Multi-agent-based intelligent financial document writing method and system

Abstract

The invention provides an intelligent bond project establishment application report compiling method and system based on multi-agent cooperative work, and the method and system employ an architecture based on multi-agent cooperative work, can simulate the working process of an actual operator, and cooperatively complete the task decomposition, data arrangement, content generation and structure verification of financial document compiling work. The automation level is improved. A model agent for simulating the division roles of six financial project groups is designed, and the model agent comprises a project manager, data investigation, material examination, content writing, quality control auditing and content correction. Each role plays different functions and tasks in the research and report writing process, uses different tools, and is mutually matched from multiple dimensions, so that the quality and efficiency of the bond project establishment application report are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of financial data processing, and specifically relates to a multi-agent-based intelligent financial document writing method and system. BACKGROUND

[0002] In the financial market, bonds, as an important financing tool, play an indispensable role in corporate financing, government financing, and other fields. Before issuing bonds, investment banks need to write a detailed bond project application report for the proposed issuer to assess the feasibility, compliance, and risk level of the project. This report is the core basis for project review, approval, and subsequent issuance processes, and its content covers multiple professional dimensions such as issuer background, fund use, financial status, industry analysis, and risk warnings, which puts high demands on the professional ability and writing efficiency of project personnel.

[0003] However, in the traditional process, the writing of the bond project application report still mainly relies on manual operation, which gradually exposes many problems in the current environment of high-speed changes in financial information and parallel project advancement:

[0004] Information update lag: The report involves a large amount of data and industry material sorting, and manual processing is low in efficiency, making it difficult to reflect market changes and the latest policy requirements in a timely manner.

[0005] Heavy workload: From data collection, material analysis to document writing, writers often need to invest a lot of time and effort.

[0006] Quality varies: The depth and accuracy of the report content are highly dependent on the personal experience and professional ability of project personnel, making it difficult to unify standards;

[0007] Complex document structure: Bond reports need to cover multiple dimensions, with complex structure, making writing and proofreading prone to errors, affecting compliance and approval efficiency.

[0008] Currently, although natural language processing (NLP) technology has been widely applied, mainly focusing on basic tasks such as financial information extraction and public opinion analysis, there are still significant bottlenecks in the generation of structured and complex financial documents. In recent years, the development of large language models (LLM) has made it possible for intelligent document writing, which performs well in language understanding and generation. However, directly applying current large models to bond report writing still faces many challenges:

[0009] Limited semantic context: The one-time generation capability of large models is limited, making it difficult to ensure logical coherence and semantic consistency in long documents.

[0010] Industry information is outdated: The training corpus of large models is mostly based on past data, lacking the integration of real-time financial dynamics and policies and regulations, leading to hallucinations and biases in large models.

[0011] Task boundaries are blurred: Single large models struggle to balance data processing, industry analysis, and report writing, among other tasks.

[0012] Lack of integration of text and graphics: Traditional large models struggle to automatically construct charts, data summaries, risk ratings, and other business intelligence (BI) elements.

[0013] Single perspective: Single large models often lack multi-perspective analysis and cross-validation, affecting the completeness of reports and the value of decision-making references.

[0014] Currently, there have been explorations of large models in the financial field, such as Bloomberg GPT, FinGPT, and PIXIU. However, these systems mainly focus on task evaluation, question answering, sentiment analysis, and other sub-tasks, and still lack support for document-level content understanding and structured generation. Additionally, current financial institutions mainly use "single model calling" for large models, without forming a collaborative cognitive and role division mechanism, making it difficult to adapt to the needs of complex document generation tasks.

[0015] Therefore, there is an urgent need for a systematic solution that can coordinate various financial sub-tasks to provide financial institutions with efficient, accurate, and compliant intelligent writing capabilities for financial documents such as bond project application reports. SUMMARY

[0016] To solve the above problems, the present application provides an intelligent bond project application report writing method and system based on multi-agent collaborative work to realize more efficient and intelligent automatic writing of financial documents. The present application is a systematic solution that can collaborate with various financial sub-tasks, adopts a multi-agent collaborative work-based architecture, can simulate the actual investment bank personnel's work process, collaboratively complete the task decomposition, data organization, content generation and structure review of financial document writing work, and improve the automation level. Among them, six model agents simulating the division of labor roles of financial project groups are designed, including project manager, data investigation, material review, content writing, quality control review and content correction. Each role plays a different function and task in the report writing process, uses different tools, and cooperates from multiple dimensions to improve the quality and efficiency of the bond project application report. Among them, the tools that these roles can use include surfing the Internet, calling the python interpreter, drawing charts, etc. The use of these tools helps the role to integrate data more timely and comprehensively and ensures that the bond project application report accurately extracts the issuer's company business data, audit data and business development related data from multiple dimensions, ensuring that the bond project application report objectively evaluates the comprehensive information of the issuer. At the same time, each agent collaborates to realize the phased writing of multiple chapters of the bond project report, thereby avoiding the length limitation of the context, maintaining the coherence of the semantics of the bond project report, and the cognitive collaboration of multiple agents is expected to improve the comprehensive cognition of complex financial problems, thereby further improving the quality and depth of the bond project application report.

[0017] Specifically, the present application adopts the following technical solutions:

[0018] The present application provides an intelligent financial document writing method based on multi-agent, which has the technical features that the method comprises the following steps: step S1, knowledge structured extraction is performed on the draft file data of a financial project to obtain structured financial draft data, and a corresponding vectorized knowledge base is constructed; step S2, based on the vectorized knowledge base, a financial document required by the financial project is written by multi-agent collaborative work. Among them, the multi-agent is used to simulate the division of labor structure of the investment bank project team, including a project manager agent, a data investigation agent, a material review agent, a content writing agent, a quality control review agent and a content correction agent.

[0019] The multi-agent-based intelligent financial document writing method provided by the application can also have the following technical features: step S1 includes the following sub-steps: S1-1, integrating the draft file data by using a draft file system to obtain integrated data; S1-2, comprehensively analyzing the integrated data by using a word segmenter, a document parser, and an OCR scanner to obtain parsed data; S1-3, detecting and screening the parsed data by using a language detection screening algorithm, and simultaneously detecting the parsed data by using a repeated detection algorithm to filter redundant content to obtain structured financial draft data; S1-4, constructing a semantic label system of the core elements of the financial project, and setting labels for the structured financial draft data based on the semantic label system; and S1-5, constructing the vectorized knowledge base of the financial project based on the structured financial draft data with labels.

[0020] The multi-agent-based intelligent financial document writing method provided by the application can also have the following technical features: in S1-2, the word segmenter is a word segmenter based on a vector semantic model. In S1-3, the language detection screening algorithm is an algorithm of a language detection mechanism based on the fusion of rules and a FastText model, which screens non-Chinese text and unstructured data in the parsed data, and only retains structured and language-complete Chinese data as the structured financial draft data; and the repeated detection algorithm is a MiniHash algorithm, which detects repeated paragraphs in the parsed data to filter redundant content.

[0021] The multi-agent-based intelligent financial document writing method provided by the application can also have the following technical features: the writing process of the financial document includes a design stage, a writing stage, an auditing stage, and an archiving stage. In the design stage, the chapter task of writing the financial document is disassembled and a chapter-level writing outline is generated by the project manager agent. In the writing stage, the content writing agent is dominant, and the material investigation agent and the material review agent are coordinated to complete the generation of each chapter draft. In the auditing stage, the quality control and auditing agent is dominant, and the content correction agent is coordinated to complete the content correction of each chapter draft to obtain each chapter finished product. In the archiving stage, the project manager agent is dominant, and each chapter finished product is combined to generate a complete financial document.

[0022] The multi-agent-based intelligent financial document writing method provided by the application can also have the following technical features: step S2 includes the following sub-steps: S2-1, in the design stage, the project manager agent disassembles the chapter tasks of the financial document according to the key information of the financial project provided by the user and a standardized template, and formulates a chapter-level writing outline; S2-2, in the design stage, the data investigation agent collects supplementary materials required for writing the financial document from external data sources according to the outline and the key information; S2-3, in the writing stage, the content writing agent plays a core role and cooperates with the data investigation agent and the material review agent to generate each chapter draft according to the chapter tasks and the outline, wherein the material review agent is used for standardizing and integrating underlying data and consistency checking; S2-4, in the review stage, the quality control review agent reviews the semantic coherence, account index calculation correctness and coverage of regulatory requirements of each chapter draft, and generates corresponding quality control feedback information; S2-5, in the review stage, the content revision agent revises the content of each chapter draft according to the quality control feedback information to generate each chapter finished product; S2-6, in the archiving stage, the project manager agent merges and arranges each chapter finished product according to the outline, and combines each chapter finished product to generate a complete financial document.

[0023] The multi-agent-based intelligent financial document writing method provided by the application can also have the following technical features: the project manager agent is realized by a question and answer model based on a retrieval enhancement generation mechanism combined with a knowledge graph, the question and answer model is constructed by a large language model based on a predetermined financial document structure and compliance specification corpus through instruction prompt engineering optimization, so that the project manager agent has project coordination and industry understanding ability. The data investigation agent is composed of a retriever module and a retrieval engine of the vectorized knowledge base. The content writing agent is realized by a large language model based on a retrieval enhancement generation mechanism, which generates each chapter draft based on the retrieval results of the retrieval enhancement generation mechanism, the chapter tasks and the outline, and the context arrangement of the prompt engineering. The material review agent, the quality control review agent and the content revision agent are all constructed by a large language model through instruction prompt engineering optimization.

[0024] The intelligent financial document writing method based on multiple agents provided by the application can also have the technical features that, in the process of generating the chapter draft by the content writing agent, multi-round interactive prompting and fragment rewriting are supported to ensure the language accuracy and logical reasonableness of the chapter draft. The content writing agent also generates chart content through a corresponding auxiliary tool module, and the generated chart content is output synchronously with the chapter text.

[0025] The intelligent financial document writing method based on multiple agents provided by the application can also have the technical features that, in the process of generating the chapter draft by the content writing agent, multi-round interactive prompting and fragment rewriting are supported to ensure the language accuracy and logical reasonableness of the chapter draft. The content writing agent also generates chart content through a corresponding auxiliary tool module, and the generated chart content is output synchronously with the chapter text.

[0026] The application provides an intelligent financial document writing system based on multiple agents, which has the following technical features: a vectorized knowledge base is constructed based on structured financial draft data obtained by knowledge structuring extraction on draft file data of a financial project; and a multiple-agent collaboration module is used to collaboratively write a financial document required by the financial project based on the vectorized knowledge base. The multiple agents are used to simulate the division of labor structure of a project team of an investment bank, including a project manager agent, a data investigation agent, a material review agent, a content writing agent, a quality control and review agent, and a content correction agent.

[0027] Effects of the application

[0028] Compared with the traditional financial document writing method relying on manual work, the intelligent financial document writing method and system based on multiple agents provided by the application has the following beneficial effects:

[0029] Greatly improve the writing efficiency: through task decomposition and multiple-agent parallel processing, the automatic writing of the bond project report is realized, the writing period is significantly shortened, and the manual input is reduced.

[0030] Improve the report quality and compliance: the division of labor of each agent is clear, and each agent has the ability of information retrieval, data analysis, chart generation, and semantic proofreading, which effectively ensures the logical rigor, financial accuracy, and compliance and standardization of the report content.

[0031] Optimize the context coherence and global consistency: through the context sharing mechanism supported by the vectorized knowledge base, the information consistency and semantic coherence are maintained between the agents, and the limitations of traditional large models in processing long documents are overcome. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 is a schematic diagram of the intelligent financial document writing method based on multiple agents in the embodiment of the application;

[0033] Figure 2 is a flow chart of the multi-agent based intelligent financial document writing method in the embodiments of the present application;

[0034] Figure 3 is a flow chart of step S1 in the embodiments of the present application;

[0035] Figure 4 is a schematic diagram of a typical division of labor structure of an investment bank project team in the prior art;

[0036] Figure 5 is a flow chart of step S2 in the embodiments of the present application;

[0037] Figure 6 is a structural block diagram of the multi-agent based intelligent financial document writing system in the embodiments of the present application. DETAILED DESCRIPTION

[0038] In order to make the technical means, creative features, purposes and effects achieved by the present application easy to understand, the multi-agent based intelligent financial document writing method and system of the present application will be specifically described below in conjunction with embodiments and drawings.

[0039] <EMBODIMENT>

[0040] Figure 1 and Figure 2 are a schematic diagram and a flow chart of the multi-agent based intelligent financial document writing method in the embodiments, respectively.

[0041] As shown in Figure 1 and Figure 2 , the multi-agent based intelligent financial document writing method of the embodiments comprises the following steps:

[0042] Step S1, knowledge structured extraction is performed on the draft file data of the financial project to obtain structured financial draft data, and a corresponding vectorized knowledge base is constructed.

[0043] Step S2, based on the vectorized knowledge base, the financial document required for the financial project is written by multiple agents working collaboratively.

[0044] Exemplarily, the following will take a financial project of bond establishment and a financial document of bond establishment application report to be written as an example to explain the above steps in detail. However, the method and system of the present application are not limited thereto, and can also be applied to the writing of financial documents of other types of projects by analogy.

[0045] Step S1, knowledge structured extraction is performed on the draft file data of the financial project to obtain structured financial draft data, and a corresponding vectorized knowledge base is constructed.

[0046] In the process of writing the bond project application report, the report content usually covers the issuer background, financial indicator analysis, project purpose description, industry environment assessment, credit risk disclosure and other aspects. These contents not only depend on the project personnel's proficiency in basic information, but also require them to have high sensitivity and analytical judgment to the latest market dynamics, policy changes and enterprise operating conditions. Especially in the bond project submission link, the project draft as the key work underlying material, its integrity and accuracy directly determine the quality and compliance of the final document.

[0047] In order to solve the problems of high manual dependence, poor timeliness and chaotic content structure in the traditional draft information collection process, a structured financial project draft extraction method is proposed in this embodiment. With the help of large language model and lightweight classification model, real financial project materials are deeply processed to extract high-quality structured semantic fragments, thereby providing a stable foundation for subsequent intelligent document generation.

[0048] Figure 3 The flowchart of step S1 in this embodiment is shown in FIG. 1.

[0049] As shown in FIG. 1, step S1 specifically includes the following sub-steps: Figure 3

[0050] Step S1-1, integrating the draft file data by using the draft file system to obtain integrated data.

[0051] Among them, the draft file system may be, for example, the existing draft file system of Guotai & Junan Securities.

[0052] Step S1-2, using a tokenizer, a document parser and an OCR scanner to comprehensively analyze the integrated data to obtain parsed data.

[0053] Among them, the tokenizer is a tokenizer based on a vector semantic model. Taking the bond project application report as an example, the issuer's project draft is comprehensively analyzed in this step.

[0054] Step S1-3, detecting and screening the parsed data by a language detection screening algorithm, and simultaneously detecting the parsed data for redundancy by a repeated detection algorithm to filter out redundant content and obtain structured financial draft data.

[0055] Among them, the language detection screening algorithm is an algorithm based on the fusion of rule and FastText model language detection mechanism, which automatically screens the non-Chinese text and unstructured data in the parsed data, and only retains the structured and semantically complete Chinese bond material data as the above structured financial draft data. The repeated detection algorithm is a MiniHash algorithm, which detects the paragraphs of text in the parsed data for redundancy to filter out highly redundant content and improve the diversity and representativeness of the sample corpus.​

[0056] Step S1-4, a semantic label system of core elements of the financial project is constructed, and the structured financial draft data is labeled based on the semantic label system.

[0057] For example, for a bond project, the semantic label system of its core elements mainly targets the issuer's business information, financial audit status, industry to which the issuer belongs, and related important subsidiaries. Based on these labels, intelligent scheduling of multiple agents (Agents) can be realized to complete the corresponding content generation and provide support for subsequent comprehensive evaluation.

[0058] Step S1-5, a vectorized knowledge base of the financial project is constructed based on the structured financial draft data labeled.

[0059] Step S2, based on the vectorized knowledge base, the financial documents required for the financial project are written by multiple agents working collaboratively.

[0060] For example, for a bond project, under the support of structured bond project draft data, a multi-agent collaborative solution is designed for the task of writing a bond project application report, which is complex in content, strict in format, and clear in compliance requirements. This embodiment simulates the typical division of labor structure of the investment bank project team to efficiently generate professional reports that meet regulatory standards from multiple perspectives.

[0061] Figure 4 is a schematic diagram of the typical division of labor structure of the investment bank project team in the prior art.

[0062] As shown in Figure 4 , the roles in the traditional investment bank project team usually include project managers (Design part in the figure), data investigation and audit personnel (Compile part in the figure), quality control audit personnel (Examination part in the figure), and content summary writers (Summary part in the figure).

[0063] As shown in Figure 1 , in the method and system of the embodiment, the typical division of labor structure of the traditional investment bank project team is simulated, and six types of agents (Agents) are preset, each type of agent has a clear functional division and tool capability, as follows:

[0064] Project Manager Agent: The Project Manager Agent is implemented by combining a question-answering model based on the Retrieval Augmentation Generation (RAG) mechanism with knowledge graph reasoning capabilities. The core question-answering model is optimized by large language models based on pre-defined financial document structures (such as bond project application report structures) and compliance specification corpus through instruction prompting engineering. The Project Manager Agent uses customized prompts and standardized templates to generate chapter task decomposition solutions after retrieving the underlying knowledge in the vectorized knowledge base, and to develop the outline of financial documents (such as bond project application reports). The optimized prompting engineering for each outline is read in, and the corresponding prompting engineering is distributed for the next step of unit decomposition to achieve content writing. Therefore, it has strong project coordination and industry understanding capabilities.

[0065] Investigation Agent: The Investigation Agent is composed of a Retriever module with retrieval capabilities and the retrieval engine of the above-mentioned vectorized knowledge base. The Agent is familiar with various external data sources and is responsible for collecting supplementary materials needed for writing financial documents from external data sources, such as bond project application reports. Supplementary materials can include issuance-related information, industry affiliation, industry news, and corporate announcements.

[0066] Review Agent: The Review Agent is proficient in financial and compliance information recognition and data structuring, responsible for standardizing and integrating underlying data such as business registration, financial statements, and audit reports, and conducting consistency checks. The Review Agent is also responsible for reviewing the search results from the underlying knowledge sources by the Investigation Agent to ensure the accuracy of the search vector results and thus ensure accurate content generation. Similar to the Project Manager Agent, the Review Agent can be optimized by large language models through instruction prompting.

[0067] Writing Agent: The Writing Agent is implemented by a large language model based on the RAG mechanism. It calls the vectorized knowledge base based on the RAG mechanism and combines the large language model capability to complete the first draft writing of each chapter of the financial document. Specifically, the Agent provides the retrieval results of the RAG mechanism to the large language model, which generates the first draft of each chapter based on the retrieval results, the chapter tasks generated by the Project Manager Agent, and the outline, using context arrangement based on prompt engineering. During the generation process of the large language model, multi-round interactive prompts and piecewise rewriting are supported to ensure the language accuracy and logical rationality of the first draft. In addition, the Agent also undertakes the generation of chart content in the financial document, and implements drawing tasks such as asset and liability structure chart and solvency analysis chart through corresponding auxiliary tool modules (such as Python interpreter calling matplotlib library), and the generated charts are output synchronously with the chapter text.

[0068] Reflection Agent: The Reflection Agent takes the logic and compliance of the document as the core, and is responsible for checking the semantic coherence of each chapter of the first draft of the financial document, the correctness of the accounting index calculation, and the coverage of the regulatory requirements, and generating corresponding quality control feedback information. Similar to the Project Manager Agent, the Reflection Agent can be optimized and constructed by a large language model based on instruction prompts.

[0069] Revise Agent: The Revise Agent refines the language, logic, and format of each chapter of the first draft of the financial document generated according to the quality control feedback information, to ensure the uniformity of the text style and the compliance of the professional expression of the financial document. Similar to the Project Manager Agent, the Revise Agent can be optimized and constructed by a large language model based on instruction prompts.

[0070] In addition, in this embodiment, the Agents also include a Unit Parser, a Process Subtask, and a Refine query. The Unit Parser is used to analyze the chapter tasks decomposed by the Project Manager Agent and pass the information that needs to be supplemented to the Material Investigation Agent. The Process Subtask is used to check and judge whether the supplementary materials required for writing each chapter have been obtained, and pass the corresponding check and judgment results to the Unit Parser or the Writing Agent. The Refine query is used to generate an optimized query for the Material Review Agent. These Agents can be implemented using existing technologies, so they will not be described in detail.

[0071] The plurality of agents can communicate with each other to form an agent network, thereby cooperating as needed during the writing process.

[0072] Based on the plurality of agents, in the method and system of the embodiment, a multi-stage, agent role clear writing process is designed, and the overall process is divided into four key stages: design, writing, review, and archiving. Each stage is led by a specific agent and collaborates with each other to advance. Among them, the design stage is led by the project manager agent, the writing stage is led by the content writing agent, the review stage is led by the quality control agent, and the archiving stage is led by the project manager agent.

[0073] Figure 5 is a flowchart of step S2 in the embodiment.

[0074] As Figure 5 shown, based on the agent role division and writing process design, step S2 specifically includes the following sub-steps:

[0075] Step S2-1, in the design stage, the project manager agent disassembles the specific chapter tasks of the required financial document according to the key information of the financial project provided by the user and the standardized template, and formulates a chapter-level writing outline.

[0076] Among them, in the design stage, everything starts from the initial demand of the user. For bond project, the key information provided by the user includes the proposed issuer, financing amount, product type, issuance place, etc.

[0077] Step S2-2, in the design stage, the data investigation agent collects the supplementary materials required for writing the financial document from external data sources according to the outline and the key information provided by the user.

[0078] Among them, for the bond project, the data investigation agent quickly collects the relevant policy background of the issuer, macro-industry information, regional support documents, etc., as data support for subsequent content generation.

[0079] Step S2-3, in the writing stage, the content writing agent plays a core role, and jointly with the data investigation agent and the material review agent, generates the first draft of each chapter of the financial document according to the chapter task and the outline.

[0080] The material investigation agent obtains the data of the industry status and policy dynamics by using the draft data system. The material review agent calls the business platform and financial data system to process the enterprise operation and account structure information. The content writing agent calls the RAG module to match the relevant knowledge text in the vectorized knowledge base based on the chapter task and outline, and organizes the language and generates the first draft of each chapter. At the same time, the content writing agent can call various auxiliary tool modules during content writing, such as calling the Python interpreter to calculate the debt repayment ability, calling the chart generation template to draw the asset-liability graph or operating cash flow trend graph, etc. These auxiliary tools are triggered and executed by the system unified scheduling engine (such as LangGraph or custom Executor framework), realizing context controllability and task traceability.

[0081] In step S2-4, in the review stage, the quality control review agent reviews the semantic coherence of the first draft of each chapter, the correctness of the account index calculation, and the coverage of the regulatory requirements, and generates corresponding quality control feedback information.

[0082] The quality control review agent performs multi-dimensional checks on the first draft of each chapter, including semantic consistency, calculation accuracy, reference data timeliness, and text language specification. At the same time, it also checks whether there are sensitive expressions that violate regulatory rules, missing risk prompts, etc. in the first draft of each chapter. Through strict quality review, it is ensured that the final financial document is not only logically smooth, but also compliant.

[0083] In step S2-5, in the review stage, the content revision agent revises the first draft of each chapter according to the quality control feedback information, and generates the final draft of each chapter.

[0084] In step S2-6, in the archiving stage, the project manager agent merges and arranges the final draft of each chapter according to the outline, and combines the final draft of each chapter to generate a complete financial document.

[0085] After generating the complete financial document, the system supports one-click export to a standard format (such as PDF / Word) file, so as to facilitate subsequent uploading on the regulatory platform, internal approval or customer submission.

[0086] In the entire financial document generation process as described above, each agent shares context information through the vectorized knowledge base to ensure the consistency of the content between chapters and the semantic continuity, solving the semantic discontinuity problem caused by the context window limitation of traditional large models. At the same time, the multi-agent (multi-role) collaborative work also significantly improves the task generalization ability and complex document processing ability of the system.

[0087] The embodiment also provides a system corresponding to the above method.

[0088] Figure 6 is a structural block diagram of the multi-agent based intelligent financial document writing system in this embodiment.

[0089] As shown in Figure 6 , the multi-agent based intelligent financial document writing system 10 includes a vectorized knowledge base (database) 10, a multi-agent collaborative writing module 11, and a control module 12. Among them, the vectorized knowledge base 10 is constructed by the above step S1, and the multi-agent collaborative writing module 11 writes the financial document according to the method of the above step S2. The control module 12 is used to control the reading and writing of the vectorized knowledge base 10 and the work of the multi-agent collaborative writing module 11.

[0090] The multi-agent collaborative writing module 11 at least includes a project manager agent 111, a data investigation agent 112, a material review agent 113, a content writing agent 114, a quality control audit agent 115, a content correction agent 116, an auxiliary tool set 117, a unified scheduling engine 118, and a process management unit 119. Among them, the division of labor of each agent is as described above. The auxiliary tool set 117 includes a plurality of auxiliary tool modules, including the above-mentioned Python interpreter and library, etc., and the unified scheduling engine 118 is used for the auxiliary tools required by each agent. The process management unit 119 controls the work of each agent according to the sub-step process of the above step S2.

[0091] Effects of the embodiment

[0092] According to the multi-agent based intelligent financial document writing method and system provided in this embodiment, compared with the traditional financial document writing method relying on manual work, the following advantages are obtained:

[0093] Greatly improve the writing efficiency: through task decomposition and multi-agent parallel processing, realize the automatic writing of the bond project report, significantly shorten the writing period and reduce the manual input.

[0094] Improve the quality and compliance of the report: each agent has a clear division of labor and has the ability of information retrieval, data analysis, chart generation, and semantic proofreading, which effectively guarantees the logical rigor, financial accuracy, and compliance and standardization of the report content.

[0095] Enhance timeliness and response capability: with the help of RAG mechanism, real-time access to policy, industry and enterprise dynamic information ensures that the generated financial document content is close to the current market environment and avoids information lag.

[0096] Optimize the coherence and global consistency of the context: through the context sharing mechanism supported by the vectorized knowledge base, the information consistency and semantic coherence between each agent are maintained, overcoming the limitations of traditional large models in processing long documents.

[0097] The system architecture is flexible, and since the on-demand calling of the modules of the multiple auxiliary tools is realized through the unified scheduling engine, the intelligent agent or the new auxiliary tool module can be conveniently added according to actual business requirements, the system is suitable for intelligent writing scenes of various structured financial documents, and has wide popularization value.

[0098] The above examples are only used for illustrating the specific embodiments of the present application, and the present application is not limited to the description range of the above examples. It should be understood by those skilled in the art that the present application is not limited by the above examples, and the above examples and the description in the specification only illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A multi-agent based intelligent financial document drafting method, characterized in that, The method comprises the following steps: Step S1, knowledge structuring extraction is performed on the draft file data of the financial project to obtain structured financial draft data, and a corresponding vectorized knowledge base is constructed; Step S2, based on the vectorized knowledge base, a financial document required by the financial project is written by multiple agents working cooperatively, Wherein, the multiple agents are used to simulate the division structure of the project team of the investment bank, including a project manager agent, a data investigation agent, a material review agent, a content writing agent, a quality control review agent and a content correction agent.

2. The intelligent financial document writing method based on multiple agents according to claim 1, wherein: wherein Step S1 comprises the following sub-steps: S1-1, the draft file data is integrated by using a draft file system to obtain integrated data; S1-2, the integrated data is comprehensively parsed by using a word segmenter, a document parser and an OCR scanner to obtain parsed data; S1-3, the parsed data is detected and screened by a language detection screening algorithm, and the parsed data is repeatedly detected by a repeated detection algorithm to filter redundant content to obtain structured financial draft data; S1-4, a semantic label system of core elements of the financial project is constructed, and the structured financial draft data is labeled based on the semantic label system; S1-5, the vectorized knowledge base of the financial project is constructed based on the structured financial draft data with labels.

3. The intelligent financial document writing method based on multiple agents according to claim 2, wherein: wherein In S1-2, the word segmenter is a word segmenter based on a vector semantic model, In S1-3, the language detection screening algorithm is an algorithm of a language detection mechanism based on the fusion of rules and a FastText model, which screens non-Chinese text and unstructured data in the parsed data, and only retains structured and language complete Chinese data as the structured financial draft data; the repeated detection algorithm is a MiniHash algorithm, which repeatedly detects paragraph-level text in the parsed data to filter redundant content.

4. The intelligent financial document writing method based on multiple agents according to claim 1, wherein: wherein The writing process of the financial document comprises a design stage, a writing stage, an audit stage and an archiving stage, In the design stage, the project manager agent is dominant, the chapter task of writing the financial document is disassembled and a chapter-level writing outline is generated, In the writing stage, the content writing agent is dominant, and the data investigation agent and the material review agent are cooperated to complete the generation of each chapter draft, In the audit stage, the quality control review agent is dominant, and the content correction agent is cooperated to complete the content correction of each chapter draft to obtain each chapter finished product, In the archiving stage, the project manager agent is dominant, and each chapter finished product is combined to generate a complete financial document.

5. The intelligent financial document writing method based on multiple agents according to claim 4, wherein: wherein, Step S2 includes the following sub-steps: S2-1, in the design phase, the project manager agent disassembles the chapter tasks of the financial document according to the key information of the financial project provided by the user and the standardized template, and formulates a chapter-level writing outline; S2-2, in the design phase, the data investigation agent collects supplementary materials required for writing the financial document from external data sources according to the outline and the key information; S2-3, in the writing phase, the content writing agent plays a core role and cooperates with the data investigation agent and the material review agent to generate each chapter draft according to the chapter tasks and the outline, wherein the material review agent is used for standardized integration and consistency checking of underlying data; S2-4, in the review phase, the quality control review agent reviews the semantic coherence, account index calculation correctness and coverage of regulatory requirements of each chapter draft, and generates corresponding quality control feedback information; S2-5, in the review phase, the content revision agent revises the content of each chapter draft according to the quality control feedback information to generate each chapter finished product; S2-6, in the archiving phase, the project manager agent merges and arranges each chapter finished product according to the outline, and combines each chapter finished product to generate a complete financial document.

6. The multi-agent-based intelligent financial document writing method according to claim 5, wherein: wherein the project manager agent is implemented by a question and answer model based on a retrieval enhancement generation mechanism combined with a knowledge graph, the question and answer model is constructed by a large language model based on a predetermined financial document structure and compliance specification corpus through instruction prompting engineering optimization, so that the project manager agent has project coordination and industry understanding ability, the data investigation agent is composed of a retriever module and a retrieval engine of the vectorized knowledge base, the content writing agent is implemented by a large language model based on a retrieval enhancement generation mechanism, which generates each chapter draft based on the retrieval results of the retrieval enhancement generation mechanism, the chapter tasks and the outline, and uses context arrangement of prompting engineering, the material review agent, the quality control review agent and the content revision agent are all constructed by a large language model through instruction prompting engineering optimization.

7. The multi-agent-based intelligent financial document writing method according to claim 6, wherein: wherein during the generation of the chapter draft by the content writing agent, multi-round interactive prompting and fragment rewriting are supported to ensure the language accuracy and logical rationality of the chapter draft, the content writing agent also generates chart content through a corresponding auxiliary tool module, and the generated chart content is output synchronously with the chapter text.

8. The multi-agent-based intelligent financial document writing method according to claim 1, wherein: wherein the financial project is a bond project, the financial document to be written is a bond project application report.

9. A multi-agent based intelligent financial document drafting system, characterized in that, including: The vectorized knowledge base is constructed based on structured financial draft data obtained by knowledge structuralization extraction on draft file data of the financial project; And The multi-agent collaborative writing module is configured to write financial documents required by the financial project by multi-agent collaboration based on the vectorized knowledge base. The multi-agent is configured to simulate the division of labor structure of the project team of the investment bank, including a project manager agent, a data investigation agent, a material review agent, a content writing agent, a quality control review agent, and a content correction agent.

Images