Bidding agent based on HippoRAG knowledge graph

Abstract

The invention discloses a bidding agent based on a HippoRAG knowledge graph, and the agent comprises a bid invitation file analyzer which receives a bid invitation file in a document format, blocks the bid invitation file according to chapters and sections, and completes the structural processing of the document; the knowledge graph construction module is used for extracting entities, relationships and constraints from the bid inviting file and constructing a domain-specific knowledge graph; converting the analyzed structured bid invitation information into a knowledge graph structure exclusive to the bid invitation field, and storing the knowledge graph structure in a graph database; a HippoRAG inference engine is combined with a domain exclusive knowledge graph to realize multi-hop retrieval, constraint inference and context aggregation; and the bidding file generator is used for generating a bidding file meeting the requirement of the bidding file on the basis of the retrieval result of the HippoRAG and the constraint guidance. According to the method, discrete text blocks can be converted into a reasonable semantic network, information aggregation along constraint links is retrieved and generated, and the problem of knowledge islands is thoroughly solved.

Description

Technical Field

[0001] This invention relates to a bidding intelligent agent, and more particularly to a bidding intelligent agent based on the HippoRAG knowledge graph. Background Technology

[0002] In existing technologies, the core objective of intelligent bidding agents for the intelligent assistance needs in the bidding field is generally to use artificial intelligence technology to automatically parse bidding documents and generate bids that meet the requirements of the bidding documents based on the parsing results. Unlike general generative AI (such as models that directly generate text based on large-scale corpora), the generation of bids must strictly follow the constraints of the bidding documents (such as qualification requirements, technical parameters, commercial terms, etc.). Therefore, existing technologies usually adopt a technical path of "first parsing the bidding documents to form a knowledge base, and then combining the knowledge base with a large language model (LLM) to generate content." Among these, the Retrieval-Augmented Generation (RAG) framework is the mainstream solution in the current industry.

[0003] RAG technology, as an AI framework that integrates information retrieval and traditional generative models, is widely used in the generation of tender documents in existing technologies. Its core design concept is to combine the precise positioning capabilities of traditional information retrieval systems (such as keyword-based database retrieval and semantic vector-based vector database retrieval) with the content generation and generalization capabilities of Large Language Models (LLM), thereby improving the matching degree between the output content and the requirements of the tender documents through a collaborative "retrieval-generation" mechanism. Specifically, the application process of RAG technology in existing technologies typically includes the following steps: (1) Standardization of original documents; The original text is formatted uniformly and cut into appropriately sized chunks according to certain rules. This prepares "fine-grained, high-quality" text units for subsequent vector retrieval.

[0004] (2) Vector conversion and storage; The Embedding model is used to convert text blocks into high-dimensional vectors (capturing semantic features) and store them in a vector database.

[0005] (3) Text retrieval; Then, the user's query is encoded using the same embedding model. In the vector library, algorithms such as cosine similarity and Euclidean distance are used to find the text block with the highest matching degree to the query, and the most useful and relevant information fragments are quickly filtered out from the long document.

[0006] (4) Enhanced generation of large models; By using retrieved document fragments as context for prompts in the large model, the model incorporates external knowledge from the vector library when answering user questions, improving the credibility of the answers and outputting more accurate and professional results, thus avoiding illusions.

[0007] The overall process of the RAG framework is as follows: Figure 1 As shown, through a closed loop of "text -> vector -> retrieval -> generation," RAG addresses the pain points of large models, such as "outdated knowledge, insufficient professionalism, and susceptibility to illusions." It transforms large models into knowledge integrators, calling upon external knowledge bases in real time and using the retrieved, precise knowledge to assist in generating answers. This ensures timeliness and professionalism while making the output traceable and more credible, making it highly suitable for application in bidding and tendering intelligent agents.

[0008] In existing technologies, while bidding agents based on traditional RAG improve the flexibility of bid document generation through the "retrieval-generation" paradigm, their underlying logic of "document segmentation → independent retrieval → surface generation" leads to a serious "knowledge silo" problem in the process of parsing bidding documents, specifically manifested in the following three levels of defects: 1. Physically dividing a document into blocks disrupts logical connections, creating isolated fragments of information; The first step in traditional RAG (Research and Evaluation Group) is to divide the tender document (such as a PDF / Word document) into independent text blocks according to chapter titles (such as "Instructions to Bidders 2.1", "Technical Specifications 3.2", "Evaluation Method Appendix 4.1", etc.), and use these text blocks as the basic units of the search database. This block-based approach is essentially a crude disassembly of the tender document's logic, resulting in the clause information being stored as separate vectors with no connection between them in the vector database. This makes it impossible for the search module to perceive the logical dependencies between clauses, and it can only recall isolated text blocks based on keywords, unable to reconstruct the constraint chain of the tender document.

[0009] 2. Reference relationships are not explicitly modeled, causing cross-chapter navigation to fail; The tender documents contain numerous cross-chapter references such as "see Chapter X" and "see Table X" (e.g., "For detailed bidder qualification requirements, please refer to Table 3.2 in the pre-evaluation method"). These references are key logical pointers connecting different chapters. However, traditional RAG technology does not structurally extract and model these reference relationships. For non-explicit knowledge such as "see the pre-evaluation method table," traditional RAG treats it merely as plain text, failing to extract it as a "Clause A -> Reference -> Clause B" relationship edge.

[0010] During the search, even when querying "Bidder Qualification Requirements", the system can only locate the text block "Bidder Qualification: see Table 3.2 in the Bid Evaluation Method", and cannot automatically trigger the search for "Table 3.2 in the Bid Evaluation Method", resulting in the omission of key constraints.

[0011] 3. The large model lacks context, resulting in incomplete and error-prone generated content; The quality of LLM generation is highly dependent on the completeness of the input context. Traditional RAG, due to the problem of knowledge silos, only provides isolated text blocks as the context for LLM, resulting in incomplete information and consequently incomplete generated tender documents, which seriously affects the user experience.

[0012] As can be seen from the above, while traditional RAG technology solves the rigidity problem of template systems through the process of "document segmentation + semantic retrieval + LLM generation," its core drawbacks, such as "physical segmentation destroying logical connections," "failure to model reference relationships," and "lack of context," prevent the parsing of tender documents from reproducing the true logic of "multi-hop references and constraint linkages." Large models, lacking complete contextual information, are prone to omissions, contradictions, or deviations during generation. Summary of the Invention

[0013] The technical problem to be solved by this invention is to provide a bidding agent based on HippoRAG knowledge graph, which can transform discrete text blocks into a reasonable semantic network, enabling retrieval and generation of aggregated information along constrained links, thus completely solving the problem of knowledge silos.

[0014] To address the aforementioned technical problems, this invention provides a bidding agent based on the HippoRAG knowledge graph, comprising: a bidding document parser that receives bidding documents in document format, divides them into blocks by chapter, and completes document structuring; a knowledge graph construction module that extracts entities, relationships, and constraints from the bidding documents to construct a domain-specific knowledge graph; transforms the parsed structured bidding information into a domain-specific knowledge graph structure and stores it in a graph database; a HippoRAG inference engine that, combined with the domain-specific knowledge graph, enables multi-hop retrieval, constraint inference, and context aggregation; and a bid document generator that, based on the HippoRAG retrieval results and constraint guidance, generates bid documents that meet the requirements of the bidding documents.

[0015] Furthermore, the tender document parser breaks down the uploaded tender documents into blocks according to first-level and second-level headings, providing a structured data source for the knowledge graph construction module.

[0016] Furthermore, the knowledge graph construction module uses chapters, clauses, and constraints as nodes, and reference, inclusion, and constraint relationships as edges, explicitly modeling multi-hop reference logic and constraint linkage rules.

[0017] Furthermore, the retrieval end of the HippoRAG inference engine achieves cross-chapter multi-hop inference through knowledge graph "reference edges", and the generation end of the HippoRAG inference engine injects "constraint nodes" to guide the compliant output of the large language model.

[0018] Furthermore, during the retrieval process, the HippoRAG inference engine first selects a clause node in the knowledge graph. This clause node is assigned to a chapter node through an "inclusion relationship" edge, associated with another clause node through a "reference edge," and bound to a constraint node through a "constraint edge," forming a complete link of "chapter → clause → reference clause → constraint rule."

[0019] Furthermore, the HippoRAG inference engine also includes a material library for bid content, which is extracted from the knowledge graph and associated with the search results. The bid document generator combines the recalled text blocks and bid templates, and calls the generation capabilities of the large language model to generate a draft bid document that meets the bidding requirements under the template rules and material constraints.

[0020] Furthermore, the tender document generator is equipped with a tender document template. During query parsing, a large language model is used to extract query instructions from the tender document template, and word vector embedding is performed to convert natural language into computer-understandable numerical vectors, thus adapting the format for knowledge graph retrieval.

[0021] Furthermore, the document format is either PDF or Word.

[0022] Compared with the prior art, the present invention has the following beneficial effects: The bidding agent based on HippoRAG knowledge graph provided by the present invention transforms discrete text blocks into a reasonable semantic network by explicitly modeling reference relationships and constructing a constraint linkage graph, enabling retrieval and generation to aggregate information along constraint links, thus completely solving the problem of knowledge silos. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the RAG process; Figure 2 This is a schematic diagram of the bidding intelligent agent system architecture based on HippoRAG knowledge graph of the present invention; Figure 3 This is a schematic diagram illustrating the implementation steps of the bidding agent based on the HippoRAG knowledge graph of this invention. Detailed Implementation

[0024] The present invention will now be further described with reference to the accompanying drawings and embodiments.

[0025] This invention addresses the issues of knowledge silos and multi-hop reference breaks in bid document parsing using traditional RAG technology, proposing a fusion solution of a "customized HippoRAG framework + domain-specific knowledge graph". The core logic is: 1. Knowledge Graphification: The "chapter structure, constraint relationships, and cross-chapter references" of the tender documents are explicitly modeled into a domain-specific knowledge graph, breaking the isolation of physical document blocks; 2. Introduction of HippoRAG framework: To solve the problem of information fragmentation across chapters caused by knowledge silos, the present invention introduces the latest HippoRAG framework. Its native "retrieval enhancement generation + knowledge graph fusion" feature can connect the scattered content in the bidding documents through the knowledge graph, making up for the deficiency of traditional RAG which cannot establish logical connections by relying solely on text block division, and laying the technical foundation for subsequent customized transformation. 3. HippoRAG Customization: The core of the HippoRAG customization in this invention focuses on the retrieval module. By deeply binding its native retrieval logic with a domain-specific knowledge graph, it is transformed into "semantic link retrieval guided by a domain-specific knowledge graph", making it compatible with the constraints and guidance of the domain-specific knowledge graph, and realizing the upgrade from "text retrieval" to "semantic link retrieval". 4. Multi-hop parsing generation: By connecting cross-chapter information through the "reference edges" and "constraint types" of the knowledge graph, HippoRAG can aggregate information along the link of "qualification requirements → evaluation method → ​​instructions to bidders" and finally generate a bid document that meets all constraints.

[0026] The bidding agent based on the HippoRAG knowledge graph provided by this invention consists of four layers, realizing a closed loop of the entire process of "tender document input → knowledge graph construction → multi-hop parsing → tender document generation". The overall system architecture is as follows: Figure 2 As shown, it includes: The tender document parser, as the input layer, receives tender documents in PDF / Word format, divides them into blocks by chapter, and completes document structuring processing. The knowledge graph construction module, as the knowledge graph layer, extracts entities, relationships, and constraints from the tender documents to construct a domain-specific knowledge graph. The HippoRAG inference engine, as the core engine layer, combines domain-specific knowledge graphs to achieve multi-hop retrieval, constrained reasoning, and context aggregation. The tender document generator, as the output layer, generates tender documents that meet the requirements of the tender documents based on the search results and constraints of HippoRAG.

[0027] This invention creates a domain-specific knowledge graph: using the "chapter, clause, and constraint" of the bidding documents as nodes, and "reference (cross-chapter association), inclusion (chapter-clause hierarchy), constraint (clause-rule binding)" as edges, it explicitly models multi-hop reference logic and constraint linkage rules (such as the link of "qualification requirements → pre-approval table of evaluation method → ​​performance threshold").

[0028] This invention addresses the HippoRAG framework, specifically designed for the field of bidding intelligent agents. Focusing on the core requirements of "multi-hop referencing and constraint linkage" in bid document parsing, it modifies HippoRAG's retrieval and generation modules. The retrieval end utilizes knowledge graph "reference edges" to achieve cross-chapter multi-hop reasoning (e.g., aggregating information along the "qualification requirements → evaluation method → ​​bidder instructions" link), while the generation end injects "constraint nodes" to guide LLM compliant output (e.g., generating content in linkage according to "qualification constraints → performance constraints"). This overcomes the traditional RAG's problems of "knowledge silos" and "inaccurate constraints" in the bidding field, achieving precise parsing and generation capabilities specific to this domain.

[0029] The detailed steps of the bidding intelligent agent of the present invention are as follows: Figure 3 As shown, its design covers the entire process from tender document parsing to bid document generation, realizing the construction of a fully integrated intelligent agent. This process can be divided into two main stages: tender document processing and knowledge graph construction, and bid document generation. The logic and connection of each step are as follows: Phase 1: Tender document processing and knowledge graph construction; ① Upload tender documents: Upload the target tender documents to the system as the initial input for the process.

[0030] ② Tender document parsing: The uploaded tender documents are broken down into blocks according to first-level headings and second-level headings, providing a structured data source for subsequent knowledge graph construction.

[0031] ③ Knowledge Graph Construction (Graph Database): The parsed bidding information is transformed into a knowledge graph structure specific to the bidding domain (containing elements such as "entity-relationship-attribute") and stored in the graph database to realize the structured storage and efficient association query of bidding information.

[0032] Phase Two: Tender Document Generation; ① Tender document template: Extract the basic framework of the tender document (including fixed format, required modules, etc.) from the tender document as the "skeleton" for content generation.

[0033] ② Query parsing: The LLM large language model is used to extract query commands from the tender document template and perform word vector embedding processing (converting natural language into numerical vectors that computers can understand) to adapt the format for knowledge graph retrieval.

[0034] ③ Graph Retrieval: Using the query vector generated above, accurately retrieve and locate the bidding information node that matches the query in the graph database (knowledge graph) constructed in Phase 1.

[0035] ④ Recall related text blocks: Extract text fragments (such as specific descriptions of bidding requirements, technical parameter examples, etc.) related to the search results from the knowledge graph as a "material library" for the bidding content.

[0036] ⑤ LLM Constraint Generation: Combining the recalled text blocks and bidding templates, the generation capabilities of the Large Language Model (LLM) are invoked to generate a draft of the bidding documents that meets the bidding requirements under the constraints of template rules and materials.

[0037] ⑥ Tender document output: Output the final version of the tender document.

[0038] The entire process uses knowledge graphs as its core hub: first, it uses the analysis of bidding documents and the construction of the knowledge graph to store the scattered bidding information in a structured way; then, it uses "word vector embedding + knowledge graph retrieval" to provide accurate related information support for the generation of bid documents; and finally, it uses LLM to complete the content generation and output.

[0039] The core innovation of this invention lies in the deep customization and transformation of the traditional HippoRAG framework. By deeply integrating it with a "domain-specific knowledge graph", it breaks through the inherent limitations of traditional HippoRAG in intelligent parsing of tender documents and endows it with unique technical advantages in complex tendering scenarios.

[0040] The domain-specific knowledge graph constructed in this invention is based on the "chapter structure, clause content, and constraint rules" of the bidding documents. It explicitly models the logical connections between these elements through three types of nodes and three types of edges, forming a semantic network. A typical example is as follows: Taking "Bidder Qualification Requirements" as an example, the graph will have a clause node (such as "3.1 Bidder Qualification Requirements", which contains the original text "must possess general contracting qualification for building construction"), which is assigned to a chapter node (such as "Chapter 2 Instructions to Bidders") through "inclusion relationship" edges. At the same time, since the bidding documents state that "specific qualification requirements are detailed in the appendix to the Instructions to Bidders", this clause node is associated with another clause node through "reference edges" (such as "Appendix to Instructions to Bidders 3.2 Qualification Scoring Standards", which contains the detail "Level 1 qualification requires 100,000㎡ of public building performance in the past 3 years"). In addition, this clause node is also bound to constraint nodes through "constraint edges" (such as "Qualification level ≥ Level 1", "Performance type = public building", "Performance scale ≥ 100,000㎡"), transforming abstract requirements into structured rules. In this way, when parsing "Bidder Qualifications", the graph can automatically jump to the specific scoring criteria in the preceding table through "reference edges" and extract rigid thresholds through "constraint edges", forming a complete link of "chapter → clause → reference clause → constraint rule", which completely solves the knowledge silo problem of "finding the index but not finding the content" in traditional RAG.

[0041] In summary, this invention, by customizing a domain-specific knowledge graph (with chapters, clauses, and constraints as nodes, and reference, inclusion, and constraint relationships as edges), endows HippoRAG with stronger multi-hop reasoning and constraint guidance capabilities, breaking through the knowledge silo dilemma of traditional RAGs. It achieves complete aggregation of cross-chapter information and accurate generation of tender document constraints, significantly improving the parsing accuracy of tender documents and the efficiency of tender document generation, while reducing labor costs.

[0042] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications and improvements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be defined by the claims.

Claims

1. A bidding agent based on the HippoRAG knowledge graph, characterized in that, include: The tender document parser receives tender documents in document format, divides them into blocks by chapter, and completes document structuring processing. The knowledge graph construction module extracts entities, relationships, and constraints from the tender documents to build a domain-specific knowledge graph. The parsed and structured bidding information is transformed into a knowledge graph structure specific to the bidding field and stored in a graph database; The HippoRAG inference engine, combined with domain-specific knowledge graphs, enables multi-hop retrieval, constrained reasoning, and contextual aggregation. The tender document generator, guided by HippoRAG search results and constraints, generates tender documents that meet the requirements of the tender documents.

2. The bidding agent based on the HippoRAG knowledge graph as described in claim 1, characterized in that, The tender document parser breaks down the uploaded tender documents into blocks according to first-level and second-level headings, providing a structured data source for the knowledge graph construction module.

3. The bidding agent based on the HippoRAG knowledge graph as described in claim 1, characterized in that, The knowledge graph construction module uses chapters, clauses, and constraints as nodes, and reference, inclusion, and constraint relationships as edges, explicitly modeling multi-hop reference logic and constraint linkage rules.

4. The bidding agent based on the HippoRAG knowledge graph as described in claim 3, characterized in that, The retrieval end of the HippoRAG inference engine achieves cross-chapter multi-hop inference through knowledge graph "reference edges," while the generation end of the HippoRAG inference engine injects "constraint nodes" to guide the compliant output of the large language model.

5. The bidding agent based on the HippoRAG knowledge graph as described in claim 4, characterized in that, When the HippoRAG inference engine performs a retrieval, it first selects a clause node in the knowledge graph. This clause node is assigned to a chapter node through an "inclusion relationship" edge, associated with another clause node through a "reference edge", and also bound to a constraint node through a "constraint edge", forming a complete link of "chapter → clause → reference clause → constraint rule".

6. The bidding agent based on the HippoRAG knowledge graph as described in claim 1, characterized in that, The HippoRAG inference engine also includes a library of text fragments related to the search results extracted from the knowledge graph as material for the bidding content. The bidding document generator combines the recalled text blocks and bidding templates, and calls the generation capabilities of the large language model to generate a draft bidding document that meets the bidding requirements under the template rules and material constraints.

7. The bidding agent based on the HippoRAG knowledge graph as described in claim 1, characterized in that, The tender document generator is equipped with a tender document template. During query parsing, a large language model is used to extract query instructions from the tender document template, and word vector embedding is performed to convert natural language into numerical vectors that computers can understand, thus adapting the format for knowledge graph retrieval.

8. The bidding agent based on the HippoRAG knowledge graph as described in any one of claims 1-7, characterized in that, The document format is either PDF or Word.

Images