Work report generation system and method based on multi-path retrieval and large model

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a government work report generation system and method based on multi-path retrieval and a large model, and the system comprises a data engineering module, a database system module, a multi-path retrieval module, a cue word engineering module and a report generation system module. The system has the beneficial effects that database management and work report generation capabilities are considered, a user not only can store, manage and check own government affair data, but also can be linked to a report generation module to generate a government work report based on a knowledge base constructed by a client, and the data accuracy and personalization of report content are greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a work report generation system and method based on multi-path retrieval and large models. Background Art

[0002] With the rapid development of large language models, the field of text generation has flourished, with a wide variety of writing tools emerging. The quality of official document writing directly impacts the image of government agencies, the implementation of decisions, and the effectiveness of their work. Its accuracy, logic, and standardization are crucial, reflecting the authority and professionalism of the agency. With the rapid development of artificial intelligence technologies such as large models and generative AI, official document writing, long considered a serious discipline, is undergoing a profound, technology-driven transformation.

[0003] Work reports are a key use case for official document writing. Currently, there are several products on the market that assist with writing work reports, such as those from iFlytek and Xinhua News Agency. These products can assist civil servants in writing reports and streamline their workflow. However, some issues remain, such as incorrect data, short length, and unprofessional sentence structure. Summary of the Invention

[0004] The purpose of the present invention is to provide a work report generation system and method based on multi-way retrieval and large model to solve the problems raised in the above background technology.

[0005] To achieve the above-mentioned purpose, the present invention provides the following technical solution: a work report generation system based on multi-path retrieval and large model, comprising:

[0006] The data engineering module is used to perform general cleaning and data formatting on documents generated in the workflow of government institutions, generating standardized data containing fields such as time, type, title, text, outline, summary, affiliated unit, keywords, and confidentiality level;

[0007] The database system module is used to store the processed data in a vector database and a full-text retrieval database, and to build a rights management and data management system based on roles and document classification;

[0008] Multi-way retrieval module, which is used to retrieve documents related to the input topic from the database through the fusion of full-text retrieval and vector retrieval combined with the re-ranking model;

[0009] The prompt word engineering module is used to configure the prompt words for the outline, paragraphs, foreword, and conclusion elements required to generate work reports;

[0010] The report generation system module is used to obtain materials through multi-channel retrieval based on the subject, keywords, time, issuing unit and length entered by the user, generate outlines and paragraphs in combination with prompt word engineering, and finally integrate them into a complete work report.

[0011] Preferably, the data formatting processing in the data engineering module specifically includes: using a large model to perform semantic understanding of the document, extracting and generating document type, outline, summary, affiliated unit, and keyword field information; setting a fixed field format, including time, title, text, and confidentiality level, where the time field is the issuance time, the title field is the document name, the text field is the document body content, and the confidentiality level field is divided into top secret, confidential, secret, and general files.

[0012] Preferably, the data storage and indexing process in the database system module specifically includes: preprocessing the data to ensure that the data format and content meet the requirements of subsequent processing; dividing the preprocessed data into two processing paths: on the one hand, converting the data into vector form through vector extraction technology and storing it in a vector database to support efficient semantic search; on the other hand, extracting structured information from the data and storing it in a full-text retrieval database to realize full-text retrieval function; building a data management system, including a permission management module and a document classification display module. The permission management module is responsible for defining and assigning user roles to ensure that different users obtain appropriate access rights according to their roles; the document classification display module realizes document classification and more efficient retrieval through interaction with the full-text retrieval database and the vector database.

[0013] Preferably, the multi-way fusion retrieval method in the multi-way retrieval module specifically includes: full-text retrieval, segmenting the documents and building an inverted index, scoring the queries according to the BM25 algorithm, and finding the best matching documents; vector retrieval, using the embedding model to convert documents and queries into high-dimensional dense vectors, and finding the document vector closest to the query in the vector space by calculating the inner product; fusion retrieval, performing weighted averaging on the results of full-text retrieval and vector retrieval, assigning different weights and then fusing them to obtain the final score and output it; reranking, using the reranker model to rerank the fusion retrieval results, placing the best matching documents in the front position, so as to reduce the noise in the large model input and improve the accuracy of the generated report.

[0014] Preferably, the report generation process in the report generation system module specifically includes:

[0015] The user enters the subject, keywords, time, publishing unit and length;

[0016] Generate outline prompts based on the outline, generate the outline of the work report in three steps through the large model, and support users to manually modify the outline content;

[0017] The outline is split into first-level, second-level, and third-level headings and processed to form structured data containing the title, heading level, first-level heading, whether it contains subheadings, and filler word number fields;

[0018] Based on the parsed outline, use the multi-channel search module to search for materials, and summarize and refine the search content as reference materials for filling in titles and generating paragraphs;

[0019] Based on the results of outline parsing and information retrieval, paragraph generation prompt words are used to generate paragraph content through the large model;

[0020] Generate paragraph content in sequence according to the title order of the outline, and add standard content such as title, preface, conclusion, etc. to finally form a complete work report.

[0021] A method for a work report generation system based on multi-path retrieval and large models, comprising the following steps:

[0022] Data engineering processing: Perform general cleaning and data formatting on documents generated in the workflow of government institutions, remove useless characters, summarize and classify data, and set fields, including time, type, title, text, outline, summary, affiliated unit, keywords, and confidentiality level; among them, the document type, outline, summary, affiliated unit, and keyword fields are obtained through the big model;

[0023] Database system construction: store the processed data in the database and build data management and rights management systems;

[0024] Multi-way retrieval and recall: A multi-way fusion retrieval method based on a rearrangement model is used, including full-text retrieval, vector retrieval, fusion retrieval, and re-ranking steps;

[0025] Work report prompt word project configuration: debug prompt words for work report related elements, including title, outline, paragraphs, foreword and conclusion;

[0026] Report generation, including user input, outline generation, outline parsing, information retrieval, paragraph generation, and report body.

[0027] Preferably, in the data engineering processing step, general cleaning specifically involves removing a large number of useless characters contained in some samples in the data, including invalid spaces, line breaks, HTML characters, and characters that are completely unrelated to official documents; data formatting specifically involves combining the excellent semantic understanding ability of the large model and metadata content to summarize and classify the data, and set the fields according to the fields set in Table 1.

[0028] Preferably, in the multi-way retrieval recall step, the specific process of full-text retrieval is: segmenting the document and building an inverted index, then segmenting the query and matching keywords, scoring according to the BM25 algorithm, and finding the best matching document; the specific process of vector retrieval is: using the embedding model, converting the document and query into high-dimensional dense vectors, storing them in the vector space, and finding the document vector closest to the query in the vector space by calculating the inner product; the specific process of fusion retrieval is: normalizing the results of full-text retrieval and vector retrieval, assigning weights, and then weighted fusion to obtain the final score and output it; the specific process of reranking is: using the reranker model to place the best matching document as close to the front as possible, reducing the document noise input to the large model, and improving the accuracy.

[0029] Preferably, in the prompt word project configuration step, the prompt word for the outline's previous years' work summary requires the title format to be "I. Review of 20XX's Work", with the time being the year before the time in the theme. The secondary title must contain specified keywords, and there should be two secondary titles, respectively, focusing on gratitude for the hard work of the staff in the region and welcoming new challenges in the future, without sub-titles. The title format and number are clearly specified; the prompt word for this year's overall work requirements and plans requires the title format to be "II. 20XX...", focusing on the overall work requirements, work goals and work plans, with the time being the corresponding time in the theme, and with only one title without sub-titles; this year The work arrangement prompt requires the title format to be "III. Work Arrangements in 20XX", focusing on the key work of this year. The time is the corresponding time in the theme. The secondary title must contain specified keywords, and the title format and quantity are clearly stipulated; the paragraph prompt requires the designated title to be filled with content based on the reference content, maintaining coherence with the previous text, not outputting repeated events and data, and only supplementing the content of the designated title. It should be written using imperative sentences and subject-verb grammar, including specified keywords, and filled according to the title background and paragraph word count requirements. It should reasonably use reference content data and event information to generate a concise and concise paragraph in the tone of official government documents.

[0030] Preferably, in the report generation process steps, the outline parsing step performs data processing according to the set outline parsing fields; the information retrieval step is based on the parsed outline, and uses a fusion multi-way retrieval method based on the rearrangement model to retrieve and summarize the materials for each title to provide reference materials for paragraph generation; the paragraph generation step is based on the outline parsing and information retrieval results, uses paragraph generation prompt words, and generates paragraph content through the large model; the report body step generates paragraph content in the order of outline titles, and adds standard content of title, preface, and conclusion to form a complete work report body.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] The work report generation system and method based on multi-channel retrieval and large models proposed in the present invention take into account both database management and work report generation capabilities. Users can not only store, manage and view their own government data, but also link to the report generation module to generate work reports based on the knowledge base built by the customer, greatly improving the data accuracy and personalization of the report content. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of data storage and indexing in the present invention;

[0034] Figure 2 This is the architecture diagram of the data management system of the present invention;

[0035] Figure 3 This is a multi-path search framework diagram of the present invention;

[0036] Figure 4 Generate a system flow chart for this invention report;

[0037] Figure 5 This is a system architecture diagram of the present invention. DETAILED DESCRIPTION

[0038] In order to clearly and completely describe the objectives and technical solutions of the present invention and make the advantages more clearly understood, the embodiments of the present invention are further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are part of the embodiments of the present invention, not all of them, and are only used to explain the embodiments of the present invention, not to limit the embodiments of the present invention. All other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0039] For example 1, please refer to Figures 1 to 5 The present invention provides a technical solution: a work report generation system based on multi-channel retrieval and a large model. Based on the acquired data, the system, through a series of data engineering processes, stores the processed data in a database (vector database, full-text retrieval database), and adds functions such as permission management and document classification. Relevant content is recalled through multi-channel retrieval, combined with prompt word engineering, and work report elements are generated using a large model. Finally, the above processes are logically connected to form a complete system. Specifically, the system can be divided into: a data engineering module, a database system module, a multi-channel retrieval module, a prompt word engineering module, and a report generation system module.

[0040] Data Engineering Module:

[0041] Data is the document generated in the workflow, and its quality determines the accuracy and reliability of the data in the final report.

[0042] (1) General cleaning

[0043] Some samples in the data contain a large number of useless characters, such as invalid spaces, line breaks, HTML characters, and characters that are completely unrelated to official documents. This content will affect the output effect of the model.

[0044] (2) Data formatting

[0045] Combining the excellent semantic understanding ability of the large model and metadata content, the data is summarized and classified. The specific fields are set as follows:

[0046] Table 1 Data field settings

[0047]

[0048] Among them, the document type, outline, abstract, affiliated unit and keyword fields are obtained using the large model.

[0049] Database system module:

[0050] This module stores the processed data in the database and builds a data management and rights management system.

[0051] (1) Data storage

[0052] 1) Data storage and indexing process

[0053] First, the data undergoes a preprocessing phase to ensure that its format and content meet the requirements for subsequent processing. Next, the data is divided into two processing paths: on the one hand, the data is converted into vector form using vector extraction technology and stored in a vector database to support efficient semantic search, allowing users to search based on the semantic content of documents; on the other hand, structured information in the data is extracted and stored in Elasticsearch to implement full-text search capabilities, allowing users to quickly find relevant documents using keywords. This design enables the system to handle both complex semantic queries and traditional keyword search needs, providing users with comprehensive and flexible data retrieval services. This coexistence of structured and unstructured data processing can fully leverage the advantages of different databases and improve data retrieval efficiency and accuracy.

[0054] 2) Document Classification

[0055] The classification system is based on the unit and type of the document. Through the dual classification of the unit and article type, efficient management and accurate retrieval of documents can be achieved, providing users with more convenient and personalized services.

[0056] Document Affiliation: A key dimension of document classification, it reflects the document's origin and management. Based on the level and function of government agencies, document affiliations can be categorized as follows: central agencies, local agencies (documents issued by these agencies are primarily focused on local administration and services), and special agencies.

[0057] Document type: reflects the nature and purpose of its content. Based on common types of government documents, documents can be divided into various types, such as "notice," "request for instructions," "report," "decision," "meeting minutes," and "speech."

[0058] (2) Data management system

[0059] In this system, the permission management module is responsible for defining and assigning user roles, ensuring that different users obtain appropriate access rights based on their roles; the document classification and display module achieves document classification and more efficient retrieval through interaction with Elasticsearch and vector databases.

[0060] Multi-way search module:

[0061] This technology uses a multi-way fusion retrieval method based on a rearrangement model to address the ineffectiveness of a single retrieval strategy. By integrating full-text retrieval with vector retrieval, it leverages the advantages of both keyword matching and semantic matching to retrieve documents that better match the input from massive amounts of data, thereby providing more effective information. The framework is as follows:

[0062] Full-text search: segment the document and build an inverted index. Then segment the query and match the keywords, score them according to the BM25 algorithm, and find the best matching document.

[0063] Vector retrieval uses an embedding model to convert documents and queries into high-dimensional dense vectors, which are stored in a vector space. By calculating the inner product, it searches for the document vector closest to the query in the vector space, prioritizing semantic similarity. Even if the words or phrases differ, documents with similar meanings can be retrieved.

[0064] Fusion retrieval filters and combines the results of full-text and vector searches through a weighted average. The document scores from both methods are normalized, weighted, and then combined to produce a final score, which is then output. This method combines the strengths of keyword and semantic matching, significantly improving search performance and enabling a more comprehensive search for relevant documents.

[0065] Reranking: Due to the limited input length and analytical capabilities of large models, the quality of the input data determines the quality of the output. Using a reranker model, we can place the most matching documents as close to the top as possible, reducing the number of documents fed into the large model and improving accuracy.

[0066] Prompt word engineering module:

[0067] Prompt words affect the quality and effectiveness of the elements related to the work report generated by the large model, so they need to be debugged, including elements such as the title, outline, paragraphs, foreword, and conclusion. The prompt word configuration of the main elements is as follows:

[0068] (1) Outline

[0069] The outline is divided into three parts: a summary of previous years' work, this year's overall requirements and plans, and this year's work arrangements. The prompts after debugging are as follows:

[0070] 1) Summary of previous years' work

[0071] [Based on the given topic and keywords, and referring to some given outline materials, write the first-level title and sub-titles of the table of contents outline of the official document type "Work Report" in accordance with the given requirements, ensuring the organization and logic of the writing process.

[0072] The following are the requirements given:

[0073] 1. The title format is "I. Review of work in 20XX", which is mainly a summary of last year's work. Note that the time of this title is one year before the time in "{input topic}".

[0074] 2. The secondary title must include the following keywords: "macroeconomic development, rural areas, cities, environment, people's livelihood, scientific and technological innovation, industrial development, high-quality development, transformation, reform, industrial structure, modern industrial system, {enter keywords}";

[0075] 3. Finally, there should be two sub-headings, one focusing on gratitude for the efforts of the staff in the region and the other on welcoming new challenges in the future. These two sub-headings should not have sub-headings.

[0076] 4. The format of the secondary heading is "(I)", "(II)", and so on, and the number shall not exceed eight;

[0077] 5. The format of the third-level title is "1.", "2." and so on, and the number shall not exceed two;

[0078] 6. There is a line break between each title. Do not output extra line breaks.

[0079] 7. Please output the outline content directly, do not summarize;

[0080] Here are the topics given: "{enter topic}";

[0081] The following are the reference materials:

[0082] "{Input reference material}"

[0083] 2) Overall requirements and plans for this year's work

[0084] [Based on the given topic and in accordance with the given requirements, write the second first-level title of the table of contents outline of the official document type "Work Report" to ensure the organization and logic of the writing process.

[0085] The following are the requirements given:

[0086] 1. The title format is "II. 20XX...", and the content mainly revolves around "overall work requirements, work goals and work plans". Please rewrite it appropriately. Please note that the time of the title is the corresponding time in "{input subject}".

[0087] 2. The section has only one title and no subtitles;

[0088] 3. Please output the title directly without summarizing;

[0089] The following are the given topics: "{enter topic}"]

[0090] 3) This year's work arrangements

[0091] [Based on the given topics and keywords, and referring to some given outline materials, write the third first-level title and its sub-titles of the table of contents outline of the official document type "Work Report" in accordance with the given requirements, ensuring the organization and logic of the writing process.

[0092] The following are the requirements given:

[0093] 1. The title format is "III. Work arrangements for 20XX", which mainly revolves around the work priorities of this year. Note that the time of the title is the corresponding time in "{input topic}".

[0094] 2. The secondary title must include the following keywords: "rural, urban, environment, people's livelihood, scientific and technological innovation, industrial development, high-quality development, transformation, reform, industrial structure, modern industrial system, {enter keywords}";

[0095] 3. The format of the secondary heading is "(I)", "(II)", and so on, and the number shall not exceed eight;

[0096] 4. The format of the third-level title is "1.", "2." and so on, and the number shall not exceed two;

[0097] 5. There is a line break between each title. Do not output extra line breaks.

[0098] 6. Please output the outline content directly, do not summarize;

[0099] Here are the topics given: "{enter topic}";

[0100] The following are the reference materials:

[0101] "{Input reference material}"

[0102] (2) Paragraph

[0103] Paragraphs are used to fill in the content of each heading in the outline. The prompt words are as follows:

[0104] [Please fill in the content of the specified excerpt outline based on the newly given reference content and the given requirements.

[0105] The following are the requirements given:

[0106] 1. Please keep the content consistent with the above, and be careful not to output events and data that are repeated in the above;

[0107] 2. Please only supplement the specified title. Do not output additional titles. When supplementing, please output the original title and its serial number in their original form and separate them from the original title.

[0108] 3. Please use imperative sentences, using verb-object grammar. Do not mention specific places or government units. Instead, use "our district," "our city," or "our province."

[0109] 4. Please include the following keywords when filling in: "{enter keywords}";

[0110] 5.{Enter title background};

[0111] 6. The content should be rich and full, {enter the required number of words for the paragraph};

[0112] 7. Please select and use the reference content reasonably based on the given keywords, and try to retain the data and event information of the reference content;

[0113] 8. Use an official tone, keep the sentences concise, and try to fill in a whole paragraph. Do not use transition words such as "first, second, first, second, last, etc."

[0114] The following is the designated title given: "{enter designated title}";

[0115] The following is the reference content given: "{input topic}, {input reference material}";

[0116] The following is the given context: "{enter context}"]

[0117] Report generation system modules:

[0118] This module introduces the process arrangement and implementation system of report generation. The flow chart is as follows:

[0119] (1) User input

[0120] Users enter relevant content based on the front-end page, including subject, keywords, time, publishing unit and length.

[0121] (2) Generate an outline

[0122] Use the corresponding prompt words for generating an outline to generate the outline of the work report in three steps based on the large model, and support users to manually change the outline content.

[0123] (3) Outline Analysis

[0124] The outline is split into first-level, second-level, and third-level headings and processed to form structured data. The fields are set as follows:

[0125] Table 2 Outline parsing field settings table

[0126]

[0127] (4) Information Retrieval

[0128] Based on the parsed outline, a fusion multi-way retrieval method based on a rearrangement model is used to retrieve materials for each title, and the retrieved content is summarized and refined as reference material when filling in the title and generating paragraphs.

[0129] (5) Paragraph Generation

[0130] Based on the results of outline parsing and information retrieval, paragraph generation prompt words are used to generate paragraph content through a large model.

[0131] (6) Report Content

[0132] Generate paragraph contents in sequence according to the title order of the outline, and add other contents required by the work report standards (title, preface, conclusion, etc.) to finally form the main body of the report.

[0133] Example 2, based on Example 1, a method for generating a work report based on multi-way retrieval and a large model, comprising the following steps:

[0134] Data engineering processing: Perform general cleaning and data formatting on the documents generated in the workflow of government institutions, remove useless characters, summarize and classify the data, and set fields, including time, type, title, text, outline, summary, affiliated unit, keywords, and confidentiality level; among them, the document type, outline, summary, affiliated unit, and keyword fields are obtained through the big model; in the data engineering processing steps, general cleaning specifically removes a large number of useless characters contained in some samples in the data, including invalid spaces, line breaks, HTML characters, and characters that are completely unrelated to official documents; data formatting specifically combines the excellent semantic understanding ability of the big model and metadata content to summarize and classify the data, and set the fields according to the fields set in Table 1.

[0135] Database system construction: store the processed data in the database and build data management and rights management systems;

[0136] Multi-way retrieval recall: A multi-way fusion retrieval method based on a rearrangement model is adopted, including full-text retrieval, vector retrieval, fusion retrieval and re-ranking steps; the specific process of full-text retrieval is: segment the document and build an inverted index, then segment the query and match the keywords, score according to the BM25 algorithm, and find the best matching document; the specific process of vector retrieval is: use the embedding model to convert the document and query into high-dimensional dense vectors, store them in the vector space, and calculate the inner product to find the document vector closest to the query in the vector space; the specific process of fusion retrieval is: normalize the results of full-text retrieval and vector retrieval, assign weights, and then perform weighted fusion to obtain the final score and output it; the specific process of re-ranking is: use the reranker model to place the best matching document as close to the front as possible, reduce the document noise input to the large model, and improve the accuracy.

[0137] Configuration of work report prompt words project: Prompt words are debugged for relevant elements of the work report, including title, outline, paragraphs, foreword and conclusion; the prompt words for the outline's summary of previous years require the title format to be "I. Review of Work in 20XX", with the time being the year before the time in the theme. The secondary title must contain specified keywords, and there should be two secondary titles, respectively focusing on gratitude for the hard work of the staff in the region and welcoming new challenges in the future, without sub-titles. The title format and number are clearly specified; the prompt words for this year's overall work requirements and plans require the title format to be "II. 20XX...", focusing on the overall work requirements, work goals and work plans, with the time being the time in the theme. The time should be the same, and there should be only one title without sub-titles; the work arrangement prompt for this year requires the title format to be "III. Work Arrangements for 20XX", which should revolve around the key work of this year. The time should be the corresponding time in the theme. The secondary title should contain specified keywords, and the title format and quantity are clearly specified; the paragraph prompt requires the designated title to be filled with content around the reference content, maintaining consistency with the previous text, not outputting repeated events and data, only supplementing the content of the designated title, using imperative sentences and subject-verb grammar, including specified keywords, filling in according to the title background and paragraph word count requirements, reasonably using reference content data and event information, and generating a concise and concise paragraph in the tone of an official document.

[0138] Report generation includes user input, outline generation, outline parsing, information retrieval, paragraph generation and report body; the outline parsing step performs data processing according to the set outline parsing fields; the information retrieval step is based on the parsed outline, using a fusion multi-way retrieval method based on the rearrangement model to retrieve and summarize materials for each title to provide reference materials for paragraph generation; the paragraph generation step is based on the outline parsing and information retrieval results, using paragraph generation prompt words, and generating paragraph content through a large model; the report body step generates paragraph content in the order of outline titles, and adds standard content of title, foreword, and conclusion to form a complete work report body.

[0139] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A work report generation system based on multi-channel retrieval and large model, characterized by: include: The data engineering module is used to perform general cleaning and data formatting on documents generated in the workflow of government institutions, generating standardized data containing fields such as time, type, title, text, outline, summary, affiliated unit, keywords, and confidentiality level; The database system module is used to store the processed data in a vector database and a full-text retrieval database, and to build a rights management and data management system based on roles and document classification; Multi-way retrieval module, which is used to retrieve documents related to the input topic from the database through the fusion of full-text retrieval and vector retrieval combined with the re-ranking model; The prompt word engineering module is used to configure the prompt words for the outline, paragraphs, foreword, and conclusion elements required to generate work reports; The report generation system module is used to obtain materials through multi-channel retrieval based on the subject, keywords, time, issuing unit and length entered by the user, generate outlines and paragraphs in combination with prompt word engineering, and finally integrate them into a complete work report.

2. A work report generation system based on multi-path retrieval and large model according to claim 1, characterized in that: The data formatting processing in the data engineering module specifically includes: using a large model to semantically understand the document, extracting and generating document type, outline, summary, affiliated unit, and keyword field information; setting fixed field formats, including time, title, text, and confidentiality level, where the time field is the issuance time, the title field is the document name, the text field is the document content, and the confidentiality level field is divided into top secret, confidential, secret, and general files.

3. A work report generation system based on multi-path retrieval and large model according to claim 2, characterized in that: The data storage and indexing process in the database system module specifically includes: preprocessing the data to ensure that the data format and content meet the requirements of subsequent processing; dividing the preprocessed data into two processing paths: on the one hand, converting the data into vector form through vector extraction technology and storing it in the vector database to support efficient semantic search; on the other hand, extracting structured information from the data and storing it in the full-text retrieval database to realize full-text retrieval function; building a data management system, including a permission management module and a document classification and display module. The permission management module is responsible for defining and assigning user roles to ensure that different users obtain appropriate access rights according to their roles; the document classification and display module realizes document classification and more efficient retrieval through interaction with the full-text retrieval database and the vector database.

4. A work report generation system based on multi-path retrieval and large model according to claim 3, characterized in that: The multi-way fusion retrieval method in the multi-way retrieval module specifically includes: full-text retrieval, which segments documents and builds an inverted index, scores queries according to the BM25 algorithm, and finds the best-matching documents; vector retrieval, which uses the embedding model to convert documents and queries into high-dimensional dense vectors, and calculates the inner product to find the document vector closest to the query in the vector space; fusion retrieval, which performs weighted averaging on the results of full-text retrieval and vector retrieval, assigns different weights, and then fuses them to obtain the final score and output it; reranking, which uses the reranker model to rerank the fusion retrieval results, placing the best-matching documents in the front position to reduce noise in the large model input and improve the accuracy of generated reports.

5. A work report generation system based on multi-path retrieval and large model according to claim 4, characterized in that: The report generation process in the report generation system module specifically includes: The user enters the subject, keywords, time, publishing unit and length; Generate outline prompts based on the outline, generate the outline of the work report in three steps through the large model, and support users to manually modify the outline content; The outline is split into first-level, second-level, and third-level headings and processed to form structured data containing the title, heading level, first-level heading, whether it contains subheadings, and filler word number fields; Based on the parsed outline, use the multi-channel search module to search for materials, and summarize and refine the search content as reference materials for filling in titles and generating paragraphs; Based on the results of outline parsing and information retrieval, paragraph generation prompt words are used to generate paragraph content through the large model; Generate paragraph content in sequence according to the title order of the outline, and add standard content such as title, preface, conclusion, etc. to finally form a complete work report.

6. A method for a work report generation system based on multi-path retrieval and large model according to claim 5, characterized in that: The following steps are involved: Data engineering processing: Perform general cleaning and data formatting on documents generated in the workflow of government institutions, remove useless characters, summarize and classify data, and set fields, including time, type, title, text, outline, summary, affiliated unit, keywords, and confidentiality level; among them, the document type, outline, summary, affiliated unit, and keyword fields are obtained through the big model; Database system construction: store the processed data in the database and build data management and rights management systems; Multi-way retrieval and recall: A multi-way fusion retrieval method based on a rearrangement model is used, including full-text retrieval, vector retrieval, fusion retrieval, and re-ranking steps; Work report prompt word project configuration: debug prompt words for work report related elements, including title, outline, paragraphs, foreword and conclusion; Report generation, including user input, outline generation, outline parsing, information retrieval, paragraph generation, and report body.

7. A method according to claim 6, characterized in that: In the data engineering processing steps, general cleaning specifically involves removing a large number of useless characters contained in some samples in the data, including invalid spaces, line breaks, HTML characters, and characters completely unrelated to official documents; data formatting specifically involves combining the excellent semantic understanding capabilities of the large model and metadata content to summarize and classify the data, and set the fields according to the fields set in Table 1.

8. A method according to claim 7, characterized in that: In the multi-way retrieval recall step, the specific process of full-text retrieval is as follows: segment the document and build an inverted index, then segment the query and match keywords, score according to the BM25 algorithm, and find the best matching document; the specific process of vector retrieval is as follows: use the embedding model to convert the document and query into high-dimensional dense vectors, store them in the vector space, and calculate the inner product to find the document vector closest to the query in the vector space; the specific process of fusion retrieval is as follows: normalize the results of full-text retrieval and vector retrieval, assign weights, and then perform weighted fusion to obtain the final score and output it; the specific process of reranking is as follows: use the reranker model to place the best matching document as close to the front as possible, reduce the document noise input to the large model, and improve accuracy.

9. A method according to claim 8, characterized in that: In the prompt word project configuration step, the prompt word for the outline's previous years' work summary requires the title format to be "I. Review of 20XX's Work", with the time being the year before the time in the theme. The secondary title must contain the specified keywords, and there should be two secondary titles, one focusing on the efforts of the staff in the region and the other on meeting new challenges in the future, without sub-titles. The title format and number are clearly specified; the prompt word for this year's overall work requirements and plans requires the title format to be "II. 20XX...", focusing on the overall work requirements, work goals and work plans, with the time being the corresponding time in the theme, and with only one title without sub-titles; this year's work The arrangement prompt requires the title format to be "III. Work Arrangements for 20XX", focusing on the key work of this year, with the time corresponding to the time in the theme. The secondary title must contain specified keywords, and the title format and number are clearly specified; the paragraph prompt requires the designated title to be filled with content based on the reference content, maintaining coherence with the previous text, not outputting repeated events and data, and only supplementing the content of the designated title. Imperative sentences and subject-verb grammar should be used for writing, including specified keywords, and filling in according to the title background and paragraph word count requirements. Reference content data and event information should be used reasonably to generate a concise and concise paragraph in the tone of an official government document.

10. A method according to claim 9, characterized in that: In the report generation process, the outline parsing step processes data according to the set outline parsing fields; The information retrieval step uses a fusion multi-way retrieval method based on the parsed outline to retrieve and summarize materials for each title based on the rearrangement model, providing reference materials for paragraph generation. The paragraph generation step uses paragraph generation prompts based on the outline parsing and information retrieval results to generate paragraph content through the large model. The report body step generates paragraph content in the order of outline titles, and adds standard content such as title, foreword, and conclusion to form a complete work report body.

Images