CN-122022717-A - Purchasing questioning full-flow intelligent processing system based on large language model

Abstract

The invention provides a purchasing and questioning full-flow intelligent processing system based on a large language model, which concretely comprises a knowledge base management module, a questioning function analysis module, an intelligent reply generation module, an intelligent consultation interaction module and a collaborative management module, wherein the knowledge base management module is responsible for converting rules, cases and purchasing files into structured data and establishing a retrievable vector database, the questioning function analysis module is responsible for receiving the files, extracting elements and checking compliance through OCR and LLM multi-mode recognition, the intelligent reply generation module is responsible for matching rules and case bases from the knowledge base and external data sources according to analyzed sub-problems and automatically generating reply content based on templates so as to support manual adjustment and verification. The invention solves the problems of low efficiency, poor accuracy, weak traceability, disordered knowledge base management, insufficient interaction experience and the like in the existing purchasing and questioning processing flow.

Inventors

• TIAN SHENGLI
• Liao Yaqiao
• MA JINGYUAN
• LUO MINGKANG
• Li Zhuomiao
• Zeng Yingyan
• ZOU XIAOHONG
• WU WENHUI
• WANG SHINA
• WU YUGE
• YE LUGAO
• ZHU HUAN
• ZHAO JINGYING
• WEN YIFENG
• WANG HONG
• OuYang Jinjie

Assignees

• 广州市政府采购中心
• 广东南方信息安全产业基地有限公司

Dates

Publication Date

20260512

Application Date

20251229

Claims (10)

1. 1. A purchase challenge whole-flow intelligent processing system based on a large language model is characterized by comprising the following steps: The knowledge base management module is responsible for converting rules, cases and purchase files into structured data and establishing a retrievable vector database; the questioning function analysis module is responsible for extracting elements and checking compliance through OCR (optical character recognition) and LLM (web language) multi-mode recognition after receiving the file; The intelligent reply generation module is responsible for matching rule and case bases from a knowledge base and an external data source according to the analyzed sub-problems, automatically generating reply content based on a template, and supporting manual adjustment and auditing; The intelligent consultation interaction module supports full-flow natural language dialogue, a manager can initiate consultation at any time through the intelligent consultation interaction module in the processing process, and provides accurate guidance through 'one-question one-answer library+LLM semantic understanding', so that complex problems can be automatically related to historical cases or regulation clauses; And the collaborative management module is used for setting different operation authorities according to roles, automatically recording a full-flow operation log, wherein the full-flow operation log comprises information of 'the uploading time of a suspicious function, the element extraction confirmation person, the reply content modification record and the audit node opinion', and all operations cannot be tampered.
2. 2. The purchasing challenge full-flow intelligent processing system based on the large language model of claim 1, wherein the challenge function analyzing module converts the challenge function from unstructured to structured based on the multi-modal identification capability of LLM, and the specific process is as follows: S1, file receiving and analysis, namely extracting text information in a scanned piece/picture by adopting an OCR technology, converting the file into editable plain text by combining a LLM document segmentation algorithm, and establishing an image-text comparison relation of 'original text-analysis text'; S2, element one-key extraction, namely automatically extracting core elements through a named entity recognition and intention recognition algorithm of LLM, generating a 'questioning function analysis result list', and entering the next link after confirmation by a manager; S3, checking the compliance and the integrity, namely firstly comparing the time limit and the form requirements about the questioning and submitting in the purchasing rule, judging whether the questioning function accords with the acceptance condition, and checking the compliance, then checking whether the extracted elements are complete through a preset rule, reminding sponsors of supplementing the missing items with red marks, and finally, precisely matching the extracted rule terms with a purchasing rule base and a bidding rule base in a knowledge base, and judging whether the terms are valid and quoted accurately; s4, problem disassembly and indexing, namely disassembling the questioned total problem into a plurality of sub-problems through a text splitting algorithm of LLM, and adding a label for each sub-problem to realize accurate matching with a knowledge base case and a rule.
3. 3. The large language model-based purchase challenge whole-flow intelligent processing system according to claim 2, wherein the core elements extracted in step S2 specifically include: basic information, namely questioning the name of a unit, unifying social credit codes, and whether to participate in the purchase of the project, contact persons and contact ways; the questioning attribute is questioning property, submitting time and submitting form; core content, question the total questions, specific fact basis and cited legal and regulation terms.
4. 4. The large language model based purchase challenge whole flow intelligent processing system according to claim 1, wherein the knowledge base management module works as follows: (1) The data structuring process comprises segmenting and indexing the PDF/WORD file by adopting a LLM semantic understanding algorithm, converting the parsed text into a vector, storing the vector into a vector database to realize 'semantic level retrieval'; (2) And the intelligent retrieval comprises the steps of automatically sending a retrieval request to a knowledge base after the intelligent reply generation module receives the disassembled sub-problems, calculating the similarity between the sub-problem vectors and the knowledge base data vectors, sequencing the sub-problem vectors from high to low according to the similarity, and outputting a related result.
5. 5. The intelligent purchasing and questioning full-flow processing system based on a large language model according to claim 1, wherein the intelligent reply generation module is based on the structured output of a 'questioning function analysis module', and automatically generates a reply function draft with standard format, rich arguments and traceable contents by intelligently combining an internal and external knowledge base with a rule base through a retrieval enhancement generation technology.
6. 6. The large language model based purchase challenge whole flow intelligent processing system according to claim 1, wherein the intelligent reply generation module works as follows: (1) Receiving final output data from a 'suspicious function analysis module', packaging the input data into a unified processing task, and triggering an intelligent reply generation pipeline; (2) Converting each sub-problem into a high-dimensional vector, carrying out similarity retrieval in a knowledge base, and generating an 'evidence package' for each sub-problem, wherein the evidence package comprises the retrieved related text fragments and the related scores; (3) The intelligent content generation and tracing label is that natural language replies are synthesized by using a large language model, and tracing information is automatically embedded; (4) Template automatic typesetting and integration, namely integrating the generated sub-question replies into a complete reply conforming to an official format; (5) And (3) manually checking and editing, namely, finally checking the draft by a manager through a man-machine interaction interface.
7. 7. The large language model based purchase challenge whole flow intelligent processing system according to claim 6, wherein the final output data from the challenge function parsing module comprises: structuring questioning data, namely a questioning function analysis result list; A disassembled sub-problem set, namely a plurality of sub-problems with labels; And (5) an original text association index, namely corresponding position information of the sub-questions and the questioning letter original text.
8. 8. The intelligent processing system for purchasing challenge whole flow based on big language model according to claim 6, wherein after similarity searching is performed in knowledge base, the data after searching is prioritized, the rule of priority ordering is that for the sub-questions with labels of "rule quotation" and "compliance", the corresponding and related legal original text and official explanation are preferentially searched from "rule base", and for the sub-questions with labels of "parameter rationality" and "fact identification", the case similar, well-defined case or reply paragraph is preferentially searched from "knowledge base-history case".
9. 9. The intelligent purchasing and questioning full-flow processing system based on large language model according to claim 6, wherein in the intelligent content generation and tracing labeling process, firstly, constructing structured prompt words, clearly instructing LLM, then, replying to each sub-question based on the provided evidence package, and finally, inputting the sub-questions and the corresponding evidence package into LLM through a retrieval enhancement generation framework, and calling LLM to generate content.
10. 10. The intelligent processing system for purchasing and testing complete flow based on large language model as set forth in claim 6, wherein in the process of templatizing automatic typesetting and integrating, the sub-question replies containing the traceability marks are organized and filled according to the preset logic structure of the templates by calling the built-in reply function templates, and simultaneously, the basic information in the "test function analysis result list" is automatically filled to the corresponding positions of the templates, and then a structurally complete reply draft containing the official letter head-up, the deposit positions, the main reply contents and all contents which are pre-typeset according to the official document format is generated.

Description

Purchasing questioning full-flow intelligent processing system based on large language model Technical Field The invention relates to the technical field of text reply generation, in particular to a purchase challenge full-flow intelligent processing system based on a large language model. Background In the fields of government purchasing, enterprise purchasing and the like, when a provider makes a question about a purchasing file, a purchasing result and the like, a purchasing manager (such as a purchasing agency and a purchasing person) needs to conduct compliance review and reply. The question processing in the current industry mainly depends on manual whole-flow operation, and the specific technical scheme and flow are as follows: (1) The receiving and analyzing stage of the questioning function includes that a purchasing manager manually receives the questioning function submitted by a provider (mostly PDF, WORD format or paper scanning piece), extracts key information through manual reading, and includes questioning unit names, whether to participate in project purchasing, questioning properties (aiming at bid-drawing files/results/processes and the like), questioning core contents, fact bases, law and regulation references and the like, and meanwhile, compliance of the questioning function submitting (such as whether submitting time is within legal period and whether form accords with regulations) needs to be manually checked. (2) And in the basic information matching and checking stage, sponsors need to manually inquire the station account of the purchased item, match the corresponding item number, item name, purchaser, purchase file and other background information, check the integrity of the questioning information one by one (if whether the questioning unit qualification and questioning problem description are omitted clearly or not), and manually check whether the legal and regulation clauses of the questioning reference are accurate or not. The reply basis retrieval and arrangement stage is that, aiming at the problem of question, the sponsor needs to acquire the reply basis by manually browsing or retrieving a locally stored purchasing rule base and a bid-posting rule base (most of PDF integrated files are unstructured indexes), manually searching historical question cases (most of files are filed in a document form, no classification labels or similarity matching functions are needed), and if external information such as enterprise qualification, product parameters and the like needs to be verified, the sponsor needs to manually log in corresponding official websites to inquire and screen capture and save. (3) And the reply content writing and auditing stage comprises the steps that a manager manually writes pieces of reply content according to the retrieved basis and referring to a general question reply template, and after completion, the accuracy of reply logic, regulation quotation and case matching is confirmed through multi-level manual auditing, and finally a question reply letter is formed and led out. (4) And in the archiving stage, the processed suspicious functions, the reply functions and the related materials are manually classified and stored in an archive (electronic or paper), and no mechanism for automatically associating the suspicious functions, the reply functions and the related materials with a case library exists. The prior art has the core defects that: (1) The efficiency is low, the whole process depends on manual operation, only the extracting of the questioning function elements and the searching link of the legal case take a plurality of hours (complex projects even 1-2 days), and the legal requirement of short-term junction handling of purchasing questioning cannot be met (for example, the requirement of 7 working days after the questioning is received is specified in a government purchasing questioning and complaining method). (2) The accuracy is insufficient, key information (such as incorrect judgment of questioning properties and incomplete extraction of facts basis) is easy to miss by manually extracting questioning elements, rules and regulations are dependent on experience of sponsors, problems such as clause outdated and wrong quotation can occur, and the fact that historical case matching lacks a standardized method easily causes inconsistent reply logic and past cases. (3) The traceability is poor, the basis of the reply content (such as cited regulation clauses and historical cases) needs manual marking, the reply content cannot be rapidly positioned to the original file, the query result of the external information (such as enterprise qualification validity) lacks an automatic retention and association mechanism, and repeated operation is needed in the subsequent auditing or review. (4) The knowledge base management is disordered, core data such as purchasing regulations, historical cases and the like are stored in the form of unstructured documents (s