Bond product issuing document auditing method and system based on large model agent

Abstract

The invention provides a bond product issuing document auditing method and system based on a large model agent, and the method comprises the following steps: converting unstructured auditing knowledge into structured auditing rules through an auditing rule construction agent, and storing the structured auditing rules into a prompt word bank, content positioning information is generated based on historical document statistics and stored in a position library; obtaining a to-be-audited document and identifying a bond product type to match the audit key point; respectively calling rules and positioning information from a prompt word library and a position library according to key points, and extracting related content fragments; and the rules and the fragments are combined to form dynamic cues, the dynamic cues are submitted to a large language model (LLM) for analysis, and a standardized auditing conclusion is output. The system comprises a knowledge construction module, a database module, a document processing module, a content extraction module, a dynamic cue word generation module and an LLM processing module, the auditing efficiency and consistency can be improved, and the maintenance cost can be reduced.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing technology, specifically relating to a method and system for reviewing bond product issuance documents based on a large model intelligent agent. Background Technology

[0002] As a core channel for direct corporate financing, the bond market plays a vital role in supporting the real economy and optimizing resource allocation. In recent years, with the increasing complexity and diversification of bond products, and the issuance venues covering the National Association of Financial Market Institutional Investors (NAFMII) and stock exchanges in the interbank market, a large number of specialized products have emerged in the market.

[0003] To address the challenges posed by the high diversity and complexity of bond products to the review of issuance documents, the industry currently relies heavily on manual review by professionals from intermediaries such as securities firms, law firms, and accounting firms. In addition, to improve efficiency, some basic automation tools have been introduced, such as scripts based on keyword matching or fixed rules, to assist in completing some review tasks.

[0004] However, the aforementioned existing technologies have significant limitations: manual review is inefficient and costly, requiring substantial manpower and time to review a prospectus that can easily reach hundreds of pages. Furthermore, standardized review criteria are difficult to unify, and differing interpretations of regulations and rules by different personnel lead to subjectivity and inconsistent quality in the review process. Simultaneously, the depth and breadth of manual processing are limited; handling large volumes of text information in a short time can easily result in oversights and errors, making comprehensive screening and complex analysis difficult. While traditional automated tools offer some assistance, they lack the ability to understand deep semantic context, cannot identify complex business logic and potential risks, and are inflexible. When document format, chapter structure, or expression changes slightly, fixed rules easily become ineffective, resulting in extremely high maintenance costs and failing to meet the actual needs of current bond market reviews. Summary of the Invention

[0005] This invention proposes a method and system for reviewing bond product issuance documents based on a large model intelligent agent, which solves the problems of low efficiency and poor flexibility of existing review methods.

[0006] To address the aforementioned technical problems, this invention provides a method for reviewing bond product issuance documents based on a large-scale intelligent agent model, comprising the following steps:

[0007] Step S1: Construct an intelligent agent through the review rules to transform unstructured review knowledge into structured review rules and store them in the prompt word library. Perform statistical analysis on historical documents through the content location statistics intelligent agent to generate content location information and store it in the location library.

[0008] Step S2: Obtain the bond product issuance documents to be reviewed, identify the bond product type of the documents to be reviewed, and match the corresponding set of review points according to the bond product type;

[0009] Step S3: For each audit point in the audit point set, perform the following operations: retrieve the corresponding structured audit rules from the prompt word library according to the audit point, retrieve the corresponding content positioning information from the location library, and extract content fragments related to the audit point from the document to be audited according to the content positioning information;

[0010] Step S4: Combine the structured review rules with the content fragments to generate dynamic prompt words for LLM (Large Language Model) processing. Send the dynamic prompt words to the LLM. The LLM analyzes the content fragments based on the dynamic prompt words and outputs the review results corresponding to the review points.

[0011] Preferably, the step S1 of transforming unstructured audit knowledge into structured audit rules includes the following steps: decomposing the unstructured audit knowledge into data analysis paths and logical judgment rules, transforming the logical judgment rules into a multi-step instruction set executable by the large language model, defining a standardized format for the output of the large language model, and integrating the multi-step instruction set with the standardized format to obtain structured audit rules.

[0012] Preferably, the statistical analysis of historical documents by the content positioning statistical agent in step S1 includes the following steps: performing statistical analysis on the chapters, titles, and table names of the historical documents, and summarizing the location patterns of the information required for the review points in the documents to be reviewed.

[0013] Preferably, step S3, which involves extracting content fragments related to the review points from the document to be reviewed based on the content location information, includes the following steps: extracting the content fragments from specific chapters, specific titles, or specific tables of the document to be reviewed based on the location information stored in the location library.

[0014] Preferably, the dynamic prompt word in step S4 includes several placeholders, and the content fragment is filled into the placeholders to generate the dynamic prompt word.

[0015] Preferably, in step S4, the LLM analyzes the content segment based on the dynamic prompt words, including the following steps: performing a preset quantitative calculation on the content segment to obtain a risk index, comparing the risk index with a preset risk threshold, and determining the risk level.

[0016] Preferably, the audit result corresponding to the audit points in step S4 is a text conforming to a preset standard format, which includes a conclusion, core indicators, and in-depth analysis.

[0017] Preferably, the set of audit points in step S2 includes at least one of the following types of points: corporate governance and organizational structure, debt, cash flow, profit, assets, business, or compliance and quality.

[0018] This invention also provides a review system for bond product issuance documents based on a large model intelligent agent, which is based on the above-mentioned review method for bond product issuance documents based on a large model intelligent agent, and includes: a knowledge construction module, a database module, a document processing module, a content extraction module, a dynamic prompt word generation module, and a large language model processing module;

[0019] The knowledge construction module transforms unstructured audit knowledge and historical documents into structured data and stores them in a database.

[0020] The database module includes a prompt word library and a location library. The prompt word library is used to store structured review rules, with each rule associated with a bond product type and review points. The location library is used to store content location information, including descriptions of common locations of each review point in the document.

[0021] The document processing module: processes newly input documents to be reviewed and determines the key points that need to be reviewed;

[0022] The content extraction module extracts relevant content fragments from the document to be reviewed for each review point.

[0023] The dynamic prompt word generation module combines the review rules and content fragments to generate dynamic prompt words for each review point, and sends the dynamic prompt words to the large language model;

[0024] The large language model processing module analyzes content fragments based on dynamic prompts to determine whether they are compliant or have problems.

[0025] Preferably, the knowledge construction module includes an audit rule construction agent and a content positioning and statistics agent. The audit rule construction agent is used to parse unstructured audit knowledge, extract key rules through natural language processing technology, and transform the key rules into structured audit rules. The content positioning and statistics agent is used to perform statistical analysis on historical documents, identify common content patterns and data points of different types of bond products, and generate content positioning information.

[0026] The beneficial effects of the present invention include at least the following:

[0027] 1. By transforming unstructured, expert-experience-dependent review knowledge into structured review rules and content positioning information and storing it in a database, the originally scattered and implicit knowledge becomes centralized, explicit, and reusable. This not only ensures that all document reviews are based on the same set of standard rules, effectively eliminating the problem of inconsistent standards caused by differences in subjective understanding among different reviewers, but also greatly improves the standardization and consistency of review results. Furthermore, it solidifies the core experience of experts into the system, reducing the organization's over-reliance on individual experts and preventing knowledge loss.

[0028] 2. An efficient and precise positioning and extraction process has been built. The system automatically identifies bond types, matches key review points, and directly extracts specific content fragments that need to be reviewed based on the location database information, instead of sending the entire document to LLM. This avoids reviewers having to manually flip through hundreds of pages of documents to find relevant clauses. At the same time, only the most relevant content fragments are sent to LLM, reducing interference from irrelevant information. This reduces token consumption costs and improves the focus and accuracy of LLM analysis.

[0029] 3. By using dynamic prompts, the structured review rules and the specific content to be reviewed are handed over to the LLM for processing. This enables the LLM to understand the context and deep semantics of the content. It can identify logical contradictions, potential risks, and imprecise expressions that cannot be found by fixed rules or keyword matching. When encountering new document formats or expressions, there is no need to rewrite complex rules. The system can quickly adapt to changes simply by updating the rule descriptions in the prompt word library or the location information in the location library. The maintenance cost is far lower than that of traditional script programs. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0031] Figure 2 This is a schematic diagram of the structured prompt word construction process according to an embodiment of the present invention;

[0032] Figure 3 This is a schematic diagram of the system structure according to an embodiment of the present invention;

[0033] Figure 4 This is a schematic diagram of the review process in Embodiment 1 of the present invention;

[0034] Figure 5 This is a schematic diagram of the review process in Embodiment 2 of the present invention;

[0035] Figure 6 This is a schematic diagram of the review process in Embodiment 3 of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0037] like Figure 1 As shown, this embodiment of the invention provides a method for reviewing bond product issuance documents based on a large model intelligent agent, including the following steps:

[0038] Step S1: Construct an intelligent agent through the review rules to transform unstructured review knowledge into structured review rules and store them in the prompt word library. Perform statistical analysis on historical documents through the content location statistics intelligent agent to generate content location information and store it in the location library.

[0039] Specifically, the intelligent agent constructing the review rules transforms newly added laws and regulations or texts formed from the accumulated experience of review experts into standardized review rules, which are then stored in or updated in the prompt word library. The content positioning and statistics intelligent agent analyzes a large number of similar historical documents, generates or updates content positioning statistics on different review points, and stores them in the location database.

[0040] Step S2: Obtain the bond product issuance documents to be reviewed, identify the bond product type of the documents to be reviewed, and match the corresponding set of review points according to the bond product type.

[0041] Step S3: For each audit point in the audit point set, perform the following operations: retrieve the corresponding structured audit rules from the prompt word library based on the audit point, retrieve the corresponding content positioning information from the location library, and extract content fragments related to the audit point from the document to be audited based on the content positioning information.

[0042] Step S4: Combine the structured review rules with the content fragments to generate dynamic prompt words for LLM (Large Language Model) processing. Send the dynamic prompt words to the LLM. The LLM analyzes the content fragments based on the dynamic prompt words and outputs the review results corresponding to the review points.

[0043] Specifically, the bond product type of the document to be reviewed is identified, and the corresponding set of review points is matched. Based on the review points, review rules are retrieved from the prompt word library, and content location information is retrieved from the location library. Based on the content location information, relevant content is extracted from the document to be reviewed. The review rules and the extracted content are combined into dynamic prompt words, which are then processed by a large language model to generate analysis results.

[0044] The prompt word library in this embodiment of the invention contains a prompt word library for internal control auditing. It makes explicit and structured the implicit experience of financial auditing experts, and transforms the implicit experience of financial auditing experts into explicit, structured, and machine-executable intelligent auditing programs.

[0045] like Figure 2 As shown, the method for constructing this suggestion dictionary specifically includes the following steps:

[0046] S1: Domain Knowledge-Driven Audit Target Definitions: Each prompt begins with a clear and specific audit target, jointly extracted by senior experts in finance, accounting, and law based on the latest regulatory provisions, industry practices, and historical risk cases. This ensures that all audit tasks stem from genuine and compliant business needs.

[0047] S2: Expert Logic Deconstruction and Structuring: Systematically deconstruct the thought process of domain experts for the defined audit points. That is, transform the analytical path, judgment basis, data correlation, and risk threshold in the expert's mind when conducting the audit into a clear, unambiguous logical chain and structured framework.

[0048] S3: Large Language Model (LLM) Instruction Adaptation and Task Decomposition: The structured expert logic formed in the previous step is further translated and adapted into instructions that the LLM can understand and execute. Complex review points are decomposed into a series of logically clear and progressive sub-tasks, guiding the LLM to complete complex analysis according to a preset path.

[0049] S4: Standardized Output Specification Design: To ensure that the analysis results generated by LLM can be directly used by downstream systems or quickly interpreted by auditors, this embodiment of the invention designs strict standardized output specifications for each prompt word. It requires LLM to organize its output content in Markdown format and a specific paragraph structure of conclusion first and arguments second, ensuring the standardization, consistency and usability of the results.

[0050] S5: Prompt Message Integration and Placeholders: This feature organically integrates core task instructions adapted to LLM directives and standardized output specifications into a complete prompt message template. More importantly, it includes multiple pre-defined placeholders. These placeholders are used to dynamically fill in specific content extracted in real-time from the document to be reviewed, such as financial data and paragraph text, during the actual review process.

[0051] Through the aforementioned systematic construction process, the prompt word library constructed in this embodiment of the invention comprehensively covers seven major audit categories: corporate governance and organizational structure, debt, cash flow, profit, assets, business operations, and compliance and quality. Each prompt word becomes an intelligent subroutine that encapsulates complete business logic, analysis paths, and output standards. Through placeholder design, it can flexibly adapt to countless specific audit scenarios, thereby ensuring the professionalism, depth, and automation level of the audit.

[0052] like Figure 3 As shown, this embodiment of the invention also provides a review system for bond product issuance documents based on a large model intelligent agent, which is implemented based on the above-mentioned review method for bond product issuance documents based on a large model intelligent agent, including: a knowledge construction module, a database module, a document processing module, a content extraction module, a dynamic prompt word generation module, and a large language model processing module.

[0053] The knowledge construction module includes a review rule construction agent and a content location statistics agent. The review rule construction agent is used to parse unstructured review knowledge. It contains a rule specification agent, which receives review rules uploaded from external sources, including laws and regulations, regulatory Q&A, departmental rules, etc., as well as the experience of review experts. That is, the review experience of senior review experts is precipitated and solidified into text form, and automatically processed and normalized into standardized review rules containing specific business logic, judgment criteria and output format requirements, and stored in the prompt word library for subsequent use.

[0054] The content positioning and statistical agent includes a location statistics agent that receives and processes a large number of similar historical bond issuance documents. By statistically analyzing information such as the structure, chapters, titles, and table names of these documents, it summarizes the high-probability location distribution patterns of key information within the documents, targeting specific review points such as bank credit analysis, non-operating transaction risks, and working capital accumulation risks. These patterns are quantified as content positioning statistics and stored in a location database, forming an information positioning index.

[0055] The database module includes a cue word library and a location library. The cue word library stores structured review rules, with each rule associated with a bond product type and review criteria. The location library stores content location information, including descriptions of common locations of each review criterion in the document.

[0056] Document processing module: Analyzes uploaded documents awaiting review, identifies the specific type of bond product, and automatically matches all relevant review points applicable to that type of product based on the identified product type.

[0057] Content extraction module: For each matched review point, it sends requests to both the prompt word library and the location library. It retrieves the standardized review rules corresponding to that review point from the prompt word library and the location index of the information needed to determine that review point within the document from the location library. Based on the location information obtained from the location library, it accurately extracts relevant text, data, and other content from the corresponding location in the document to be reviewed.

[0058] The dynamic prompt word generation module combines static review rules obtained from the prompt word library with content dynamically extracted from the document to generate a complete dynamic prompt word that includes contextual information and a clear review task.

[0059] Large Language Model Processing Module: The dynamic prompt word is sent to the Large Language Model (LLM) for in-depth analysis. The LLM performs logical judgments, calculations, comparisons, and risk assessments on the extracted content according to the instructions, and finally generates structured analysis results.

[0060] Example 1: "Analysis of Remaining Credit and Liquidity Risk"

[0061] This embodiment details how the present invention can be applied to the automated review of the issuer's "residual credit and liquidity risk" in bond issuance documents. The entire process follows the five-layer progressive prompt framework proposed in this invention, transforming the complex review task into a precise and executable set of instructions. Figure 4 As shown, it includes the following steps:

[0062] S1: Definition of Key Audit Points Driven by Domain Knowledge

[0063] Senior financial experts and internal reviewers, based on industry best practices, regulatory requirements, and experience from historical review cases, precisely extracted the core review point: remaining credit and liquidity risk analysis. At this stage, experts identified key risk clues and quantifiable thresholds, forming the foundational knowledge framework for this review point. This includes:

[0064] Bank classification standards: To facilitate statistical analysis, bank categories are clearly defined. For example, banks can be divided into policy banks, state-owned commercial banks, joint-stock banks, and regional banks, among which regional banks include city commercial banks, rural commercial banks, and private banks.

[0065] Risk Quantification Indicators: Multi-level risk indicators are established. For example, risk levels for credit concentration are defined, such as classifying regional bank credit lines exceeding 50% as high-risk; and liquidity risk indicators are supplemented, such as indicating "severely insufficient available credit" when the total credit utilization rate exceeds 80%. These definitions are based on extensive real-world case analysis and risk assessment, providing clear and quantifiable standards for subsequent logical decomposition and model analysis.

[0066] S2: Expert Logic Decomposition and Structuring

[0067] Experts have broken down the review logic of "remaining credit and liquidity risk analysis" in depth, structuring it into a clear "data analysis path" and "risk assessment rules" to ensure the rigor and reproducibility of the analysis.

[0068] The analysis path includes bank classification mapping, core indicator calculation, and risk assessment rules.

[0069] Bank classification mapping is used to establish mapping rules from the full name of a bank to a standard classification, such as a branch of a bank in a certain city of a certain province should be accurately classified as a regional bank.

[0070] Core metrics calculation is used to determine the key metrics that need to be calculated, including:

[0071] The percentage of total credit line granted by bank category = (total credit line granted by that bank category / total credit line granted) × 100%;

[0072] Credit balance percentage by bank category = (Credit balance of that bank category / Total credit balance) × 100%;

[0073] Total credit utilization rate = (Total credit balance / Total credit amount) × 100%.

[0074] Risk assessment rules are used to transform expert judgment experience into a series of explicit conditional statements:

[0075] If the total amount or balance of regional banks accounts for more than 50%, regional concentration risk is triggered.

[0076] If the total credit utilization rate is greater than 80%, the risk of insufficient credit space will be triggered.

[0077] If none of the high-risk conditions are met, the risk is deemed controllable.

[0078] S3: LLM Instruction Adaptation and Task Decomposition

[0079] The structured, manually reviewed logic is transformed into a multi-step, multi-level instruction set that a Large Language Model (LLM) can accurately understand and execute. This includes:

[0080] Data processing instructions: Explicitly require LLM to perform data cleaning and standardization before calculation, such as: strictly matching bank classification, such as classifying a branch of a rural commercial bank as a regional bank, automatic unit conversion, such as retaining 2 decimal places when converting 10,000 yuan to 100 million yuan, and calculating the percentage to 0.01%.

[0081] Logical analysis instructions: Directly convert the analysis paths and calculation formulas decomposed in S2 into LLM task steps.

[0082] Risk Assessment and In-Depth Analysis Instructions: This step not only requires LLMs to conduct an initial risk assessment according to the S2 rules, but also further requires them to conduct in-depth analyses from specific perspectives for different risk types. For example, for regional concentration risk, the instructions require that an analysis of regional financial dependence and the risks of non-standard financing be included.

[0083] S4: Standardized Output Specification Design

[0084] To ensure the professionalism, usability, and consistency of LLM outputs, this step employs a strict hierarchical and field-based output specification. The output must include four main modules: risk type conclusions, core indicators, in-depth analysis, and management recommendations. An example of the actual output after LLM processing is shown below, which fully conforms to the pre-defined output specifications:

[0085] I. Regional Concentration Risk

[0086] Key indicators: Regional banks accounted for 63.30% of the total amount and 60.18% of the balance; policy banks accounted for 0%, state-owned banks accounted for 0.84%, and joint-stock banks accounted for 35.86%; the total credit utilization rate was 78.18%.

[0087] In-depth analysis: Regional financial dependence: 63.3% of credit lines are concentrated in a local provincial bank, and regional economic fluctuations may lead to a synchronized tightening of credit in the banking system.

[0088] Non-standard financing penetration risk: The credit line of a certain joint-stock bank X accounts for 35.47%, but the bank has experienced frequent risk events in recent years, and it is necessary to verify whether there is any non-standard channel business.

[0089] Liquidity buffer calculation: Of the 443 million yuan of unused credit, 310 million yuan is dependent on joint-stock bank X, accounting for 70%. If the bank reduces the credit line, it will directly lead to a decrease in the liquidity coverage ratio.

[0090] Management suggestions:

[0091] 1. Prioritize using unused credit limits from banks with higher stability.

[0092] 2. Establish strategic cooperation with state-owned banks to reduce the concentration of regional banks to below 50%.

[0093] 3. Conduct stress tests on credit lines granted by joint-stock banks and prepare contingency financing plans.

[0094] To ensure the traceability and verifiability of the results, a data processing verification process will also be generated:

[0095] II. Data Processing Verification

[0096] 1. Classification verification:

[0097] Regional bank = a provincial bank (535 million) + a rural commercial bank (90 million) + a municipal bank (158 million) → total 783 million / total credit line 1.237 billion ≈ 63.30%.

[0098] Unit conversion: Joint-stock bank x credit balance of RMB 310 million automatically converts to RMB 310 million.

[0099] Anomaly alert: A provincial bank's credit line has been 100% utilized, and its willingness to renew the credit line needs to be closely monitored.

[0100] S5: Prompt word integration and placeholders

[0101] All the above instructions, rules, and specifications are logically integrated into a complete and self-consistent prompt word template. The core of this template is a placeholder {{info}}. During actual execution, the system will extract data relevant to the current review point from the document to be reviewed, such as a bank credit details table, and populate this parameter location, thus creating a dynamic prompt word that can be executed immediately.

[0102] A complete, finalized prompt template is as follows:

[0103] Task objective: To assess the stability of the issuer's debt structure and identify risks associated with concentrated credit lines and liquidity pressures.

[0104] Input data: {{info}}.

[0105] Analysis steps: Ensure that the information extracted from the input data is correctly processed and used for calculations during this process.

[0106] 1. Bank Classification:

[0107] Policy banks: China Development Bank, Export-Import Bank of China, Agricultural Development Bank of China and their branches; state-owned commercial banks; joint-stock banks; regional banks: city commercial banks, rural commercial banks, private banks, etc., other than those mentioned above.

[0108] 2. Data Calculation:

[0109] The percentage of total credit granted by each type of bank is calculated as follows: (Total credit granted by that type / Total credit granted) × 100%.

[0110] Calculate the percentage of outstanding credit balance for each type of bank = (Outstanding credit balance for that type / Total outstanding credit balance) × 100%;

[0111] Calculate the total credit utilization rate = (Total credit balance / Total credit amount) × 100%.

[0112] Risk assessment rules:

[0113] 1. If the total amount of regional banks accounts for more than 50% or the balance accounts for more than 50%, regional concentration risk will be triggered.

[0114] 2. If the total amount of policy banks accounts for more than 50% or the balance accounts for more than 50%, then the concentration risk of policy banks will be triggered.

[0115] 3. If the total credit utilization rate is greater than 80%, the risk of insufficient credit space will be triggered;

[0116] 4. If none of these conditions are met, the risk is considered controllable.

[0117] Output specifications:

[0118] [Risk Type Conclusion]

[0119] Key metrics: Presentation of key data, such as: regional banks' total share of X%, balance share of Y%, and utilization rate of Z%.

[0120] In-depth analysis:

[0121] Perspective 1: [Specific Analysis];

[0122] Perspective 2: [Quantitative Measurement];

[0123] Management recommendations: [Targeted measures].

[0124] Implementation requirements:

[0125] 1. Data Processing

[0126] It strictly matches bank classifications, automatically converts units, and calculates percentages with an accuracy of 0.01%.

[0127] 2. Analytical dimensions that must be included in the risk assessment perspective: regional concentration risk, including regional financial dependence and non-standard financing penetration risk; policy concentration risk, including the impact of policy continuity and fiscal subsidy gap; credit space, i.e. debt maturing in the next 12 months vs. unused credit gap; and controllable risk, including credit maturity distribution and the strategic stability of major banks.

[0128] The final output obtained at this audit point is:

[0129] 1. Analysis of Remaining Credit Line and Liquidity Risk

[0130] Regional banks have excessively high credit concentration: total credit line accounts for 63.05%, and outstanding credit line accounts for 62.70%.

[0131] The credit resources are sufficient, with a remaining margin of 21.82%, providing an effective liquidity buffer. However, there is a risk of concentration in the credit structure.

[0132] LLM analysis results:

[0133] Regional concentration risk: Remaining credit lines are concentrated in joint-stock bank X, accounting for 70%; regional banks collectively account for over 60% of credit lines, with one provincial bank alone accounting for 26.35%, reflecting a high dependence on local financial institutions for financing channels. The credit utilization rate of 78.18% is close to the 80% warning line. Unused credit is concentrated in joint-stock bank X, amounting to 310 million yuan, while other banks have only 53 million yuan in standby credit, indicating weak emergency withdrawal capabilities. Short-term credit accounts for approximately 12.4%, requiring assurance of timely renewal upon maturity.

[0134] Demand Gap: Regional banks typically have lower capital adequacy ratios. An economic downturn in the issuer's region could trigger a collective tightening of credit lines by these banks. It is recommended to review the credit agreements with joint-stock banks for clauses that could trigger credit line reductions, establish a dynamic dashboard for credit utilization to closely monitor institutions with smaller unused credit lines, and conduct stress tests: if the largest creditor contracts by 30%, the remaining capacity will drop to 5.8%, necessitating advance planning for alternative financing channels. Current unused credit lines can cover approximately RMB 443 million in liquidity needs, but this must be matched with the scale of debt maturing in the next six months. It is recommended to optimize the pace of credit line usage, prioritizing the use of credit lines from non-major banks to reduce reliance on concentrated lending.

[0135] Risk control: It is also recommended to convert 50% of the unused credit line of joint-stock banks into liquidity standby agreements, explore cooperation with policy banks to optimize the credit structure, and pay attention to the non-performing loan ratio and industry correlation risks of regional banks such as a certain provincial bank.

[0136] Example 2: Automated review of "non-operating transaction risks"

[0137] This embodiment details how the present invention can be applied to reviewing a key financial risk in bond prospectuses: non-operating transaction risk. The entire process follows the five-layer progressive prompt framework proposed in this invention, transforming the complex review task into a precise and executable set of instructions. Figure 5 As shown, the specific steps include:

[0138] S1: Definition of Key Audit Points Driven by Domain Knowledge

[0139] The core task of this step is defined as: quantifying non-operating transaction risks and triggering tiered regulatory responses based on risk levels. This definition stems directly from financial experts' distillation of regulatory requirements and risk control practices. Based on this core task, experts further defined key quantitative standards, namely, risk grading thresholds. These thresholds are the cornerstone of the review and judgment process and are directly embedded in subsequent instructions: "High Risk" corresponds to a risk ratio greater than 10%, "Warning" to a risk ratio greater than 5%, "Watchlist" to a risk ratio greater than 3%, and "Safe" to a risk ratio less than or equal to 3%.

[0140] S2: Expert Logic Decomposition and Structuring

[0141] This step breaks down the expert's review process into a rigorous and unambiguous chain of logical judgments, ensuring the uniqueness and stability of the analysis path. It includes core data calculation logic, anomaly handling logic, and risk rating logic.

[0142] Core data calculation logic: Experts defined priority rules for calculating the total non-operating balance.

[0143] Priority 1: Total amount of non-operating receivables = Other non-operating receivables + Balance of inter-company loans.

[0144] Priority 2: If the data required for Priority 1 is incomplete, the alternative plan will be activated, and the total amount of non-operating intercompany receivables will be calculated as: non-operating intercompany receivables plus outstanding funds loans.

[0145] Exception handling logic: To avoid the model from becoming disillusioned when there is insufficient data, experts have set clear termination conditions: if there is no data to calculate for both priority 1 and priority 2, the analysis process must be terminated immediately and the unique text "No relevant data was extracted, unable to calculate and analyze" will be output.

[0146] Risk rating logic: After successfully calculating the total non-operating balance, the following logical steps are executed: First, calculate the core risk indicators:

[0147] Risk ratio = Total amount of non-operating receivables / Total audited assets at the end of the most recent year × 100%.

[0148] The calculated risk percentage is then compared with the four-level threshold defined in S1 to form a multi-branch judgment structure, thereby determining the final risk level.

[0149] S3: LLM Instruction Adaptation and Task Decomposition

[0150] This step transforms the structured expert logic into a sequence of instructions that a Large Language Model (LLM) can execute precisely. This instruction sequence comprises the following parts:

[0151] Data Extraction Instructions: The instructions explicitly require the LLM to extract a series of key fields from the input data, including multiple alternative field names such as total assets, non-operating receivables / payables, and inter-company loans, to address the differences in wording across different documents. The instructions also include specific data processing rules: if the inter-company loan balance is not extracted, it will be automatically assigned a value of 0.

[0152] Logical Analysis Instruction: This instruction explicitly defines the calculation logic of S2. The instruction requires the LLM to strictly adhere to the priority order: first attempt calculation using priority 1, and if that fails, then attempt priority 2. Simultaneously, the instruction includes exception handling logic; that is, if neither formula can be calculated, the process must terminate and a specified text output.

[0153] Risk Assessment Instruction: This instruction provides a fixed output text template for each risk level. For example, for the "High Risk" level, the instruction requires the model to return a string in the format "Percentage [Value]% > 10%, disclosure required: Debtor credit → Repayment arrangements → Necessity of borrowing → Impact on debt repayment".

[0154] LLM Analysis Instructions: These instructions are strictly tied to risk levels and are used to guide LLM in conducting in-depth value-added analysis. For example, when a risk level is determined to be "high-risk," the instruction will drive the model to perform additional analysis on the debt-to-equity ratios and compliance of the fund flows of the top three debtors; while when a risk level is determined to be "safe," the instruction will guide the model to propose management recommendations such as setting a balance restriction in the prospectus.

[0155] Regulatory Response Instruction: To ensure compliance during the review process, this instruction clearly defines the mandatory disclosure items that need to be matched according to different risk levels. For example, when the risk level is greater than 10%, the system is instructed to display four items: the debtor's qualifications, repayment arrangements, the necessity of the borrowing, and the impact on debt repayment.

[0156] S4: Standardized Output Specification Design

[0157] This step aims to ensure the consistency, usability, and professionalism of the system output. Based on the prompts, the final output must be in plain text format, and its content and structure must strictly adhere to the preset specifications. A complete output report includes the following sections:

[0158] The core data extracted: clearly list the raw data values ​​used for calculations.

[0159] Calculation results: The calculation results of the total non-operating balance and risk ratio are clearly displayed.

[0160] Analysis conclusion: Based on the risk ratio, strictly apply the risk assessment instruction template defined in step three to output a conclusion with risk level and regulatory prompts.

[0161] Commentary: Includes additional risk control perspective analysis content that matches the risk level.

[0162] Regulatory response: Disclosures required by regulatory response instructions that are appropriate to the risk level.

[0163] S5: Prompt word integration and placeholders

[0164] This step integrates all the instructions, rules, and specifications defined in the first four steps into a complete, reusable master prompt template. The core of this template lies in its parameterized design, with the most important placeholder being {{info}}. In the actual workflow, the system first extracts text and data fragments related to non-commercial transactions from the document to be reviewed, and then injects this content as a whole into the {{info}} parameter position. In this way, a general, static prompt template is combined with specific document content to form a dynamic prompt that can be executed immediately and contains complete context and task instructions.

[0165] The final complete prompt is as follows:

[0166] Task: Quantify the risks of non-operating transactions and trigger a tiered regulatory response.

[0167] Input data: {{info}}.

[0168] Output plain text: core extracted data, calculation results, analysis conclusions + commentary, showing the process.

[0169] Strictly follow these analytical steps:

[0170] Data Extraction: Extract the following key fields from the input data, and select data from the most recent fiscal year: total assets or total assets, non-operating receivables and payables or non-operating receivables and payables, inter-company loans, non-operating receivables and payables and inter-company loans, and other non-operating receivables.

[0171] Data processing rules: If the outstanding balance of the inter-company loan has not been withdrawn, it will be automatically set to 0; ensure that all withdrawal operations are strictly based on the input data.

[0172] Calculate: Total non-operating balance.

[0173] Calculation logic: A priority formula is used; calculation is performed as long as any condition is met.

[0174] Priority 1: Total amount of non-operating receivables = Other non-operating receivables + Balance of inter-company loans.

[0175] Priority 2: Total amount of non-operating intercompany receivables = non-operating intercompany receivables plus outstanding inter-company loans.

[0176] If neither formula has any data to calculate, only "No relevant data was extracted, unable to calculate and analyze" will be output. Core data, calculation results, analysis conclusions, and evaluations will not be output, and all subsequent analysis calculations and other steps will be skipped. At the same time, it is prohibited to provide other information such as examples, simulation processes, suggestions, and analyses.

[0177] Risk ratio = Total amount of non-operating receivables / Total audited assets at the end of the most recent year × 100%. Here, the total amount of non-operating receivables needs to be directly calculated from the previous formula.

[0178] Risk Assessment and Output: Output conclusions in descending order of risk level.

[0179] (1) Greater than 10%: High risk:

[0180] "If the percentage is greater than 10%, the following information needs to be disclosed: the debtor's creditworthiness, repayment arrangements, necessity of borrowing, and impact on debt repayment."

[0181] (2) Greater than 5%: Warning:

[0182] "Percentage [value]% > 5%, commitment: no new balance will be added during the term of the product".

[0183] (3) Greater than 3%: Pay attention to:

[0184] "Percentage [value] > 3%, Disclosure: Decision-making process → Top five related parties → Repayment plan"

[0185] (4) Less than or equal to 3%: Safe:

[0186] "The percentage is ≤3%, and the risk is controllable."

[0187] Risk control perspective analysis:

[0188] High risk: Verify the debt-to-asset ratio of the top three debtors, analyze the compliance of fund flows, and issue an alert when the debt-to-asset ratio is greater than 70%.

[0189] Warning: Assess the historical collection rate of related parties.

[0190] Focus: Examine the reasonableness of interest rate pricing.

[0191] Security: It is recommended that the offering letter include a balance restriction.

[0192] Regulatory Response Instructions: Mandatory Disclosure Items Based on Risk Level:

[0193] >10%: Debtor's creditworthiness + repayment arrangements + necessity of borrowing + impact on debt repayment.

[0194] >5%: Decision-making process + top five related parties + repayment plan.

[0195] >3%: Decision-making authority + pricing mechanism.

[0196] Risk analysis of non-operating transactions:

[0197] Based on data from the end of 2024, total assets amounted to RMB 9,513,902,300, and non-operating receivables and payables and inter-company loans totaled RMB 193,188,700. The total amount of non-operating receivables and payables equals the total amount of non-operating receivables and payables and inter-company loans, i.e., RMB 193,188,700. The risk ratio is calculated to be 2.03% ≤ 3%, indicating that the risk is controllable.

[0198] Analysis:

[0199] The risk ratio is calculated to be 2.03%, which is considered to be controllable. There is no mandatory disclosure requirement. It is recommended to set a ban on non-operating accounts receivable / payable in the prospectus to strengthen risk control.

[0200] Compliance check: Non-operating receivables have undergone the decision-making procedures of the Party Committee / Board of Directors and comply with the requirements of the "Funds Management Measures". Attention should be paid to the credit risk of the top five debtors, but the current proportion is dispersed, with each transaction ≤13.7%.

[0201] Fund Flow: The outstanding loan balance is 0, and the funds are mainly used for related-party transactions. It is necessary to verify whether the repayment plan matches the project cycle. It is recommended to supplement the ownership improvement plan for properties without certificates to avoid potential disputes affecting asset liquidity, and monitor the sales revenue collection progress of inventory development costs to ensure that cash flow covers potential repayment needs for non-operating transactions.

[0202] Key data shows that total assets amounted to RMB 9,513,902,300, and non-operating receivables totaled RMB 193,188,700.

[0203] Specific Implementation Example 3: Taking the Automated Audit of "Working Capital Stagnation Risk" as an Example

[0204] This embodiment aims to illustrate how the present invention can be applied to the scenario of reviewing the issuer's "working capital lock-up risk" in bond prospectuses, and to assess the liquidity and convertibility of its assets. The entire process follows the five-layer progressive prompt framework proposed in this invention, transforming the complex review task into a precise and executable set of instructions. Figure 6 As shown, the specific steps include:

[0205] S1: Definition of Key Audit Points Driven by Domain Knowledge

[0206] The core task of this step is precisely defined as: identifying the risk of working capital tied up in cash and assessing asset liquidity and convertibility. This definition is based on a deep understanding of the quality of a company's balance sheet and asset turnover efficiency. To achieve this task, financial experts have defined key quantitative analysis dimensions and risk thresholds:

[0207] The core risk concept introduces the operating asset ratio, which is the ratio of accounts receivable and inventory to total assets, as a core indicator to measure the degree of capital tied up.

[0208] Key risk thresholds have been set with clear warning standards, including: operating assets accounting for more than 70%, accounts receivable aged more than one year accounting for more than 30%, and finished goods accounting for more than 50% of total inventory.

[0209] The configurable industry benchmark introduces the industry average percentage as a configurable parameter to provide a horizontal comparison benchmark when assessing the issuer's operating asset structure.

[0210] S2: Expert Logic Decomposition and Structuring

[0211] This step breaks down the expert's thought process in assessing asset liquidity into a clear set of core indicator calculations and a risk assessment matrix, ensuring the rigor and hierarchy of the analytical logic. In the core indicator calculation logic section, the experts defined three essential financial ratios that must be calculated and clarified their calculation formulas:

[0212] 1. Operating assets ratio = (Accounts receivable + Inventory) / Total assets × 100%;

[0213] 2. Long-term receivables ratio = Receivables over 1 year old / Total receivables × 100%;

[0214] 3. Unsold inventory ratio = Finished goods / Total inventory × 100%.

[0215] In the risk assessment logic section, experts designed a descending risk assessment matrix that prioritizes immediate cessation of risk assessment, ensuring that the highest level of risk is identified first.

[0216] 1. High-risk status determination: First, check whether there is a situation with the highest risk, that is, the simultaneous occurrence of operating assets ratio >70% and long-term receivables ratio >40% or slow-moving inventory ratio >60%.

[0217] 2. Warning status determination: If the high-risk status has not been reached, then check whether any single indicator has reached the warning line, such as the ratio of operating assets > 70% or the ratio of long-term receivables > 30%.

[0218] 3. Safety Status Determination: If none of the above risk and warning conditions are met, the status is determined to be safe.

[0219] S3: LLM Instruction Adaptation and Task Decomposition

[0220] This step transforms the structured expert logic into a sequence of instructions that the Large Language Model (LLM) can execute precisely. This instruction set is clearly decomposed into the following parts:

[0221] Data extraction instructions: Explicitly require LLM to obtain the ending balances of receivables, inventory, and total assets from the consolidated balance sheet of the input data; and further require attempting to extract more granular information from the notes, such as the aging distribution of receivables and inventory categories, to support in-depth analysis.

[0222] Logical analysis instruction: Provide the calculation formulas for the three core indicators defined in S2 directly to the LLM, requiring it to strictly follow the formulas for calculation.

[0223] Risk assessment instruction: Transforms the risk assessment matrix of S2 into LLM executable conditional judgment logic, and provides fixed output text templates for high-risk, early warning, and safety risk levels. For example, for the high-risk status, the instruction requires the output format to be: "High asset liquidity risk: Operating asset ratio [value]% + [long-term / slow-moving] risk".

[0224] Analysis Instructions: These instructions are strictly tied to risk levels and are used to guide LLM (Local Management Analyst) to conduct value-added analysis from a professional risk control perspective after reaching initial conclusions. For example, when a high-risk situation is determined, the instruction will drive the model to conduct additional analysis on the solvency of large accounts receivable, the adequacy of inventory write-downs, and require them to submit recommendations for disclosing recovery plans and disposal schemes.

[0225] S4: Standardized Output Specification Design

[0226] This step aims to ensure the professionalism and consistency of the system's output. According to the prompt, a complete analysis report must include two main parts: risk assessment and in-depth analysis.

[0227] The risk assessment section strictly follows the risk assessment template defined in S3, and outputs a conclusive statement that includes the risk level, core indicator values, and risk summary.

[0228] Following the conclusion, an in-depth analysis and recommendations from a risk control perspective, commensurate with the risk level, are provided. For the output of the safety status, not only is a conclusion such as "Healthy operating asset structure: Percentage [value]%" given, but further comparisons with industry averages are provided, such as "(Industry average: 50%)", offering users richer decision-making information.

[0229] S5: Prompt word integration and placeholders

[0230] This step integrates all the instructions, rules, and specifications defined in the first four steps into a complete and reusable prompt template. The template's flexibility lies in its inclusion of a series of placeholders. {{info}}: Core input data parameters, used to dynamically inject financial statements and notes extracted from the documents to be reviewed during actual use; Industry average percentage allows users to input corresponding benchmark values ​​based on different industry backgrounds, making the analysis more valuable; High-risk threshold ensures the model can flexibly adapt to potentially changing internal risk control standards or regulatory requirements in the future.

[0231] The final complete prompt template, delivered to the LLM for execution, is as follows:

[0232] Task: Identify the risk of working capital stagnation and assess asset liquidity and convertibility.

[0233] Input data: {{info}}.

[0234] Parameter settings: industry average percentage =

[50] , high risk threshold = 70%.

[0235] Analysis steps:

[0236] 1. Data Extraction: Obtain from the consolidated balance sheet: ending accounts receivable, other receivables, long-term receivables, ending inventory balance, and ending total assets;

[0237] Obtain from the notes: the aging distribution of receivables older than 1 year, and inventory categories, including raw materials, work in progress, and finished goods.

[0238] 2. Core Computing:

[0239] (1) Operating asset ratio: (accounts receivable + inventory) / total assets × 100%, warning when operating asset ratio > 70%.

[0240] (2) Long-term receivables ratio: receivables over 1 year old / total receivables × 100%. Warning is issued when the long-term receivables ratio is > 30%.

[0241] (3) Unsold inventory ratio: finished goods / total inventory × 100%. Warning is issued when the unsold inventory ratio is > 50%.

[0242] Risk assessment matrix: Output in descending order of risk level, stop when the risk level is met.

[0243] 1. High-risk status:

[0244] (1) Operating assets ratio > 70% + long-term receivables ratio > 40%;

[0245] (2) Operating assets ratio > 70% + unsold inventory ratio > 60%;

[0246] If any of the above conditions are met, the output will be: "High asset liquidity risk: Operating assets as a percentage of [value] + [Long-term / Unsold inventory] risk".

[0247] 2. Warning Status: Operating assets ratio > 70%, long-term receivables ratio > 30%, slow-moving inventory ratio > 50%;

[0248] If any of the above conditions are met, the output will be: "Asset quality under pressure: Operating ratio [value]% / [Aging / Slow-moving goods] abnormal".

[0249] 3. Safety Status: Output: "Healthy operating asset structure: Percentage [value]% (Industry average: [reference])".

[0250] It also provides in-depth analysis from a risk control perspective:

[0251] High-risk status: Verify the debt repayment ability of large accounts receivable, calculate the adequacy of inventory write-down provisions, and require disclosure of: accounts receivable collection plan and inventory clearance plan.

[0252] Warning status: Analyze the concentration and payment cycle of the top five customers, and check whether there is any inflated revenue through accounts receivable.

[0253] Safety Status: Pay attention to the trend of accounts receivable turnover days.

[0254] The final output of this audit point is:

[0255] Analysis of the risks of working capital stagnation

[0256] 1. Data extraction and core indicator calculation

[0257] Data at the end of 2024:

[0258] Accounts receivable = Accounts receivable 8,845.22 + Other receivables 60,170.99 = 690,162,100 yuan;

[0259] Inventory = RMB 297,194,380;

[0260] Total assets = RMB 951,390,230;

[0261] Operating assets as a percentage of total assets = (69,016.21 + 297,194.38) / 951,390.23 × 100% = 38.5%;

[0262] Long-term receivables ratio (assuming no data in the notes, it cannot be calculated at this time);

[0263] The ratio of slow-moving inventory (assuming there is no inventory classification data, it cannot be calculated at this time);

[0264] Data as of the end of 2023, pro forma consolidated financial statements:

[0265] The percentage of operating assets is:

[0266] (1,772.92+58,454.02+194,331.58) / 810,045.66×100%=31.4%;

[0267] 2. Risk Assessment

[0268] Conclusion: The operating asset structure is healthy: it accounted for 38.5% at the end of 2024, which is lower than the industry average of 50%, and the pre-assessment ratio in 2023 was 31.4%, neither of which reached the 70% warning line.

[0269] 3. In-depth analysis

[0270] Structural highlights:

[0271] Ample liquidity buffer: Cash and cash equivalents surged 346% year-on-year, from RMB141 million to RMB627 million, coupled with a conservative allocation of no trading financial assets, indicating enhanced short-term debt repayment capacity.

[0272] Debt maturity optimization: The proportion of non-current liabilities due within one year increased from 0.6% to 15.7%, which may reflect proactive debt structure adjustments.

[0273] Potential risk points:

[0274] Potential risks from surging inventory: Inventory balance increased by 81.6% year-on-year, from 2.97 billion to 1.64 billion, a growth rate far exceeding that of total current assets, requiring further investigation.

[0275] Whether the transfer of construction in progress to fixed assets is due to capacity expansion, with an additional 852 million yuan in construction in progress;

[0276] To determine whether sluggish sales have led to an accumulation of finished products, additional inventory classification data is needed for verification.

[0277] Other receivables are unusual: they account for 6.3% of total assets (602 million yuan), nearly 7 times the total accounts receivable, requiring thorough investigation.

[0278] The possibility of related-party fund misappropriation is relatively low, compared to minority shareholders' equity of only 234 million.

[0279] Whether there are hidden non-operating transactions, such as shareholder loans or security deposits.

[0280] Industry comparison recommendations:

[0281] While the current operating asset ratio of 38.5% is lower than the set industry average of 50%, if the actual industry average is higher, such as the retail industry which is usually 60%+, the parameters need to be recalibrated.

[0282] Suggested additions:

[0283] Does the proportion of accounts receivable over one year exceed 20%?

[0284] Inventory turnover days compared with historical data and industry peers.

[0285] Example 4: Workflow of the Automated Review System

[0286] This embodiment is intended to combine Figure 1 This invention provides a detailed description of a method and system for reviewing bond issuance documents based on a large-scale intelligent model. This embodiment uses the review of a "green bond prospectus" regarding the "compliance of the use of proceeds" as an example. The entire process can be divided into two stages: the knowledge base construction stage and the review task execution stage.

[0287] Phase 1: Knowledge Base Construction

[0288] Before executing specific audit tasks, the system needs to build its core knowledge base through two parallel paths: a prompt word library and a location library. In this embodiment, a financial compliance expert needs to add a new audit rule for green bonds to build the prompt word library. The expert inputs a rule through the system interface, such as: "Verify whether the use of funds promised in the prospectus is completely consistent with the list of certified green projects in the appendix." After this new audit rule is uploaded to the system, it is processed by the rule specification agent. This agent transforms the natural language rule input by the expert into a structured, standardized audit rule containing clear analytical logic and output format requirements, i.e., a prompt word template, and stores it in the prompt word library for later use. To let the system know where to find relevant information, the system administrator uploads a large number of historical "green bond prospectus" documents as training samples to build the location library. The location statistics agent starts and analyzes these documents. Through statistical learning, it finds that detailed descriptions of "use of proceeds" usually appear in a chapter titled "Chapter 5, Use of Proceeds," while the "list of certified green projects" is most likely located in "Appendix II." The location information obtained from these statistics is integrated into content location statistics and stored in a location database, forming a knowledge index about the location of information.

[0289] Phase Two: Execution of Review Tasks

[0290] Once the knowledge base is built, the system can execute online automated review tasks. An auditor uploads a new "Green Bond Prospectus" to the system and triggers a review instruction. The system's intent recognition module receives the instruction and determines the current task is "review." Subsequently, the product type recognition module analyzes the document content and accurately identifies the document as belonging to the "Green Bond" type. After recognition, the review agent is activated. Based on the product type "Green Bond," it determines that "compliance of the use of proceeds" is one of the core review points that must be performed. The review agent simultaneously accesses two knowledge bases built in the first phase: from the prompt word library, it retrieves a prompt word template matching the "compliance of the use of proceeds" review point; from the location library, it obtains the location of the information required to complete this review, namely "Chapter 5" and "Appendix II." According to the instruction obtained from the location library, the review agent automatically and accurately extracts all the text content of "Chapter 5" and the item list of "Appendix II" from the uploaded document. It combines these dynamically extracted contents with the static prompt word template obtained from the prompt word library to generate a dynamic prompt word containing complete context and a clear analysis task. The dynamic suggestion is sent to the backend analytics engine (LLM). The engine compares and analyzes the extracted content according to the instructions, and finally generates the analysis results.

[0291] The present invention provides a method and system for reviewing bond product issuance documents based on a large-scale intelligent agent, which has the following advantages:

[0292] 1. Improve review efficiency and reduce review costs: Through automated processes, reviewers are freed from heavy and repetitive document review work, and the review time is shortened from several days to several minutes, which greatly reduces labor costs and improves business processing capabilities.

[0293] 2. Ensure standardization and consistency in audits: All audits are performed based on a unified, expert-defined rule base, eliminating the subjectivity and individual differences of manual audits, ensuring a high degree of consistency in audit standards for the same type of product, and improving the reliability of risk control.

[0294] 3. High flexibility and scalability: By decoupling the review logic from information location, the system possesses strong adaptability. When new bond products emerge or regulatory provisions change, the system's review capabilities can be quickly expanded simply by adding new rules or documents to the corresponding knowledge base through the intelligent agent. This requires no modification to the core code, resulting in low maintenance costs and rapid iteration.

[0295] 4. Realize the accumulation and reuse of expert knowledge: Solidify the valuable experience and knowledge of top audit experts into calculable and executable digital assets, realize the accumulation, inheritance and large-scale reuse of high-value knowledge, and bring a qualitative improvement to the overall risk management capabilities of the institution.

[0296] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; only preferred embodiments of the present invention are illustrated. The descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. As long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0297] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for reviewing bond product issuance documents based on a large-scale intelligent agent model, characterized in that, Includes the following steps: Step S1: Construct an intelligent agent through the review rules to transform unstructured review knowledge into structured review rules and store them in the prompt word library. Perform statistical analysis on historical documents through the content location statistics intelligent agent to generate content location information and store it in the location library. Step S2: Obtain the bond product issuance documents to be reviewed, identify the bond product type of the documents to be reviewed, and match the corresponding set of review points according to the bond product type; Step S3: For each audit point in the audit point set, perform the following operations: retrieve the corresponding structured audit rules from the prompt word library according to the audit point, retrieve the corresponding content positioning information from the location library, and extract content fragments related to the audit point from the document to be audited according to the content positioning information; Step S4: Combine the structured review rules with the content fragments to generate dynamic prompt words for LLM (Large Language Model) processing. Send the dynamic prompt words to the LLM. The LLM analyzes the content fragments based on the dynamic prompt words and outputs the review results corresponding to the review points.

2. The method for reviewing bond product issuance documents based on a large model intelligent agent according to claim 1, characterized in that: The step S1 of transforming unstructured audit knowledge into structured audit rules includes the following steps: decomposing the unstructured audit knowledge into data analysis paths and logical judgment rules; transforming the logical judgment rules into a multi-step instruction set executable by the large language model; defining a standardized format for the output of the large language model; and integrating the multi-step instruction set with the standardized format to obtain structured audit rules.

3. The method for reviewing bond product issuance documents based on a large model intelligent agent according to claim 1, characterized in that: The statistical analysis of historical documents by the content positioning statistical agent described in step S1 includes the following steps: performing statistical analysis on the chapters, titles, and table names of the historical documents, and summarizing the location patterns of the information required for the review points in the documents to be reviewed.

4. The method for reviewing bond product issuance documents based on a large model intelligent agent according to claim 1, characterized in that: Step S3 involves extracting content fragments related to the review points from the document to be reviewed based on the content location information, including the following steps: extracting the content fragments from specific chapters, specific titles, or specific tables of the document to be reviewed based on the content location information stored in the location library.

5. The method for reviewing bond product issuance documents based on a large model intelligent agent according to claim 1, characterized in that: The dynamic prompt word mentioned in step S4 includes several placeholders. Content fragments are filled into the placeholders of the dynamic prompt word to generate a complete dynamic prompt word.

6. The method for reviewing bond product issuance documents based on a large model intelligent agent according to claim 1, characterized in that: In step S4, LLM analyzes the content segment based on the dynamic prompt words, including the following steps: performing a preset quantitative calculation on the content segment to obtain a risk index, comparing the risk index with a preset risk threshold, and determining the risk level.

7. The method for reviewing bond product issuance documents based on a large model intelligent agent according to claim 1, characterized in that: The audit results corresponding to the audit points mentioned in step S4 are texts that conform to a preset standard format, which includes conclusions, core indicators, and in-depth analysis.

8. The method for reviewing bond product issuance documents based on a large model intelligent agent according to claim 1, characterized in that: The set of audit points mentioned in step S2 includes at least one of the following types of points: corporate governance and organizational structure, debt, cash flow, profit, assets, business, or compliance and quality.

9. A system for reviewing bond issuance documents based on a large-scale intelligent agent model, implemented based on the method for reviewing bond issuance documents based on a large-scale intelligent agent model as described in any one of claims 1-8, characterized in that, include: The module includes a knowledge construction module, a database module, a document processing module, a content extraction module, a dynamic prompt word generation module, and a large language model processing module. The knowledge construction module transforms unstructured audit knowledge and historical documents into structured data and stores it in a database. The database module includes a prompt word library and a location library. The prompt word library is used to store structured review rules, with each rule associated with a bond product type and review points. The location library is used to store content location information, including descriptions of common locations of each review point in the document. The document processing module processes newly input documents awaiting review and determines the key points requiring review. The content extraction module extracts relevant content fragments from the document to be reviewed for each review point. The dynamic prompt word generation module combines the review rules and content fragments to generate dynamic prompt words for each review point, and sends the dynamic prompt words to the large language model; The large language model processing module analyzes content fragments based on dynamic prompts to determine whether they are compliant or have problems.

10. The bond product issuance document review system based on a large model intelligent agent according to claim 9, characterized in that: The knowledge construction module includes an audit rule construction agent and a content positioning and statistics agent. The audit rule construction agent is used to parse unstructured audit knowledge, extract key rules through natural language processing technology, and transform the key rules into structured audit rules. The content positioning and statistics agent is used to perform statistical analysis on historical documents, identify common content patterns and data points of different types of bond products, and generate content positioning information.

Images