Intelligent bidding and tendering customer service for power grid project based on large language model and implementation method of intelligent bidding and tendering customer service

Abstract

The invention belongs to the technical field of integration of artificial intelligence and electric power system informatization, and particularly relates to a power grid project bidding and tendering intelligent customer service based on a large language model and an implementation method thereof. The method is used for solving the problems that existing electric power bidding consultation response is slow, information system semantic understanding is poor, and document analysis is inaccurate. The method comprises the following steps: analyzing a file into a Markdown structured text, carrying out semantic partitioning on the text, generating a high-dimensional semantic vector, and constructing a double-layer semantic index of a parent layer and a sublayer; and performing natural language understanding, intention classification and rewriting and intelligent routing to corresponding index hierarchy retrieval on user questions, and generating answers in combination with a large language model. According to the method, the analysis accuracy, retrieval efficiency and precision of the complex document are improved, the consistency and authority of the consultation result are guaranteed, the timely response of dynamic information is realized, different business intentions can be accurately matched, and the user experience is optimized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence and power system informatization fusion, and particularly relates to a power grid project bidding intelligent customer service based on a large language model and an implementation method thereof. BACKGROUND

[0002] The current power industry bidding consultation service mainly relies on the artificial customer service response mode, and in the actual application process, there is an obvious response time delay problem. Especially in the emergency scene involving tender preparation and change query, it cannot meet the business demand of high timeliness decision. In addition, due to the differences in professional background and knowledge structure of customer service personnel, there are understanding ambiguities for specific terms (such as "joint bidding qualification", "unbalanced bid", etc.), which leads to the lack of consistency and authority of the consultation results, and further may cause bidding strategy deviation or dispute.

[0003] The existing bidding information system is mainly based on keyword matching and rule retrieval algorithm, and its semantic understanding ability is limited. Such system cannot deeply analyze the semantic level of natural language questions, especially when dealing with problems containing context dependence or logical judgment (such as "whether the electronic version needs to be sealed?"), the retrieval accuracy and recall ability are not ideal, which affects the information acquisition efficiency of users. In the aspect of document processing, the traditional analysis technology is difficult to deal with the complex format of multi-source documents, especially the tables, title levels, nested structures and other contents contained in PDF and Word format files. There are often problems such as misplacement, information loss and structure loss, which leads to inaccurate extraction of key fields such as bidding deposit, technical parameters, and increases the cost of subsequent manual intervention and verification.

[0004] In the bidding process, the contents such as supplement notice, query file and technical change belong to high-frequency dynamic update information. The existing system mainly relies on manual maintenance and configuration rules for updating, and lacks efficient automatic adaptation mechanism, which leads to lag of information synchronization and cannot realize real-time response to business scenarios. Such updating lag may cause suppliers to prepare documents according to old clauses, resulting in technical deviation or bidding failure and other consequences.

[0005] Due to the incompleteness, inaccuracy or untimeliness of the system in the information processing process, it is easy to cause the suppliers to misjudge the bidding conditions or ignore the key requirements, and further lead to events such as bid cancellation, violation or dispute. At the same time, the existing electronic bidding platform has limited ability to identify irregular behaviors such as bid encirclement and bid stringing, and lacks intelligent monitoring and early warning mechanism, which is difficult to form an effective compliance supervision closed loop, thereby constituting a substantial threat to the transparency and fairness of the industry.

[0006] Although large language models (LLMs) have made significant progress in general natural language processing tasks in recent years, their application in high-professional fields such as power bidding still has obvious limitations. On the one hand, general models may produce fictitious or incorrect answers when dealing with professional policies or terms, lack of generation boundaries based on real business rules, and are prone to mislead information. On the other hand, existing models generally do not conduct directional training on the professional term library of the power industry, resulting in confusion between concepts such as "pre-bid meeting", "bid evaluation meeting", and "bid bond", affecting their usability and credibility in actual scenarios.

[0007] The existing RAG system compensates for the knowledge blind spot of general large language models to some extent through the "retrieval + generation" mechanism, but its architecture in professional scenarios such as power bidding still has significant technical bottlenecks. First, the single-layer vector indexing mechanism lacks semantic level abstraction, resulting in insufficient retrieval accuracy in large-scale text libraries, easy introduction of redundant information and interference content, and difficulty in effectively limiting professional problems; second, the use of static fixed-length text blocking strategy easily breaks the semantic association within and across documents, especially when linking and analyzing the content of bidding documents and supplementary notices, which seriously affects the context integration and recall ability; in addition, the system generally lacks a problem-oriented dynamic routing mechanism and still relies on full vector space retrieval, which cannot intelligently distribute according to problem types, resulting in increased response delay and lack of precise matching ability for user business intent. SUMMARY

[0008] The present application provides an intelligent customer service for power grid project bidding based on a large language model and its implementation method, to solve the technical problems in the prior art of high response delay of artificial customer service, insufficient consistency and authority of consultation results, weak semantic understanding and retrieval ability of information systems, inaccurate analysis of complex format documents, lagging high-frequency dynamic information updates, lack of intelligent monitoring and early warning of irregular behaviors, general large language models prone to generate incorrect answers and term confusion in professional fields, insufficient retrieval accuracy of existing RAG systems, text blocking breaking semantic association, and lack of dynamic routing mechanism leading to response delay and inaccurate intent matching.

[0009] To achieve the above purpose, the technical scheme adopted by the present application is as follows: An implementation method of an intelligent customer service for power grid project bidding based on a large language model, comprising the following steps: Obtaining relevant files and parsing the obtained files into structured text in Markdown format; Performing semantic blocking on the structured text to generate text blocks, and using an embedding model to vectorize the text blocks to generate corresponding high-dimensional semantic vectors; A semantic index structure including a child layer and a parent layer is constructed, the child layer semantic index structure stores a high-dimensional semantic vector corresponding to the text block and its structured text; the data stored in the child layer semantic index structure is clustered, different topic clusters are synthesized, and a corresponding summary semantic vector of the topic cluster is generated and stored in the parent layer semantic index structure; The natural language understanding, intent classification and question rewriting are performed on the user question, and a standard query question is output; based on the child layer semantic index structure and the parent layer semantic index structure, the standard query question is preliminarily indexed, and the routing designation of the parent layer and the child layer and the preliminary indexing range are output; The standard query question is generated into a question semantic vector through an embedding model, and the question semantic vector is subjected to K-nearest neighbor similarity retrieval in the preliminary indexing range, and a final retrieval result is output.

[0010] The obtained file is parsed into a structured text in Markdown format, specifically: the obtained related file is parsed and standardized in format, different format files call corresponding parsers, and the parsed text is automatically converted into a structured text in Markdown format, which can preserve title hierarchy, numbering logic, indentation structure and table semantics.

[0011] After completing the document parsing and Markdown format conversion, the text blocking and semantic vectorization stage is entered, the structured Markdown text is further divided into text blocks that are relatively independent in semantics and self-consistent in logic, each text block corresponds to an information unit with a clear semantic boundary in the bidding document; the blocking process is performed according to the preset blocking rule, and the syntactic structure features, paragraph natural boundaries and content granularity uniformity are comprehensively considered.

[0012] After the blocking of the structured text is completed, a pre-trained Embedding Model embedding model is called to perform deep semantic coding on the content in the text block, convert each text block into a fixed-dimensional semantic vector, form a numerical representation for efficient retrieval and similarity calculation, and output the high-dimensional semantic vector corresponding to the text block.

[0013] Taking Elasticsearch as the core vector retrieval framework, a "parent-child" two-level vector index structure is constructed, including a child layer semantic index structure and a parent layer semantic index structure, the child layer semantic index structure is responsible for storing the original high-dimensional semantic vectors generated by each text block of the embedding model and the associated structured text in Markdown format, and the child layer semantic index directly supports high-precision K nearest neighbor similarity retrieval; the parent layer semantic index structure is constructed on the child layer semantic index structure, a clustering algorithm is used to automatically classify the massive high-dimensional semantic vectors in the child layer semantic index structure, and text blocks with similar semantics are aggregated into different topic clusters, and a summary semantic vector is automatically calculated or generated for each generated topic cluster and stored in the parent layer semantic index structure.

[0014] The parent layer semantic index structure and the child layer semantic index structure are mapped and associated, a parent layer topic cluster corresponds to a group of semantic block texts in the child layer index, a mapping relationship table from the parent layer topic to the child layer text block set is established, and optimal path retrieval can be realized, when a standard query question is identified to point to a macro topic, the parent layer semantic index structure is preferentially routed to, and when a standard query question is identified to point to a specific and explicit factual question, the child layer semantic index structure is directly routed to.

[0015] An LLM Agent is used for deep semantic understanding and intent analysis of user questions, accurately identifies the core intent and key entity, and after intent analysis, the Agent will optimize and rewrite the original question, convert it into a standard query question with clear structure and explicit semantics, and make the question expression align with the child layer semantic index structure and the parent layer semantic index structure.

[0016] The standard query question is converted into a high-dimensional question semantic vector in real time through an embedding model, the question semantic vector enters the intelligent routing and index mapping stage, the LLM Agent accurately routes it to different index structures according to the analyzed question intent and semantic range, for a broad thematic inquiry, the system directs it to the parent layer semantic index structure to obtain a complete topic text segment set; for a specific factual question, it is directly routed to the child layer semantic index structure for accurate matching, and the routing of the parent layer and the child layer is specified, and the preliminary index range is obtained, the question semantic vector performs K nearest neighbor similarity retrieval in the preliminary index range, calculates the semantic distance in the vector space, accurately locates and queries the text block that best matches the query intent, and forms a preliminary candidate text segment set; a double optimization mechanism is used, a semantic correlation score threshold is set, and low correlation interference information in the candidate text segment set is filtered out, and a text length balancing mechanism is started, automatically filtering long or short text segments, intelligently integrating the paragraphs that meet the requirements, and outputting the final retrieval result.

[0017] The original question of the user is organically integrated with the structured text line in Markdown format corresponding to the final search result, and structured, a prompt is constructed, and the completed prompt is sent to a large language model deployed locally, based on context understanding, deep analysis of the semantic association between the question and the search result text, focusing on the most critical information points through an attention mechanism, and generating a natural language answer that meets the requirements.

[0018] A power grid project bidding intelligent customer service based on a large language model, comprising a text analysis module, a semantic vector module, a double-layer index module, a preliminary index module and a similarity retrieval module. The text analysis module is used to obtain relevant files and parse the obtained files into structured text in Markdown format. The semantic vector module is used to perform semantic segmentation on the structured text to generate text blocks, and use an embedding model to generate corresponding high-dimensional semantic vectors for the divided text blocks. The double-layer index module is used to construct a semantic index structure comprising a child layer and a parent layer, the child layer semantic index structure stores the high-dimensional semantic vectors and the structured text corresponding to the text blocks, the data stored in the child layer semantic index structure is clustered to synthesize different topic clusters, and the corresponding summary semantic vectors of the topic clusters are stored in the parent layer semantic index structure. The preliminary index module is used to perform natural language understanding, intent classification and question rewriting on the user question to output a standard query question, and perform preliminary indexing on the standard query question based on the child layer semantic index structure and the parent layer semantic index structure to output the routing designation and the preliminary indexing range of the parent layer and the child layer. The similarity retrieval module is used to generate a question semantic vector for the standard query question through an embedding model, and perform K-nearest neighbor similarity retrieval on the question semantic vector in the preliminary indexing range to output the final search result.

[0019] Compared with the prior art, the present application has the following beneficial effects: The text analysis module of the present application can convert different format files into structured text in Markdown format by calling corresponding parsers, which can completely retain the title level, numbering logic, indentation structure and table semantics of the document, effectively solving the problems of misplacement, information loss and structure loss when traditional parsing techniques process multi-source complex format documents. The extraction of key business fields is more accurate, reducing the subsequent manual intervention and verification cost, and providing a reliable data basis for subsequent search and answer generation.

[0020] Further, the application first divides the structured text into text blocks that are semantically independent and logically consistent, and then performs deep semantic coding through an embedded model to generate high-dimensional semantic vectors. This processing method breaks through the limitations of traditional keyword matching and rule retrieval, can deeply mine the deep semantics of the text, accurately capture the semantic associations within the document and across documents, significantly improve the understanding accuracy of natural language questions, and especially can better handle complex problems that contain context dependence or logical judgment.

[0021] Further, the application constructs a "parent-child" two-level semantic index structure, the child layer stores the original text block vectors and structured text, supporting high-precision retrieval; the parent layer forms topic clusters through clustering and stores summary vectors, realizing fast positioning of macro topics. Combined with the intelligent routing mechanism based on problem intent, broad thematic inquiries are preferentially matched to the parent layer index, and specific factual questions are directly routed to the child layer index, effectively reducing the search scope and avoiding the redundancy and inefficiency of full vector space retrieval, solving the problems of insufficient retrieval accuracy and response delay of existing RAG systems, while avoiding the semantic fragmentation caused by static partitioning.

[0022] Further, the application performs deep semantic understanding, intent classification and standardized rewriting of user questions through the LLM Agent, ensuring accurate alignment of query expression and index structure; in the answer generation stage, the structured text retrieved and the locally deployed large language model are combined to focus on key information based on context semantic association, avoiding the fabricated answers or term confusion that may occur with general large language models. This process replaces the response mode that relies on human customer service, eliminates the understanding ambiguity caused by differences in customer service professional background, improves the consistency and authority of the consultation results, and reduces the occurrence of bid strategy bias or dispute disputes.

[0023] Further, the document parsing, semantic partitioning, vector generation and index construction processes of the application can be automated, and for high-frequency dynamic update information such as supplementary notices, answer files and technical changes, there is no need for manual maintenance and configuration of rules, and new information can be quickly integrated into the index system, solving the problem of information synchronization lag in existing systems. Suppliers can obtain the latest clause information in a timely manner, avoid technical bias or bid failure due to reliance on old clauses to prepare documents, and at the same time reduce the events of bid rejection, violation or dispute caused by incomplete and inaccurate information.

[0024] Further, the application uses the dynamic routing mechanism of the LLM Agent to intelligently distribute to the corresponding index level according to the problem type and semantic range, achieving accurate adaptation to different types of user needs. Whether it is macro topic consultation or specific fact query, efficient response can be obtained, avoiding the "one-size-fits-all" retrieval mode of traditional systems, improving the matching accuracy of user business intent, and optimizing the user information acquisition experience. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 An intelligent customer service architecture diagram for a large language model of a power grid project bidding example; Figure 2 A hierarchical semantic indexing architecture diagram in the embodiment of the application; Figure 3 An intelligent Agent user question routing index flow diagram in the embodiment of the application; Figure 4 An intelligent customer service system structure diagram based on a large language model of a power grid project bidding. DETAILED DESCRIPTION

[0026] In order to further understand the content of the present application, the following will combine the drawings and specific embodiments to describe the present application in detail. It should be understood that the embodiments are only to explain the present application and not to limit it.

[0027] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0028] The embodiment proposes an intelligent customer service implementation method for a power grid project bidding based on a large language model, including the following steps: Obtain relevant files and parse the obtained files into structured text in Markdown format; Perform semantic segmentation on the structured text to generate text blocks, and use an embedding model to generate corresponding high-dimensional semantic vectors for the segmented text blocks; Construct a semantic index structure containing sublayers and parent layers, the sublayer semantic index structure stores the high-dimensional semantic vectors corresponding to the text blocks and their structured text; cluster the data stored in the sublayer semantic index structure, synthesize different topic clusters, and generate corresponding summary semantic vectors for the topic clusters and store them in the parent layer semantic index structure; Perform natural language understanding, intent classification and question rewriting on the user question, and output a standard query question; based on the sublayer semantic index structure and the parent layer semantic index structure, preliminarily index the standard query question, output the routing designation of the parent layer and the sublayer, and the preliminary indexing range; Generate a question semantic vector for the standard query question through an embedding model, and perform K-nearest neighbor similarity retrieval on the question semantic vector in the preliminary indexing range to output the final retrieval result.

[0029] Based on the above construction method, the embodiment combines Figure 1 and Figure 2 The method flow diagram is described in detail as follows, and the specific implementation method is as follows: S1: First, document access and upload, the administrator logs in the system background, uploads various types of file materials related to power grid project bidding through graphical interface or API interface, including but not limited to: bidding announcement, bidding notice, scoring standard, technical specification, business terms, contract template, supplementary file, answer notice, etc. The access document supports common office document formats such as PDF, Word, etc. At the same time, the metadata of the uploaded content is registered, recording the file name, upload time, and basic information such as the project it belongs to.

[0030] S2: Analyze and standardize the above-mentioned access documents, the system has a built-in document analysis module, which calls the corresponding parser for different formats of files, including PDF Parser, Word Parser, etc. To achieve title recognition, paragraph splitting, table structure extraction, list reconstruction, etc. The parsed text is automatically converted to Markdown format, which can preserve title hierarchy, numbering logic, indentation structure and table semantics, enhancing the context analysis capability of large language models in subsequent processing.

[0031] S3: Text blocking and semantic vector generation for the above-structured Markdown text, specifically, after completing document analysis and Markdown format standardization, enter the text blocking and semantic vectorization stage. This stage aims to further divide the structured Markdown text into semantically independent and logically consistent text blocks, each corresponding to an information unit with a clear semantic boundary in the bidding document, such as a clause item, a time node description, a scoring detail or a technical parameter description. The blocking process is performed according to pre-set blocking rules, taking into account syntactic structure features, paragraph natural boundaries and content granularity uniformity, to ensure the logical integrity of semantic blocks.

[0032] In the bidding document, common blocking examples include: division by clause item, division by scoring item, division by time node, and division by contract clause, as follows: Division by clause item: for example, "3.2 Bid Bond" in the "Bidders' Guide" as an independent semantic block, including bond amount, payment form, and refund conditions; Division by scoring item: for example, each sub-item (such as scheme integrity, technical feasibility, innovation point, etc.) under "Technical Scheme Score (Weight 40%)" in the scoring standard is treated as an independent block; Division by time node: for example, "Bid Deadline", "Opening Time", "Clarification Application Deadline" in the bidding schedule are stored separately; According to the contract terms: for example, the contract template "Article 8, breach of contract" about the calculation method of liquidated damages, exemption cases and other segments are extracted as semantic blocks.

[0033] After the completion of the segmentation of the structured text, the pre-trained Embedding Model embedding model is called, which converts each text block into a fixed-dimensional semantic vector by deeply encoding the content in the text block, forming a numerical representation that can be used for efficient retrieval and similarity calculation, to realize the mapping of each text block divided into a vector representation in a high-dimensional semantic space. To improve the adaptability of the field, a power industry professional term enhancement mechanism is introduced in this embodiment, which strengthens the features of the embedding model for professional terms such as "clarification letter", "pre-bid meeting", "joint bid", and "unbalanced bid", so that the representation of the embedding model in the vector space is more distinguishable and semantically stable, thereby supporting subsequent precise retrieval and question answering applications based on semantics.

[0034] S4: Semantic index construction and organization. On the basis of the completion of the vectorization of the text block, the system enters the semantic index construction and organization phase. This phase aims to systematically organize the massive and scattered semantic vectors to form a multi-level semantic index architecture that has both retrieval accuracy and efficiency and thematic management capabilities. The system uses Elasticsearch as the core vector retrieval framework to build a "parent-child" two-level vector index structure, including a child layer semantic index structure and a parent layer semantic index structure, to realize semantic navigation and precise positioning from coarse to fine. The child layer semantic index structure and the parent layer semantic index structure store the following content: 1. Child layer semantic index structure: The child layer semantic index, as the cornerstone of the entire retrieval system, is responsible for storing the original high-dimensional semantic vectors generated by the embedding model for each text block and the associated Markdown format structured text. This index directly supports high-precision K-nearest neighbor (K-NN) similarity retrieval and can match the semantic intent of user queries from the finest granularity. For example, when the user queries "bid bond payment form", the system can directly locate the text block storing the specific description of "bank guarantee, wire transfer, and letter of credit" in the child layer index, ensuring the accuracy of the answer.

[0035] 2. Parent-Level Semantic Index Structure: To address users' needs for macro-level themes or comprehensive information and improve the efficiency of large-scale retrieval, the system constructs a parent-level semantic index structure on top of the sub-level index. First, the system uses clustering to generate topics. It employs efficient clustering algorithms (such as K-means or graph-based HNSW algorithms) to automatically classify the massive amounts of high-dimensional semantic vectors in the sub-level semantic index structure, aggregating semantically similar text blocks into different topic clusters. For example, all clauses related to "scoring criteria" (such as technical scores, business scores, and solution innovation scores) are aggregated into one topic; all nodes involving "timelines" (such as bid deadlines, bid openings, and clarification periods) form another topic. Second, the system automatically calculates or generates a summary semantic vector for each generated topic cluster, serving as the representative of that topic in the parent-level index. This summary vector can be assigned an easily understood topic label, such as "Scoring Criteria Explanation," "Bid Bond Clause," "Project Technical Requirements," or "Contract Payment Method," thus forming a clear topic-based retrieval entry point.

[0036] 3. Hierarchical Mapping and Collaborative Retrieval: The parent-level semantic index and the child-level semantic index are closely linked through a mapping relationship. A parent-level topic cluster corresponds to a set of semantic block texts in the child-level index. One of the core purposes of this hierarchical structure is to provide an efficient pre-routing channel for the subsequent intelligent question answering module. The system pre-establishes a mapping relationship table from "parent-level topics" to "child-level text block sets," which enables the LLM Agent in step S5 to intelligently select the optimal retrieval path based on its real-time understanding of the user's question. The specific retrieval of the optimal path is as follows: like Figure 3 As shown, when the system recognizes that the user's intent points to a macro topic, it can prioritize routing to the parent index, quickly locate the relevant topic cluster, and then accurately retrieve all sub-text blocks under that cluster through the mapping relationship, thereby avoiding a full database scan.

[0037] When the system recognizes that the user's intent is a specific and clear factual question, it can instruct the retrieval engine to bypass the parent layer and directly perform high-precision similarity matching in the full quantum layer index.

[0038] This flexible routing mechanism provides solid underlying support for S5's "intelligent routing," ensuring that every user question reaches the most relevant information source via the most efficient path.

[0039] S5: User Input Processing and Intelligent Routing. When a user asks a question in natural language through the system interface, such as "I need to see the latest addendum notice" or "Can the bidding deadline be postponed?", the system initiates an intelligent query processing and routing process. This process is led by the built-in LLM Agent module, which uses a finely tuned domain-specific large language model to deeply understand and strategically reconstruct the input question.

[0040] The LLM Agent first performs deep semantic understanding and intent parsing on user queries, accurately identifying their core intent and key entities. This goes beyond shallow keyword matching, aiming to map colloquial and vague expressions to standardized concepts in the bidding and tendering field. For example, when faced with "Can I submit my bid later?", the Agent can not only parse that the user intends to inquire about "process" and "feasibility," but also pinpoint the key entities as "bid deadline" and "extension application." For "What are the qualification requirements for my company?", the Agent can understand the type of bidder referred to by "my company" in the context and confirm that the core intent is to inquire about "bidder qualification terms." After completing intent parsing, the Agent immediately optimizes and rewrites the original question, transforming it into a clearly structured and semantically explicit standard query question. This step aims to eliminate ambiguity and align the question expression with the internal knowledge structure, namely the semantic index structure of the sub-layers and parent layers constructed in step S4, thereby greatly improving the accuracy of subsequent searches. For example, rewriting "Can I submit my bid later?" as "Explanation of the process, conditions and required materials for applying for an extension of the bid deadline" results in a more comprehensive and precise search query that can be more effectively semantically associated with parent topics such as "schedule" or "change of terms".

[0041] The optimized questions then enter the intelligent routing and index mapping stage. As a crucial step connecting user question answering and vector retrieval, the LLM Agent plays a central decision-making role. Based on the parsed question intent and semantic scope, it precisely routes the question to different index levels constructed in S4: for broad topical queries, the system directs them to the parent semantic index structure to obtain a complete set of topical text segments; for specific factual questions, it directly routes them to the child semantic index structure for precise matching. This intelligent routing mechanism defines a precise target range for subsequent semantic retrieval, thus laying a solid foundation for the efficient and accurate retrieval of candidate text segments in the S6 stage.

[0042] S6: Semantic Vector Retrieval and Candidate Text Segment Recall. After the intelligent routing determines the target index domain, the system immediately initiates the semantic retrieval process. The standard query question, after standardization, is first converted into a high-dimensional semantic vector in real time through an embedding model. Then, K-Nearest Neighbor (K-NN) similarity retrieval is performed within the index range specified by the routing. By calculating the semantic distance in the vector space, the system accurately locates the text block that best matches the query intent, forming a preliminary set of candidate text segments. This process fully utilizes the previously established vector index structure to ensure that potential answer segments can be quickly identified from massive amounts of documents.

[0043] To further improve the quality of the candidate set, the system employs a dual optimization mechanism. By setting a semantic relevance score threshold, it effectively filters out low-relevance interfering information; simultaneously, it activates a text length balancing mechanism to automatically filter excessively long or short text fragments and intelligently integrates paragraphs that meet the requirements. After this meticulous screening, the final search results are not only highly relevant to the user's question but also maintain an appropriate information density and completeness, providing high-quality information raw materials for the subsequent answer generation stage.

[0044] S7: Prompt Construction and Answer Generation. After obtaining high-quality candidate texts, the system enters the crucial stage of Prompt construction and answer generation. The system first organically integrates the user's original question with the recalled relevant text blocks, structuring them through a carefully designed prompt template. This template typically includes multiple modules such as clear instruction roles, strict answer requirements, the original question, a list of candidate texts, and generation format specifications. For example: "You are an expert in the bidding and tendering field. Please answer the question strictly based on the following information; if the information is insufficient, please clearly state it. User question: {Original question}. Relevant information: {Candidate text 1} {Candidate text 2}... Please generate accurate and professional answers, prioritizing the use of clause citations and list formats." The completed prompt is sent to a locally deployed large language model for inference generation. Leveraging its powerful contextual understanding capabilities, the model deeply analyzes the semantic relationships between the question and candidate texts, focusing on the most critical information points through an attention mechanism to generate a natural language response that meets the requirements. During this process, the system not only supports single-turn question-and-answer sessions but also maintains the contextual coherence of multi-turn dialogues, ensuring that subsequent questions are accurately answered based on the previous interaction context. To ensure output quality, the system also integrates several optimization measures: after answer generation, it undergoes factual consistency verification, comparing the generated content with the original text to prevent model illusions; it also supports rich formatted output, such as accurately citing source document clause numbers, clearly presenting multiple requirements using numbered lists, and summarizing comparison information in tabular form. Finally, the verified and formatted answer is returned through the user interface, forming a complete, accurate, and easy-to-understand professional response.

[0045] Furthermore, in step S5, the LLM Agent module not only possesses natural language understanding and semantic rewriting capabilities, but also embeds a multi-class intent recognition mechanism driven by a large language model. This mechanism can perform fine-grained business classification based on the semantic features implicit in the user's question. This classification mechanism pre-defines several typical bidding and tendering business intent types, including but not limited to: Time-related questions, such as "When is the bid opening time?" and "Can I postpone submitting my bid?", involve key milestones like the bid deadline, bid opening time, and Q&A session time. Process-related questions, such as "What is the bidding process?" and "What materials are required after winning the bid?", involve stages like bid preparation, submission, clarification, bid evaluation, and contract signing. After identifying the question intent type, the LLM Agent maps the question to a predefined semantic index domain (i.e., the parent index node) in the system, such as "Time Information Index," "Technical Requirements Index," and "Bidding Process Description Index." This "intent → routing target" mapping method significantly narrows the scope of vector retrieval, improving overall retrieval efficiency and semantic matching accuracy.

[0046] Furthermore, in step S4, the text segmentation strategy employs a dynamic semantic segmentation mechanism, unlike the traditional method of truncating text based on a fixed number of tokens. The system can combine natural language processing technologies, such as syntactic analysis, paragraph recognition, and heading layering, to perform logical hierarchy judgments on the document content and dynamically determine the demarcation point of each text block. This ensures that each semantic block fully covers a business logic unit, such as a rating clause, a qualification requirement, or a legal statement, effectively avoiding information fragmentation and context loss caused by mechanical segmentation, thereby improving the quality of semantic embedding and vector retrieval performance.

[0047] In another preferred embodiment, based on the aforementioned method for implementing intelligent customer service for power grid project bidding based on a large language model, this embodiment also proposes an intelligent customer service system for power grid project bidding based on a large language model. This system combines artificial intelligence natural language processing technology with power industry bidding operations, achieving multi-functional integration of document parsing, semantic modeling, intelligent retrieval, and natural language question answering. It provides users with efficient, accurate, and traceable question-and-answer services throughout the entire bidding process. The system consists of an administrator terminal and a user terminal, operates based on a locally deployed large language model to ensure data security, and improves service efficiency and reduces manpower maintenance costs through integrated design. Figure 4 As shown, it mainly includes a text parsing module, a semantic vector module, a two-layer indexing module, a preliminary indexing module, and a similarity retrieval module; The text parsing module is used to obtain relevant files and parse the obtained files into structured text in Markdown format; The semantic vector module is used to semantically segment structured text, generate text blocks, and use an embedding model to generate corresponding high-dimensional semantic vectors for the segmented text blocks. The dual-layer index module is used to construct a semantic index structure containing a sub-layer and a parent layer. The sub-layer semantic index structure stores the high-dimensional semantic vectors and their structured text corresponding to the text blocks. The data stored in the sub-layer semantic index structure is clustered to synthesize different topic clusters, and the corresponding summary semantic vectors of the topic clusters are generated and stored in the parent layer semantic index structure. The preliminary indexing module is used to perform natural language understanding, intent classification, and question rewriting on user questions, and output a standard query question; based on the sub-layer semantic indexing structure and the parent-layer semantic indexing structure, it performs preliminary indexing on the standard query question, and outputs the routing specifications of the parent and sub-layers, as well as the preliminary index range; The similarity retrieval module is used to generate a question semantic vector from the standard query question through an embedding model, perform K-nearest neighbor similarity retrieval on the question semantic vector within the preliminary index range, and output the final retrieval result.

[0048] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can be appropriately combined to form other embodiments that can be understood by those skilled in the art. The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A method for implementing intelligent customer service for power grid project bidding based on a large language model, characterized in that, Includes the following steps: Retrieve the relevant files and parse them into structured text in Markdown format; The structured text is semantically segmented to generate text blocks. An embedding model is used to vectorize the text blocks to generate corresponding high-dimensional semantic vectors. Construct a semantic index structure containing sub-layers and a parent layer. The sub-layer semantic index structure stores the high-dimensional semantic vectors and their structured text corresponding to the text blocks. Cluster the data stored in the sub-layer semantic index structure to synthesize different topic clusters, and generate corresponding summary semantic vectors for the topic clusters and store them in the parent layer semantic index structure. Perform natural language understanding, intent classification, and question rewriting on user questions, and output standard query questions; based on the sub-layer semantic index structure and the parent layer semantic index structure, perform preliminary indexing on standard query questions, and output the routing specifications of the parent and sub-layers, as well as the preliminary index range; The standard query question is generated into a semantic vector by using an embedding model. The semantic vector is then used for K-nearest neighbor similarity retrieval within the initial index range to output the final retrieval results.

2. The method for implementing intelligent customer service for power grid project bidding based on a large language model according to claim 1, characterized in that, The process of parsing the acquired files into Markdown format structured text specifically involves: parsing and standardizing the format of the acquired files; calling the corresponding parser for different file formats; and automatically converting the parsed text into Markdown format structured text, which can retain heading levels, numbering logic, indentation structure, and table semantics.

3. The method for implementing intelligent customer service for power grid project bidding based on a large language model according to claim 2, characterized in that, After completing document parsing and Markdown format conversion, the text segmentation and semantic vectorization stage begins. The structured Markdown text is further divided into semantically independent and logically self-consistent text blocks. Each text block corresponds to an information unit with clear semantic boundaries in the bidding document. The segmentation process is executed according to preset segmentation rules, taking into account the syntactic structure features, paragraph natural boundaries, and the uniformity of content granularity.

4. The method for implementing intelligent customer service for power grid project bidding based on a large language model according to claim 3, characterized in that, After segmenting the structured text, the pre-trained Embedding Model is called to perform deep semantic encoding on the content in the text blocks, converting each text block into a fixed-dimensional semantic vector, forming a numerical representation that can be efficiently retrieved and similarity calculated, and outputting the high-dimensional semantic vector corresponding to the text block.

5. The method for implementing intelligent customer service for power grid project bidding based on a large language model according to claim 1, characterized in that, Using Elasticsearch as the core vector retrieval framework, a two-level "parent-child" vector index structure is constructed, including a child semantic index structure and a parent semantic index structure. The child semantic index structure is responsible for storing the original high-dimensional semantic vectors generated by each text block of the embedding model and their associated Markdown-formatted structured text. The child semantic index directly supports high-precision K-nearest neighbor similarity retrieval. The parent semantic index structure is built on the child semantic index structure. A clustering algorithm is used to automatically classify the massive high-dimensional semantic vectors in the child semantic index structure, aggregating semantically similar text blocks into different topic clusters. For each generated topic cluster, a summary semantic vector is automatically calculated or generated and stored in the parent semantic index structure.

6. The method for implementing intelligent customer service for power grid project bidding based on a large language model according to claim 5, characterized in that, By mapping and associating the parent-level semantic index structure with the child-level semantic index structure, a parent-level topic cluster corresponds to a set of semantic block texts in the child-level index. By establishing a mapping relationship table from the parent-level topic to the child-level text block set, optimal path retrieval can be achieved. When the standard query question is identified as pointing to a macro topic, the route is preferentially routed to the parent-level semantic index structure. When the standard query question is identified as pointing to a specific and clear factual question, the route is directly routed to the child-level semantic index structure.

7. The method for implementing intelligent customer service for power grid project bidding based on a large language model according to claim 1, characterized in that, An LLM Agent is used to perform deep semantic understanding and intent parsing of user questions, accurately identifying their core intent and key entities. After completing intent parsing, the Agent will optimize and rewrite the original question, transforming it into a standard query question with a clear structure and explicit semantics, and aligning the question description with the semantic index structure of the sub-layer and the semantic index structure of the parent layer.

8. The method for implementing intelligent customer service for power grid project bidding based on a large language model according to claim 7, characterized in that, Standard query questions are converted into high-dimensional question semantic vectors in real time through an embedding model. These semantic vectors then enter the intelligent routing and index mapping stage. The LLM Agent, based on the parsed question intent and semantic scope, precisely routes the query to different index structures. For broad topic queries, the system directs them to the parent semantic index structure to obtain a complete set of topical text segments. For specific factual queries, the system directly routes them to the child semantic index structure for precise matching, obtaining the routing specifications for the parent and child layers, as well as the initial index range. Within the initial index range, the question semantic vector performs K-nearest neighbor similarity retrieval, calculating the semantic distance in the vector space to accurately locate the text block that best matches the query intent, forming an initial set of candidate text segments. A dual optimization mechanism is employed: a semantic relevance score threshold is set to filter out low-relevance interference information in the candidate text segment set, and a text length balancing mechanism is activated to automatically filter excessively long or short text segments. The qualified paragraphs are intelligently integrated to output the final search results.

9. The method for implementing intelligent customer service for power grid project bidding based on a large language model according to claim 1, characterized in that, The system organically integrates and structures the user's original question with the relevant Markdown-formatted structured text lines corresponding to the final search results, constructs a Prompt, and sends the completed Prompt to a locally deployed large language model. Based on contextual understanding, it deeply analyzes the semantic relationship between the question and the search result text, focuses on the most critical information points through an attention mechanism, and generates a natural language answer that meets the requirements.

10. An intelligent customer service system for power grid project bidding based on a large language model, comprising the method for implementing an intelligent customer service system for power grid project bidding based on a large language model as described in any one of claims 1 and 9, characterized in that... It includes a text parsing module, a semantic vector module, a two-level indexing module, a preliminary indexing module, and a similarity retrieval module; The text parsing module is used to obtain relevant files and parse the obtained files into structured text in Markdown format; The semantic vector module is used to perform semantic segmentation on structured text, generate text blocks, and use an embedding model to vectorize the text blocks to generate corresponding high-dimensional semantic vectors. The dual-layer index module is used to construct a semantic index structure containing a sub-layer and a parent layer. The sub-layer semantic index structure stores the high-dimensional semantic vectors and their structured text corresponding to the text blocks. The data stored in the sub-layer semantic index structure is clustered to synthesize different topic clusters, and the corresponding summary semantic vectors of the topic clusters are generated and stored in the parent layer semantic index structure. The preliminary indexing module is used to perform natural language understanding, intent classification, and question rewriting on user questions, and output a standard query question; based on the sub-layer semantic indexing structure and the parent-layer semantic indexing structure, it performs preliminary indexing on the standard query question, and outputs the routing specifications of the parent and sub-layers, as well as the preliminary index range; The similarity retrieval module is used to generate a question semantic vector from the standard query question through an embedding model, perform K-nearest neighbor similarity retrieval on the question semantic vector within the preliminary index range, and output the final retrieval result.

Images