's learning of the "insufficient evidence → admitting ignorance" decision-making pattern, avoiding the illusion of no evidence. These samples are archived separately to identify knowledge base gaps and trigger the knowledge base expansion process.

[0120] Reward value = -2 (fuzzy incorrect answer): The evidence contradicts the output, or there is a deviation in the numerical calculation but it does not cause serious misleading. These samples serve as negative samples to guide the model to learn to avoid inconsistent evidence output and improve the rigor of reasoning.

[0121] Reward value = -4 (positive incorrect answer): Output with clear evidence to refute the statement, or outputting a positive conclusion without any evidence. These samples are serious negative samples, and are given high weight for reward model optimization, reinforcing the decision boundary of "no output without evidence" and "interception of contradictory evidence". Positive errors involving trading instructions, valuation opinions, and compliance issues are marked as SEV1 level events, triggering emergency responses such as alarms, human review, and version rollback.

[0122] Learning Loop 1. Sample Input and Labeling: The system randomly selects online samples according to a certain proportion and combines them with expert annotations to form a high-quality training dataset. Each sample contains complete context (including user query, RAG retrieval evidence, and LLM output), detailed scores for four levels of evaluation, 5-level reward value labels, and expert correction opinions (optional). Positive samples (reward value ≥ 0.5) and negative samples (reward value < 0) are mixed in a 1:1 or 2:1 ratio to avoid sample distribution skew.

[0123] 2. Reward Model Update: With fact consistency, retrieval coverage, and citation rate as positive optimization objectives, and minimizing divergence loss (KL Divergence), the reward model parameters are updated using PPO (Proximity Policy Optimization) or DPO (Direct Preference Optimization) algorithms. The reward model learns to distinguish the output features corresponding to the five reward values, forming a more accurate quality assessment capability. The updated reward model is validated in a pre-release environment. After confirming that metrics such as illusion rate and consistency score do not degrade, it is gradually rolled out through canary releases.

[0124] 3. Knowledge Base Governance and Cleaning: Knowledge base gaps are identified based on "no relevant knowledge found" samples (reward value = 0), triggering domain experts to supplement relevant materials and expand the knowledge base's coverage. A regular knowledge base cleansing process is implemented, including timeliness verification, contamination detection, and graph node validity filtering. Timeliness verification includes marking or deleting expired data; contamination detection includes isolating abnormal samples to a contaminated database; and graph node validity filtering (GraphRAG graph nodes have a time validity attribute, and expired nodes are automatically filtered during retrieval) also includes knowledge base versioning management, generating a new version with each major update, supporting canary releases and rollbacks.

[0125] 4. Feedback to RAG Retrieval Optimization: Query rewriting strategies, recall optimization methods, and graph traversal paths from high-quality samples are fed back into the RAG retrieval service to improve retrieval quality. For example, the query characteristics of "insufficient evidence" samples are analyzed to optimize entity recognition and keyword extraction algorithms; the graph traversal paths of "high retrieval coverage and high output quality" samples are analyzed to optimize GraphRAG's multi-hop strategy and mixed ranking weights. Through a continuous learning loop, the system continuously improves in retrieval accuracy, evidence sufficiency, and output consistency, forming a closed loop of "evaluation → feedback → optimization → re-evaluation".

[0126] Operational Constraints and Guarantee Mechanisms 1. Unified Observable Implementation: Following GenAI Semantic Conventions, collect Traces / Metrics / Events / Logs throughout the entire RAG → Prompt → Generation → Evaluation process. Standard fields include gen_ai.system, gen_ai.request.model, gen_ai.usage.input_tokens, gen_ai.usage.output_tokens, and extended fields prompt_id, prompt_version, retrieval_k, coverage_ratio, consistency_score, citation_rate, hallucination_score, temporal_drift_count, graph_hop_depth, graph_transition_weight, latency_ms, cost_usd, and decision.

[0127] GenAI Semantic Conventions: Generative AI semantic conventions are GenAI observability standard specifications defined by the OpenTelemetry community, providing a unified telemetry data acquisition format for LLM applications; Traces / Metrics / Events / Logs: Distributed traces / metrics / events / logs constitute the three pillars of observability, used for link tracing, performance measurement, state change recording, and detailed log analysis, respectively; Standard field descriptions: gen_ai.system: AI system identifier, indicating the name of the large model service provider or system currently used; gen_ai.request.model: Request model identifier, recording the inference call for this instance. The specific model name and version; `gen_ai.usage.input_tokens`: Input token count, recording the number of tokens consumed by the input prompts in this request, used for cost accounting and usage monitoring; `gen_ai.usage.output_tokens`: Output token count, recording the number of output tokens generated by the model in this request, used for cost accounting and usage monitoring; Extended field descriptions: `prompt_id`: Unique identifier for the prompt, used to trace the source and historical changes of the Prompt template; `prompt_version`: Prompt version number, recording the currently used Prompt template version, supporting canary releases. With rollback; retrieval_k: number of top-k retrieved evidence fragments, the number of evidence fragments returned by RAG retrieval, affecting coverage and inference cost; coverage_ratio: retrieval coverage, measuring the degree to which the fragments returned by RAG hit the question, ranging from 0 to 1; consistency_score: factual consistency score, measuring the semantic consistency between the LLM output and the retrieved evidence, ranging from 0 to 1; citation_rate: citation rate, measuring the proportion and normativity of evidence citations in the answer, ranging from 0 to 1; hallucination_score: hallucination score, comprehensively assessing the hallucination risk level of the output, the higher the value, the greater the risk; tem `polal_drift_count`: Temporal drift count, records the number of times expired data is referenced in the output; a value of 0 is required for financial scenarios. `graph_hop_depth`: Graph hop count, records the multi-hop depth during GraphRAG retrieval, typically 2-3 hops. `graph_transition_weight`: Graph transition weight, records the weight percentage of graph retrieval in GraphRAG hybrid sorting. `latency_ms`: Latency in milliseconds, records the end-to-end response time for this request. `cost_usd`: Cost in US dollars, records the LLM inference cost for this request. `decision`: Decision result, records the final decision of the policy engine.

[0128] 2. Automated Instrumentation and Data Acquisition: Zero-code intrusive telemetry acquisition of LangChain / LlamaIndex is performed through Auto-Instrumentation Libraries. The Batch Exporter sends the data to a time-series database, a distributed tracing backend, and a log aggregation platform.

[0129] Terminology Explanation: Auto-Instrumentation Libraries: Automated instrumentation libraries that automatically collect distributed tracing data without modifying business code; LangChain / LlamaIndex: A mainstream LLM application development framework that enables zero-intrusion telemetry data collection; Batch Exporter: A batch exporter that sends collected telemetry data to the backend storage system in batches; Time Series Database: Used to store Metrics data; Distributed Tracing Backend: Used to store Traces data; Log Aggregation Platform: Used to store and analyze Logs data.

[0130] 3. AI Agent Observability Enhancement: Implement distributed tracing for multi-agent systems, capture inter-agent interactions, decision chains, and tool call sequences, and support end-to-end latency analysis and error tracing across agents.

[0131] 4. Versioning and Canary Release Control: Versioning of prompts / routes / models.

[0132] 5. Intelligent Anomaly Response: Anomaly detection based on GNN and time-series generation model triggers the policy engine, and natural language alarm analysis automatically generates root causes of failures, scope of impact, and solution suggestions. Samples are automatically added to the database to drive the continuous learning of reward models and knowledge base governance.

[0133] Since the system in Embodiment 2 of this invention is the same system used in the method for governance and operation of multi-model collaborative large-scale models in the financial field described in Embodiment 1, those skilled in the art can understand the specific structure and variations of this system based on the method introduced in Embodiment 1 of this invention, and therefore will not be repeated here. All systems used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0134] Based on the same inventive concept, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in Embodiment 1.

[0135] Since the computer device described in Embodiment 3 of this invention is the same computer device used to implement the method for multi-model collaborative large-scale model illusion governance and operation and maintenance in the financial field in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this computer device based on the method described in Embodiment 1 of this invention, and therefore will not be described again here. All computer devices used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0136] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0137] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0138] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various modifications and variations to the embodiments of the invention without departing from the spirit and scope of the invention. Thus, if these modifications and variations of the embodiments of the invention fall within the scope of the claims of the invention and their equivalents, the invention also intends to include these modifications and variations.

Claims

1. A method for governance and operation of a large-scale collaborative multi-model illusion in the financial field, characterized in that, include: Receive user queries; perform type identification and feature extraction on the received user queries to obtain type features, and obtain the corresponding route labels and constraint configurations based on the type features; The processing path is selected based on the routing label. This path includes a computational strategy path and a RAG retrieval enhancement path. The computational strategy path invokes a pre-built computational strategy engine to generate computational strategy results. The RAG retrieval enhancement path retrieves relevant evidence fragments from the knowledge base, calculates retrieval coverage, and selects whether to add the retrieved evidence fragments to the context based on the retrieval coverage. A prompt template is constructed based on the routing label, retrieval results, and constraint configuration, embedding evidence fragments, citation format requirements, and strong mathematical constraints. Combining the RAG retrieval evidence, computational strategy results, and constructed prompt template, a pre-defined large model is invoked for inference. A multi-layered evaluation strategy is used to evaluate the output inference results. Risk grading and strategy execution are then performed based on the evaluation results.

2. The method for governance and operation of multi-model collaborative large-scale model illusion in the financial field as described in claim 1, characterized in that, The system performs type identification and feature extraction on received user queries to obtain type features, including: identifying the type of the received user queries to obtain the question intent category, including fact query, content creation, mathematical calculation, logical reasoning, and document writing; performing preset keyword detection on user queries to obtain risk markers; and statistically analyzing the frequency of numbers and the density of numerical values ​​in user queries to obtain frequency analysis results.

3. The method for governance and operation of multi-model collaborative large-scale model illusion in the financial field as described in claim 2, characterized in that, The method also includes: identifying numerically intensive problems based on frequency analysis results.

4. The method for governance and operation of multi-model collaborative large-scale model illusion in the financial field as described in claim 1, characterized in that, The RAG retrieval enhancement path is specifically used for: loading documents from the knowledge base and cutting the documents into slices of a preset length; using an embedding model to perform semantic embedding on each slice, constructing a vector database; using an embedding model to convert user queries into vectors; performing vector retrieval based on the similarity between the user query vector and the vectors in the vector database; constructing a knowledge graph from the knowledge base; embedding user queries; and performing nearest neighbor search or graph traversal to identify nodes related to the query embedding from the knowledge graph. Based on the results of vector retrieval and graph retrieval, evidence fragments are obtained.

5. The method for governance and operation of multi-model collaborative large-scale model illusion in the financial field as described in claim 1, characterized in that, Combining RAG retrieval evidence, computational strategy results, and constructed prompt templates, the system invokes a pre-defined large model for reasoning. This includes: for user queries involving mathematical calculations, the pre-defined large model is responsible for interpreting the computational steps and using the computational strategy results generated by the computational strategy path as the numerical results obtained through reasoning; for user queries of other types besides mathematical calculations, the pre-defined large model uses thought chains or thought trees to expand the logic, and the reasoning results include answer text, evidence citations, and reasoning chains.

6. The method for governance and operation of multi-model collaborative large-scale model illusion in the financial field as described in claim 1, characterized in that, A multi-layered evaluation strategy is employed to assess the output reasoning results, including: conducting a reasonableness analysis on the characteristic statements and numerical ranges in the reasoning results; analyzing the semantic consistency between the reasoning results and the retrieved evidence, and calculating a consistency score; tracing the evidence chain to verify the integrity of the reasoning chain, and calculating the citation rate to represent the proportion of evidence cited in the answer; and using the consistency score, retrieval coverage, and citation rate as a comprehensive evaluation result.

7. The method for governance and operation of multi-model collaborative large-scale model illusion in the financial field as described in claim 2, characterized in that, Based on the assessment results, risk classification and strategy execution are carried out, including: determining whether the risk meets the first preset condition for high risk based on the risk marker; if so, manual review or verification is carried out; otherwise, the data is released and marked as cited; comparing the calculation strategy results generated by the calculation strategy path with the output of the preset large model; if the values ​​are inconsistent, interception and recalculation are triggered; determining whether the search coverage is less than the coverage threshold; if it is less than the threshold, downgrading is carried out and retry is performed or an insufficient evidence prompt is returned.

8. The method for governance and operation of multi-model collaborative large-scale model illusion in the financial field as described in claim 1, characterized in that, The method also includes collecting and monitoring data during the input phase, reasoning phase, evaluation phase, and decision-making phase.

9. A system for governance and operation of a large-scale collaborative multi-model illusion in the financial field, characterized in that, include: The user query receiving module is used to receive user queries. The problem analysis and type identification module is used to identify the type and extract features from received user queries, obtain type features, and obtain corresponding routing labels and constraint configurations based on the type features. The processing path selection module is used to select the processing path according to the routing label. The processing path includes a computational strategy path and a RAG retrieval enhancement path. The computational strategy path is used to call a pre-built computational strategy engine to generate computational strategy results. The RAG retrieval enhancement path is used to recall relevant evidence fragments from the knowledge base, calculate retrieval coverage, and select whether to add the recalled evidence fragments to the context based on the retrieval coverage. The prompt template construction module is used to construct prompt templates based on routing labels, retrieval results, and constraint configurations, embedding evidence fragments, citation format requirements, and strong mathematical constraints. The reasoning module is used to combine RAG retrieval evidence, computational strategy results, and constructed prompt templates to call a preset large model for reasoning. The evaluation and decision-making module is used to evaluate the output reasoning results using a multi-layer evaluation strategy. The strategy execution and risk classification module is used to classify risks and execute strategies based on the evaluation results.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method for governance and operation of multi-model collaborative large-scale model illusion in the financial field as described in any one of claims 1 to 8.

Images