CN-121997923-A - Bid agent based on HippoRAG knowledge graph

CN121997923ACN 121997923 ACN121997923 ACN 121997923ACN-121997923-A

Abstract

The invention discloses a bidding intelligent agent based on HippoRAG knowledge maps, which comprises a bidding document analyzer, a knowledge map construction module, a HippoRAG reasoning engine, a bidding document generator and a bidding document generator, wherein the bidding document analyzer receives bidding documents in a document format, blocks the bidding documents according to chapters to complete document structuring, the knowledge map construction module extracts entities, relations and constraints from the bidding documents to construct a domain-specific knowledge map, converts the analyzed structured bidding information into a domain-specific knowledge map structure and stores the domain-specific knowledge map structure in a graph database, the HippoRAG reasoning engine is combined with the domain-specific knowledge map to realize multi-hop searching, constraint reasoning and context aggregation, and the bidding document generator is used for generating bidding documents meeting the requirements of the bidding documents based on the searching result and constraint guidance of HippoRAG. The invention can convert discrete text blocks into a reasoning semantic network, so that information is aggregated along constraint links in search and generation, and the problem of knowledge island is thoroughly solved.

Inventors

SUN YULIANG
ZHANG DIE
SUN QIANG

Assignees

上海埃威信息科技有限公司

Dates

Publication Date: 20260508
Application Date: 20251229

Claims (8)

1. A bid agent based on HippoRAG knowledge graph, comprising: the bidding document analyzer receives the bidding document in the document format, and blocks the bidding document according to chapters to complete the document structuring process; The knowledge graph construction module extracts entities, relations and constraints from the bidding documents to construct a domain-specific knowledge graph, converts the parsed structured bidding information into a domain-specific knowledge graph structure, and stores the domain-specific knowledge graph structure in the graph database; HippoRAG an inference engine, which combines the domain-specific knowledge graph to realize multi-hop retrieval, constraint reasoning and context aggregation; And a bid document generator for generating a bid document meeting the requirement of the bid document based on the search result of HippoRAG and constraint guidance.
2. The bid agent based on HippoRAG knowledge graph of claim 1 wherein the bid document parser breaks down the contents of the uploaded bid document, blocks the bid document according to a primary header and a secondary header, and provides a structured data source for the knowledge graph construction module.
3. The bid agent based on the HippoRAG knowledge graph of claim 1, wherein the knowledge graph construction module explicitly models multi-hop referencing logic and constraint linkage rules with chapters, terms, constraints as nodes, referencing, including, constraint relationships as edges.
4. The bidding intelligent agent based on HippoRAG's knowledge graph as claimed in claim 3, wherein the search end of HippoRAG inference engine implements cross-chapter multi-hop inference through knowledge graph "reference edge", and the generation end of HippoRAG inference engine injects "constraint node" to guide large language model compliance output.
5. The bid agent of claim 4 wherein said HippoRAG inference engine, upon retrieval, first selects a term node in the knowledge-graph that is assigned to a chapter node by an "inclusive" edge, is associated to another term node by a "reference edge" and also binds a constraint node by a "constraint edge" to form a complete link of "chapter→term→reference term→constraint rule".
6. The bid agent based on HippoRAG knowledge graph of claim 1 wherein said HippoRAG inference engine further comprises means for extracting text segments associated with the search results from the knowledge graph as a corpus of bid content, said bid document generator invoking the generation capabilities of the large language model in combination with recalled text blocks, bid templates, and under template rules and material constraints, generating a bid document manuscript meeting bidding requirements.
7. The bid agent based on HippoRAG knowledge graph of claim 1 wherein the bid file generator is provided with a bid file template, and when query analysis is performed, a large language model is used to extract query instructions from the bid file template, word vector embedding processing is performed to convert natural language into computer-understandable numerical vector, and format adaptation is performed for knowledge graph retrieval.
8. The HippoRAG knowledge-graph-based bid agent of any one of claims 1-7, wherein the document format is PDF format or Word format.

Description

Bid agent based on HippoRAG knowledge graph Technical Field The invention relates to a bidding intelligent agent, in particular to a bidding intelligent agent based on HippoRAG knowledge graph. Background In the prior art, aiming at the intelligent auxiliary requirement of the bidding field, the core goal of the bidding intelligent body is generally to realize automatic analysis of bidding documents by utilizing an artificial intelligence technology and generate bidding documents meeting the requirements of the bidding documents based on analysis results. Unlike the general generation AI (e.g., a model for generating text directly based on a large-scale corpus), the generation of the bidding document needs to strictly follow the constraint conditions (e.g., qualification requirements, technical parameters, business terms, etc.) of the bidding document, so that the prior art generally adopts a technical path of "parsing the bidding document to form a knowledge base first and then combining the knowledge base and a Large Language Model (LLM) to generate content", wherein the retrieval enhancement generation (RETRIEVAL-Augmented Generation, RAG) framework is a mainstream solution in the current industry. The RAG technology is used as an AI framework for fusing information retrieval and a traditional generation model, and is widely applied to the generation link of bidding documents in the prior art. The key design concept is that the accurate positioning capability of a traditional information retrieval system (such as keyword matching-based database retrieval and semantic vector-based vector database retrieval) is combined with the content generation generalization capability of a Large Language Model (LLM), and the matching degree of output content and the requirement of a bid-on file is improved through a 'retrieval-generation' collaborative mechanism. Specifically, the application flow of the RAG technology in the prior art generally includes the following steps: (1) Carrying out standardization processing on an original document; the original text is formatted uniformly and cut into length-adapted segments according to a rule (chunks). The "fine-grained, high-quality" text units are prepared for subsequent vector retrieval. (2) Vector conversion and storage; Text blocks are converted into high-dimensional vectors (capturing semantic features) using Embedding models and stored in a vector database. (3) Text retrieval; And then, encoding the query of the user by using the same Embedding model, searching a text block with the highest matching degree with the query in a vector library through algorithms such as cosine similarity, euclidean distance and the like, and rapidly screening the most useful and most relevant information fragments from the long document. (4) Generating a large model enhancement; The retrieved document fragments are used as the context of the large model prompt words, so that the large model fuses the external knowledge in the vector library when answering the questions of the user, the answering reliability is improved, more accurate and professional results are output, and the illusion is avoided. The whole flow of the RAG framework is shown in figure 1, closed loop is generated through text- > vector- > retrieval- > and the RAG solves the pain points of large model, such as outdated knowledge, insufficient expertise and easy illusion. The large model becomes a knowledge integrator, an external knowledge base is called in real time, and the retrieved accurate knowledge is used for assisting in generating answers, so that timeliness and specialty are guaranteed, output is traceable and more credible, and the method is very suitable for being applied to bidding intelligent agents. In the prior art, although the bidding agent based on the traditional RAG improves the flexibility of bidding document generation through a 'search-generation' paradigm, the underlying logic of 'document blocking, independent search and surface generation' causes serious 'knowledge island' problems in the bidding document analysis process, and the problems are particularly expressed as defects of the following 3 layers: 1. The physical partitioning of the document breaks the logical association to form isolated information fragments; The first step of the conventional RAG is to divide a bidding document (such as PDF/Word document) into independent text blocks (such as "bidder's fibrous knowledge 2.1", "technical specification 3.2", "bid evaluation method prefaced table 4.1") according to chapter titles, and the text blocks are used as basic units of a search repository. The partitioning mode is essentially rough disassembly of the bidding document logic, so that the clause information of the bidding document is respectively stored as independent vectors, the independent vectors are not related to each other in a vector library, the retrieval module cannot perceive the logic dependence among clauses, and only