Power communication professional document multi-round retrieval enhancement generation method based on domain knowledge base

Abstract

The invention relates to a power communication professional document multi-round retrieval enhancement generation method based on a domain knowledge base, and the method comprises the following steps: loading a text embedding model, a vector retrieval engine and a large language model, and obtaining a generation demand; the text embedding model performs word segmentation and coding on a domain knowledge base text to obtain a first normalized vector; the vector retrieval engine constructs partition vector indexes of the electric power communication knowledge base based on the domain knowledge base text, and stores the first normalized vectors in the corresponding partition vector indexes; the generation requirement is input into the large language model, a preliminary retrieval result is obtained, after information is extracted from the preliminary retrieval result, multiple rounds of retrieval are carried out, and a final target text meeting the generation requirement is obtained. Compared with the prior art, the method has the advantages of realizing fusion accurate retrieval, automatic information extraction, multi-round optimization and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the cross field of natural language processing, artificial intelligence and power communication professional document generation, in particular to a power communication professional document multi-round retrieval enhanced generation method based on a domain knowledge base. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, the text generation technology in the field of natural language processing has penetrated deeply from general scenarios to professional scenarios, especially in the fields of structured text generation such as industry reports, technical documents and planning outlines, the requirements for content accuracy, professionalism and logic are increasingly stringent. However, the existing technology still has many limitations in practical application, as follows: Large language models (LLM) represented by GPT series and Qwen series have been widely used in general text scenarios such as email writing and copywriting due to their powerful semantic understanding and generation capabilities, and can quickly generate fluent and grammatically correct natural language texts. However, its application in professional fields has significant shortcomings: Fact bias problem, LLM's training data covers a wide range but is difficult to update in real time, and the generation logic relies on statistical laws rather than fact checking, which leads to the generation of content that does not conform to reality in professional scenarios. In the generation of power communication planning documents, it may incorrectly reference obsolete industry standards or fabricate non-existent device parameters. Professional knowledge lag problem, technical specifications in the industry field and enterprise internal knowledge base are often updated dynamically, and LLM's static training data cannot incorporate the latest information in time, resulting in a disconnect between the generated content and current professional requirements. Insufficient structured adaptation problem, power communication professional documents usually have fixed chapter frameworks, terminology systems and logical chains, and LLM's free generation mode is difficult to strictly follow such format constraints, and requires a lot of manual adjustments to meet usage requirements.

[0003] To make up for the fact defects of LLM, retrieval augmented generation (RAG) technology has emerged, whose core logic is to provide real-time and accurate factual support for LLM by retrieving external knowledge bases, thereby reducing "hallucinations". However, the existing RAG technology still has the following key defects in the field of professional text generation: The search query (Query) is not targeted enough. The existing method is based on fixed rules or simple semantic matching to generate a search query, which is difficult to accurately locate key information combined with specific generation requirements, resulting in one-sided generated content. The second is the lack of dynamic judgment of the sufficiency of the search results. The search process is mostly "one-time" execution, and the matching degree of the search results and the generation requirements is not checked. Often, the generated content is logically broken due to incomplete information. The third point is that the search and generation processes are separated. In the traditional RAG technology, the search results are only used as "supplementary materials" for generation, and the closed-loop mechanism of "search-generation-verification-research" is not formed. When the generated content has information gaps, secondary search cannot be automatically triggered, and manual intervention is required to supplement the information, which is low in efficiency.

[0004] Text embedding and vector search are the core support of RAG technology, but their application in professional scenarios still faces bottlenecks. First, the precision of the text embedding model is limited. There are a large number of abbreviations, industry terms, and long text specifications in professional fields. The existing embedding model lacks the ability to represent the semantics of such content, resulting in related content missing or unrelated content being detected during vector search. Second, the filtering and fusion efficiency of search results is low. The search results of a professional knowledge base often contain multi-source and heterogeneous information (technical manual fragments, industry standard clauses, historical cases). The existing technology lacks an automated and effective information extraction mechanism, which requires manual screening and integration, increasing the time cost and possibly introducing errors due to human judgment bias.

[0005] In summary, the existing text generation technology has problems such as "insufficient accuracy", "disconnection between search and generation", and "lack of dynamic optimization" in professional scenarios. A closed-loop solution that integrates accurate search, automated information extraction, and multi-round optimization is urgently needed to meet the high requirements of power communication professional document generation for accuracy, efficiency, and professionalism. SUMMARY

[0006] The purpose of the present application is to provide a multi-round search enhanced generation method for professional documents of power communication based on a domain knowledge base, which integrates accurate search, automated information extraction, and multi-round optimization.

[0007] The purpose of the present application can be achieved by the following technical solutions: A multi-round search enhanced generation method for professional documents of power communication based on a domain knowledge base, the method comprising the following steps: loading a text embedding model, a vector search engine, and a large language model, and obtaining a generation requirement; The text embedding model performs word segmentation and encoding on the domain knowledge base text to obtain a first normalized vector; The vector search engine constructs a power communication knowledge base partition vector index based on the domain knowledge base text, and stores the first normalized vector in the corresponding partition vector index; In the generation demand input large language model, a structured query is generated, the structured query is converted into a second normalized vector through a text embedding model, the second normalized vector is matched with the first normalized vector in the partition indexed by the partition vector corresponding to the structured query, a preliminary retrieval result is obtained, information is extracted from the preliminary retrieval result, and multiple rounds of retrieval are performed to obtain a final target text meeting the generation demand.

[0008] Further, the specific steps of the text embedding model for segmenting and encoding the domain knowledge base text to obtain the first normalized vector are: The text embedding model splits the domain knowledge base text according to semantic complete units to obtain a segmentation result, assigns a first semantic weight to the domain term in the segmentation result, the first semantic weight is greater than the semantic weight of the non-domain term in the segmentation result, generates an initial vector of the segmentation result based on the semantic weight, and fuses and normalizes the initial vectors of single segments in the segmentation result to obtain the first normalized vector.

[0009] Further, the calculation process of the normalized vector is: The initial vectors of single segments in the segmentation result are fused by using the last-token pooling strategy, the obtained vector is subjected to L2 normalization to obtain the normalized vector.

[0010] Further, in the process of splitting according to semantic complete units, if the domain knowledge base text is a long text, after obtaining the first normalized vector, the semantic correlation degree of the first normalized vectors of adjacent segments is calculated, the weight of the first normalized vector is adjusted based on the semantic correlation degree to obtain a new first normalized vector.

[0011] Further, the format of the structured query is: <start tag> query content <end tag>.

[0012] Further, the specific steps of extracting information from the preliminary retrieval result are: The preliminary retrieval result is screened for key information by using a large language model to form a JSON format output of key information type-information-source, and the large language model generates a preliminary target text in combination with the JSON format output, historical text and a writing template.

[0013] Further, the specific steps of performing multiple rounds of retrieval to obtain a final target text meeting the generation demand are: It is determined by the large language model whether the target text includes all domain terms of the domain knowledge base text, if yes, the retrieval is terminated, and the target text at this time is taken as the final target text; Otherwise, the large language model takes the missing domain term as a supplementary retrieval query, extracts new information to obtain a new target text, and repeats the above steps until the maximum number of iterations is reached.

[0014] Further, the specific step of matching the first normalized vector in the partition indexed by the second normalized vector and the partition vector corresponding to the structured query is to find the matching first normalized vector using nearest neighbor search, and the first normalized vector is the preliminary retrieval result.

[0015] Further, the text embedding model is a pre-trained embedding model fine-tuned for terms.

[0016] Further, the pre-trained embedding model is Qwen3-Embedding-4B.

[0017] Compared with the prior art, the present application has the following beneficial effects: The present application takes the domain knowledge base as the fact basis, and enhances the representation of professional terms and the partition vector index through the domain-adapted embedding model to improve the retrieval accuracy, reduce missed and false detections, and retain the information source for easy tracing. At the same time, the present application avoids the fact bias and term errors caused by the static training data of LLM from the source, generates highly targeted structured queries, and directly uses the structured extraction results as the core material for generation. The generation process refers to historical texts and templates to realize seamless connection of retrieval-extraction-generation and solve the problem of the split between traditional RAG retrieval and generation. The present application also actively identifies information gaps and automatically triggers secondary retrieval without human intervention, controls the process with the maximum number of iterations, completes the information while ensuring efficiency, and solves the problem of incomplete information in one-time retrieval. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The figure is a system architecture diagram of the text automatic generation method based on retrieval enhancement and multi-round optimization of the present application; Figure 2 The figure is a workflow diagram of the multi-round retrieval optimization module. DETAILED DESCRIPTION

[0019] The present application will be described in detail below in conjunction with the drawings and specific embodiments. The present embodiment is implemented on the premise of the technical solution of the present application, and detailed implementation and specific operation processes are given, but the protection scope of the present application is not limited to the following embodiments.

[0020] Embodiment 1: The present application proposes a power communication professional document multi-round retrieval enhancement generation method based on a domain knowledge base, which includes the following steps: Load the text embedding model, vector retrieval engine and large language model, and obtain the generation requirement; The text embedding model performs word segmentation and coding on the domain knowledge base text to obtain the first normalized vector; The vector retrieval engine constructs a power communication knowledge base partition vector index based on a domain knowledge base text, and stores the first normalized vector in the corresponding partition vector index; In the generation requirement input large language model, a structured query is generated, the structured query is converted into a second normalized vector through a text embedding model, the second normalized vector is matched with the first normalized vector in the partition of the partition vector index corresponding to the structured query, a preliminary retrieval result is obtained, information is extracted from the preliminary retrieval result, and multiple rounds of retrieval are performed to obtain a final target text meeting the generation requirement.

[0021] The application belongs to the cross field of natural language processing, artificial intelligence and power communication professional document generation, and specifically relates to an intelligent generation technology for structured text in the field of power communication by using text embedding, vector retrieval and large language model (LLM), and is especially suitable for scenes such as power communication network planning reports, technical specification documents and operation and maintenance analysis reports that need to strictly combine the knowledge base in the field of power communication. Through the retrieval enhancement and multi-round optimization mechanism, the accuracy and adaptability of the generated content in the power communication professional terms, industry standards and logical framework are improved.

[0022] In view of the problems such as insufficient accuracy of professional content, disconnection between retrieval and generation and lack of dynamic optimization mechanism in the prior art in the generation of power communication professional documents, the application proposes the following goals: The application constructs an efficient retrieval mechanism of the knowledge base in the field of power communication, optimizes the text embedding and vector retrieval strategy, adapts to the power communication term system and knowledge base hierarchical structure, and improves the retrieval accuracy of the field knowledge; realizes automatic information processing in the power communication scene, generates a targeted retrieval query through a large language model, automatically extracts power communication key information in the retrieval result, and reduces manual intervention; designs a multi-round retrieval optimization closed loop: dynamically checks the core elements of the power communication document, automatically triggers supplementary retrieval for information gaps; improves the quality and efficiency of power communication document generation: ensures that the generated content meets the technical standards, term specifications and logical framework in the field of power communication, and focuses on optimizing the accuracy and professionalism of structured text such as communication network planning reports and technical specifications.

[0023] The application realizes automatic generation of power communication professional documents based on retrieval enhancement through the whole process mechanism of “text embedding - vector retrieval - query generation - information extraction - multi-round optimization - text generation”. The core includes four modules: a text embedding and vector retrieval module (adapted to the power communication knowledge base), a retrieval query generation module (customized for the power communication scene), an information extraction and verification module (focusing on power communication key elements), and a multi-round optimization and generation module (forming a closed loop in the field of power communication), which cooperatively complete the accurate processing from “power communication document generation requirement” to “target text”.

[0024] The technical core of the present application is: Text embedding model: Adopt a pre-trained embedding model (such as Qwen3-Embedding-4B) fine-tuned by power communication terms, and perform word segmentation and coding on domain knowledge base texts (such as chapters of “Power Communication Network Planning Specifications” and equipment manuals); For long texts (such as full texts of planning reports), adopt the “segment embedding + context association” strategy to strengthen the semantic representation of professional terms such as “SDH equipment” and “optical cable selection”; Through “last-token pooling”, generate normalized vectors, and enhance retrieval stability through L2 normalization.

[0025] Vector retrieval engine: Based on FAISS, build a power communication knowledge base partition vector index, store according to domain classification dimensions such as “backbone communication network”, “distribution communication network”, and “transmission network”, associate with text metadata, support efficient approximate nearest neighbor search (ANN) by domain partition, and quickly return power communication text fragments matching the query.

[0026] Retrieval query (Query) generation: Through LLM (such as GPT-4.1) combined with power communication document generation requirements (such as chapter title “220kV Substation Communication Link Design” and writing template “Current Situation - Scheme - Parameters”), generate structured queries (format: <|begin_search_query|>query<|end_search_query|>), for example: <|begin_search_query|>220kV Substation Communication Link Redundancy Design Specification<|end_search_query|>, <|begin_search_query|>GYTA53 Type Optical Cable Applicable Scenario in Substation<|end_search_query|>.

[0027] Effective information extraction: Use LLM to filter power communication key information from retrieval results, output in JSON format according to “submodule - information - source”, for example: { "Submodule": "Optical Cable Selection", "Information": "GYTA53 Type Optical Cable is suitable for underground pipe laying in 220kV substations and has anti-lateral pressure performance", "Source": "“Power Communication Optical Cable Selection Manual” Section 3.2" } Text generation: Based on the extracted effective information, historical texts and writing templates, generate target texts that meet the format specifications through LLM, ensuring that the content is consistent with the knowledge base information.

[0028] Multi-round retrieval optimization mechanism: Information sufficiency check: Compare the extracted information with the core elements of the power communication document using LLM to determine if there are information gaps (missing key data, incomplete logic); Dynamic supplementary retrieval: If the information is insufficient, generate a supplementary retrieval query using LLM, call the vector retrieval engine again to obtain new text segments, and extract valid information; Iteration termination condition: Set the maximum number of iterations (max_loop=3), when the information sufficiency meets the demand or reaches the maximum number of iterations, terminate the retrieval and enter the final generation link.

[0029] The overall process of the present application is: Initialization: Load the text embedding model, vector retrieval engine (FAISS index) and LLM, input the generation requirements (such as chapter title, writing template); Initial retrieval: Generate an initial retrieval query using LLM, call the vector retrieval engine to obtain relevant text segments; Information extraction: Extract valid information from the retrieval results and store them in a structured manner; Multi-round optimization: Check the information sufficiency, if there are gaps, generate a supplementary query and repeat the retrieval-extraction process; Text generation: Based on all the extracted valid information, generate the target text that meets the requirements.

[0030] The key points are: Special adaptation in the field of power communication: Optimize the embedding and retrieval strategy for domain-specific terminology, knowledge base hierarchy and standard specifications; Retrieval-generation deep collaboration: Through structured query and information extraction, realize the precise docking of power communication knowledge base and generation process; Multi-round dynamic closed loop: Check the information sufficiency based on the core elements of the power communication document, automatically supplement the retrieval, and solve the problem of incomplete information in single retrieval.

[0031] The system architecture corresponding to the method of the present application is: System architecture Hardware environment: Deploy servers (supporting GPU acceleration) for embedding models and LLMs, and database servers for storing vector indexes and knowledge bases; Software modules: Text embedding module: Realize power communication text preprocessing (terminology standardization, long text segmentation), vector generation; Vector retrieval module: Build power communication partition FAISS index, provide domain-specific retrieval interface; LLM processing module: Generate power communication-specific retrieval queries, extract valid information, and generate target text; Multi-round optimization module: controls the retrieval iteration process and judges the sufficiency of information based on the core elements of power communication; Power Communication Knowledge Base Management Module: Maintains the knowledge base by region, updates the latest State Grid standards synchronously, and supports document upload, version management, and terminology database updates.

[0032] Example of steps (using Section 2.1 of the "2026-2030 Communication Network Planning Report: Current Status of Backbone Communication Network Construction" as an example) Input requirements: The chapter title is "2.1 Current Status of Backbone Communication Network Construction". The writing template requires that it include three sub-modules: "Optical Cable Network, Optical Transmission Network, and Data Communication Network". The historical reference text is "Summary of Backbone Communication Network Construction". It must comply with the requirements of Chapter 3 of "Q / GDW 1973-2013 Technical Guidelines for Planning and Design of Power Communication Networks".

[0033] Initial search: LLM generates 3 structured queries: <|begin_search_query|> Construction Mileage of Backbone Communication Network Optical Cable Network in 2025-2026<|end_search_query|> <|begin_search_query|>Backbone communication network optical transmission network SDH equipment models and port utilization<|end_search_query|> <|begin_search_query|>Number of substations covered by the backbone data communication network<|end_search_query|> The FAISS index "Backbone Communication Network" partition search returned 5 relevant fragments (sources: "Communication Construction Plan", "2025 Operation and Maintenance Report", etc.).

[0034] Information extraction: Extract according to the format "Submodule - Information - Source": [ { Submodule: "Optical Cable Network" "Information": "150 kilometers of new GYTA53 optical fiber cable were added to the backbone communication network, covering 5 220kV substations." Source: Section 4.2 of the "Telecommunications Construction Plan" }, { Submodule: Optical Transmission Network "Information": "The existing SDH equipment model is ZXMP S385, with a port utilization rate of 85%, supporting a 10Gbps transmission rate" "Source": "Section 3.1 of the Backbone Communication Network Operation Report" Multiple rounds of optimization: Check found that the "Data Communication Network" submodule is missing "Bandwidth Status" information, and the LLM generates a supplementary query: <|begin_search_query|> Average bandwidth utilization and peak bandwidth of the backbone data communication network in 2026 <|end_search_query|> Supplementary information after secondary search: "Submodule": "Data Communication Network" "Information": "The average bandwidth utilization of the backbone data communication network is 72%, with a peak bandwidth of 1.2Tbps, covering 12 500kV substations" "Source": "Section 2.3 of the Communication Network Monitoring Report" Text generation: Based on all the information, generate chapter content that conforms to the template structure.

[0035] Performance evaluation indicators Generation accuracy: consistency with the power communication knowledge base (device parameter accuracy, standard clause matching rate); Search efficiency: single-round search response time (≤1 second), total time of multiple rounds of optimization (≤10 seconds); Content integrity: the proportion of generated text covering the core elements of the power communication document (submodule coverage rate ≥95%); Accuracy of domain terminology: correct use of power communication terminology (distinguishing accuracy ≥98%).

[0036] Standard clause matching rate: the matching proportion of generated content with the State Grid standard (clause matching rate ≥90%).

[0037] Benefits: Improve the accuracy of power communication documents: through retrieval, ensure that the content is strictly based on the domain knowledge base, reduce terminology confusion and standard reference errors; Strengthen professional adaptability: optimize retrieval and generation strategies for power communication scenarios, adhere to document frameworks and industry standards; Improve generation efficiency: automate retrieval, extraction, and optimization, reducing the time cost of manual material review (more than 60% more efficient than traditional methods); ​​​​Support dynamic optimization: multi-round retrieval mechanism can flexibly supplement information gap, ensure the content integrity of complex power communication documents.

[0038] Embodiment 2: The application also provides a computer readable storage medium, which stores a program, and the program is executed to realize the method in the embodiment 1. The computer readable medium includes permanent and non-permanent, removable and non-removable media, and can be realized by any method or technology to store information. The information can be computer readable instructions, data structure, program module or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device. According to the definition herein, the computer readable medium does not include transitory computer readable media such as modulated data signals and carriers.

[0039] The above functions, if realized in the form of software function units and sold or used as independent products, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the application or the part of the technical solutions that make essential contributions to the prior art can be embodied in the form of software products. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the embodiments of the application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

Claims

1. A method for generating multi-round search enhancement of power communication professional documents based on domain knowledge base, characterized in that, The method comprises the following steps: The method comprises the following steps: The text embedding model tokenizes and encodes the domain knowledge base text to obtain a first normalized vector. The vector retrieval engine constructs a power communication knowledge base partition vector index based on the domain knowledge base text, and stores the first normalized vector in the corresponding partition vector index. The generated requirement is input into the large language model to generate a structured query, which is converted into a second normalized vector by the text embedding model.

2. The method of claim 1, wherein the method is based on a domain knowledge base for power communication professional document multi-round retrieval enhancement generation. The second normalized vector is matched with the first normalized vector in the partition of the corresponding partition vector index of the structured query to obtain a preliminary retrieval result. The specific steps of the text embedding model tokenizing and encoding the domain knowledge base text to obtain a first normalized vector are as follows: 3.The method of claim 2, wherein, The text embedding model tokenizes and encodes the domain knowledge base text to obtain a first normalized vector. The normalization vector calculation process is as follows:

4. The method of claim 3, wherein the method further comprises: The last-token pooling strategy is used to fuse the initial vectors of single segments in the tokenization result, and the obtained vector is subjected to L2 normalization to obtain the normalized vector.

5. The method of claim 1, wherein the method further comprises: In the process of semantic unit splitting, if the domain knowledge base text is a long text, the semantic correlation degree of adjacent segment first normalized vectors is calculated after obtaining the first normalized vector, and the weight of the first normalized vector is adjusted based on the semantic correlation degree to obtain a new first normalized vector.

6. The method of claim 1, wherein the method further comprises: The format of the structured query is <start tag> query content <end tag>. The specific steps of extracting information from the preliminary retrieval result are as follows:

7. The method of claim 6, wherein the method further comprises: The large language model is used to filter key information from the preliminary retrieval result to form a JSON format output of key information type-information-source. The large language model combines the JSON format output, historical text, and writing templates to generate a preliminary target text. The specific steps of performing multiple rounds of retrieval to obtain the final target text that meets the generation requirement are as follows:

8. The method of claim 1, wherein the method further comprises: The large language model determines whether the target text includes all domain knowledge base text domain terms.

9. The method of claim 1, wherein the method further comprises: If yes, the retrieval is terminated, and the target text at this time is taken as the final target text.

10. The method of claim 1, wherein the method is based on a domain knowledge base for power communication professional document multi-round retrieval enhancement generation. Otherwise, the large language model takes the missing domain terms as a supplementary retrieval query, extracts new information, and obtains a new target text. The specific steps of matching the second normalized vector with the first normalized vector in the partition of the corresponding partition vector index of the structured query are as follows: the nearest neighbor search is used to find the matching first normalized vector, and the first normalized vector is taken as the preliminary retrieval result. The text embedding model is a term fine-tuned pre-trained embedding model. The pre-trained embedding model is Qwen3-Embedding-4B.

Images