A method, system and electronic device for tender document compliance review

Abstract

The application relates to a bidding document compliance review method and system, which solves the problems of poor flexibility of a traditional rule system, insufficient accuracy and logical consistency of a single deep learning model in professional field knowledge, and low processing efficiency of a super-long and complex bidding document in the prior art by constructing a collaborative work architecture composed of multiple professional intelligent agents such as intelligent interaction, clause compliance, quantitative indicators and real-time data verification, and combining RAG technology to realize accurate retrieval and deep fusion of regulation knowledge. Through the specialization and collaborative decision of each intelligent agent, various compliance risks such as tendency, exclusivity and logical contradiction in the document can be efficiently and comprehensively identified, and an interpretable review report containing problem positioning, judgment basis and modification suggestions is generated, which can significantly improve the accuracy and efficiency of the review, has the ability to continuously learn and quickly adapt to changes in regulations, and can provide an intelligent solution with high precision and high efficiency for the bidding supervision field.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a method, system, electronic device, and computer-readable storage medium for reviewing the compliance of tender documents. Background Technology

[0002] In procurement, engineering construction, and other bidding activities, the compliance of bidding documents is a prerequisite for ensuring the fairness, impartiality, and transparency of the bidding process. Currently, the compliance review of bidding documents mainly relies on manual methods. Reviewers need to examine each clause in the bidding documents according to relevant regulations, as well as numerous industry standards and local norms, to identify potential issues such as bias, exclusivity, or illegal provisions. However, bidding documents are typically lengthy (often exceeding hundreds of pages and hundreds of thousands of words), highly specialized, and contain numerous compliance checkpoints.

[0003] Against this backdrop, traditional manual review methods have revealed several prominent problems: First, review efficiency is low, with reviewing a complex document typically taking hours or even days; second, the rate of missed detection is high, as reviewers are easily influenced by subjective experience and fatigue, making it difficult to comprehensively identify all compliance risks, especially for highly concealed violations that rely on contextual judgment; third, review standards are difficult to standardize, as differences in the professional level and judgment criteria of different reviewers lead to insufficient consistency in review results; finally, knowledge updates are lagging behind, as laws, regulations, and industry standards are frequently updated, requiring reviewers to continuously learn, while the speed of manual knowledge updates cannot keep up with the speed of changes in laws and policies.

[0004] To address these issues, the industry has explored various technologies. For example, review systems based on keyword matching or fixed rules, while offering some interpretability, suffer from poor flexibility, struggle to adapt to varying industry and regional regulations, and incur high costs for building and maintaining rule bases. Methods based on traditional machine learning or statistical learning are limited by their ability to understand long texts, making it difficult to accurately capture complex contextual semantics and long-distance dependencies between clauses. Furthermore, their generalization ability is weak, and they heavily rely on high-quality labeled data. While deep learning models have made progress in semantic understanding in recent years, they still face challenges in terms of the accuracy and logical consistency of knowledge within specialized domains. Their "black box" nature also results in a lack of sufficient interpretability in review results, hindering effective support for subsequent regulatory decisions.

[0005] Currently, no effective solutions have been proposed to address the problems of low efficiency and poor accuracy in reviewing bidding documents and the inability to adapt to dynamic updates of regulations in related technologies. Summary of the Invention

[0006] This application provides a method, system, electronic device, and computer-readable storage medium for reviewing the compliance of bidding documents, in order to at least solve the problem of low efficiency in the review of bidding documents in related technologies.

[0007] In a first aspect, embodiments of this application provide a method for reviewing the compliance of bidding documents, the method comprising: Obtain the tender documents to be reviewed and perform structured parsing to obtain multiple text fragments in the tender documents; Based on the text fragment, a retrieval enhancement generation mechanism is used to retrieve multiple relevant compliance bases that match the text fragment from a preset legal knowledge base and industry rule base; The text fragments are reviewed in parallel using a multi-agent collaborative review system based on the relevant compliance criteria, generating compliance review results.

[0008] In some embodiments, an enhanced generation mechanism is used to retrieve multiple relevant compliance bases matching the text fragment from a preset regulatory knowledge base and industry rule base, including: The text fragment is vectorized using a pre-trained language model that matches the domain of the bidding documents to obtain a text vector representation of the text fragment; By employing a multi-path recall-fusion ranking retrieval strategy, semantic similarity retrieval is performed on the preset regulatory knowledge base and the industry rule base based on the text vector representation to obtain the relevant compliance basis.

[0009] In some embodiments, a multi-path recall-fusion ranking retrieval strategy is used to perform semantic similarity retrieval based on the text vector representation in the preset regulatory knowledge base and the industry rule base to obtain the relevant compliance basis, including: A dense retrieval tool based on BERT is used to obtain the first target fragment by calculating the similarity between the text vector representation and the query vectors in the preset legal knowledge base and industry rule base. Based on preset compliance monitoring points, sparse keyword retrieval is performed from preset legal knowledge base and industry rule base to obtain the second target fragment; By learning the ranking algorithm, the first target fragment and the second target fragment are fused and ranked according to multi-dimensional reference factors, and the relevant compliance basis is output according to the ranking result. The multi-dimensional reference factors include: semantic relevance, keyword matching degree, and text importance.

[0010] In some embodiments, prompt word engineering is used to integrate the text fragments and the relevant compliance evidence to form input information for analysis by the multi-agent collaborative review system, which includes: The intelligent interaction and semantic understanding agent is used to perform deep semantic analysis, key information extraction, background context construction and standardization processing on the text fragments through a large language model, and output standardized text fragments. A rule-based intelligent agent is used to perform format compliance review on the standardized text fragments based on the relevant compliance criteria using a large language model. The fact-checking intelligent agent is used to perform factual consistency review, correlation review, and market price review on the standardized text fragments through a trusted database; The decision-making agent is used to collaboratively judge the review results output by the rule-making agent and the fact-verification agent, and output the compliance review result.

[0011] In some embodiments, the rule-based agent includes: The clause compliance review intelligent agent is used to determine whether the standardized text fragment contains biased clauses, exclusive clauses, and ambiguous clauses, as well as whether there are contradictory clauses. A quantitative indicator review agent is used to determine whether the values ​​of the standardized text fragments conform to the corresponding mandatory specifications. The fact-checking agent is used to perform factual consistency checks on the text fragment, including: Structured fact elements are extracted from the standardized text fragments using a named entity recognition model, the fact elements are vectorized, and a fact consistency graph is constructed based on the vectorization results. A contradiction detection algorithm based on graph reasoning is performed on the fact consistency graph to identify factual logical conflicts in the standardized text fragments.

[0012] In some embodiments, the decision-making agent is used to collaboratively judge the review results output by multiple rule agents and fact-checking agents in the multi-agent collaborative review system, including: The review results output by the rule-making agent and the fact-verifying agent are compared and verified to identify conflicts or inconsistencies in the results; Furthermore, based on preset decision logic and weighted algorithm, a comprehensive decision is made according to the review results output by the rule-based intelligent agent and the fact-verification intelligent agent to obtain the compliance review result of the text fragment.

[0013] In some embodiments, the method further includes: In the multi-agent collaborative review system, a fine-tuning method is used to train and optimize the large language model running in the system, wherein the fine-tuning method includes: Construct a labeled dataset to perform fine-grained annotation of compliance issues in the bidding documents; The large language model is fine-tuned using an improved loss function, which is expressed by the following formula:

[0014] Where t is the summation iteration variable, m+n represents the total length of the model output sequence, and P(xt|x1...xt-1) represents the conditional probability. The symbol represents the indicator function, and T represents the template text content with high repetition in the bidding documents.

[0015] In some embodiments, the method further includes: After generating the compliance review results, obtain feedback from external experts on the non-compliance points in the compliance review results, wherein the feedback information includes suggestions for correction or annotation information on the non-compliance points; Based on the feedback information, the preset knowledge base, rule base, or model parameters in the multi-agent collaborative review system are updated in a targeted manner to achieve continuous optimization of the multi-agent collaborative review system.

[0016] In some embodiments, the tender documents to be reviewed are obtained and subjected to structured parsing to obtain multiple text fragments in the tender documents, including: The bidding document is processed by a pre-trained document layout analysis model to obtain regional analysis results, and each content region in the bidding document is processed by a parser corresponding to the regional analysis results to obtain parsing results. A multi-level cleaning and standardization strategy is adopted to preprocess the text in the bidding documents based on the parsing results to obtain standardized text; The standardized text is segmented, and the results of the segmentation are classified to obtain multiple text fragments in the bidding document.

[0017] Secondly, this application provides a bidding document compliance review system, the system comprising: an acquisition module, a retrieval enhancement module, and a review module, wherein: The acquisition module is used to acquire the bidding documents to be reviewed and perform structured parsing on them to obtain multiple text fragments in the bidding documents; The retrieval enhancement module is used to retrieve multiple relevant compliance bases that match the text fragment from a preset legal knowledge base and industry rule base through a retrieval enhancement generation mechanism, based on the text fragment. The review module is used to perform parallel collaborative review of the text fragments based on the relevant compliance criteria through a multi-agent collaborative review system, and generate compliance review results.

[0018] Compared to related technologies, the bidding document compliance review system and method provided in this application constructs a collaborative working architecture composed of multiple professional intelligent agents, including intelligent interaction, clause compliance, quantitative indicators, and real-time data verification. Combined with RAG technology, it achieves accurate retrieval and deep integration of regulatory knowledge, solving problems such as poor flexibility of traditional rule systems, insufficient accuracy and logical consistency of single deep learning models in professional domain knowledge, and low efficiency in processing extremely long and complex bidding documents. Through the professional division of labor and collaborative decision-making of each intelligent agent, it can efficiently and accurately identify various compliance risks in documents, such as bias, exclusivity, and logical contradictions, generating an interpretable review report containing problem location, judgment basis, and modification suggestions. This significantly improves review accuracy and efficiency while possessing the ability to continuously learn and quickly adapt to regulatory changes. It can provide a high-precision and high-efficiency intelligent solution for the bidding supervision field. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a method for reviewing the compliance of bidding documents according to an embodiment of this application; Figure 2 This is a structural block diagram of a bidding document compliance review system according to an embodiment of this application; Figure 3 This is an architecture diagram of a bidding document compliance review system according to an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0021] Obviously, the accompanying drawings described below are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, any changes to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0022] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0023] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects are in an "or" relationship. The terms "first," "second," and "third" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0024] This application provides a method and system for compliance review of bidding documents based on retrieval-enhanced generation and multi-agent collaboration. This solution aims to achieve efficient, accurate, and interpretable compliance review of bidding documents by constructing an automated intelligent review process, thereby addressing the problems of low efficiency, high missed detection rate, and inconsistent standards in existing manual reviews.

[0025] Figure 1 This is a flowchart of a method for reviewing the compliance of bidding documents according to an embodiment of this application. Figure 1 As shown, the process may include the following steps: Step S101: Obtain the tender documents to be reviewed and perform structured parsing to obtain multiple text fragments in the tender documents.

[0026] Specifically, in bidding and tendering scenarios, the bidding and tendering documents to be reviewed come in various formats, including PDF, Word, Excel, and others, and have complex internal structures containing a large amount of unstructured content such as text, tables, and charts. To enable machines to accurately understand this content, this step aims to perform in-depth analysis and standardization of the original documents, providing a high-quality, machine-readable data foundation for subsequent review and analysis.

[0027] In one specific embodiment, the execution process of this step specifically includes: First, the bidding documents are processed by a pre-trained document layout analysis model to obtain regional analysis results. Then, the parser corresponding to the regional analysis results processes each content region in the bidding documents to obtain parsing results.

[0028] Specifically, this document layout analysis model, based on deep learning technology, can automatically identify different logical regions in a document, such as titles, body paragraphs, tables, charts, headers, and footers. Then, a specialized parser is invoked to process each identified region accordingly. For example, for table content, graph neural network-based table recognition technology is used to ensure that the original row and column structure and semantic information are fully preserved while extracting data from the table; for text in formulas and charts, optical character recognition (OCR) technology and image recognition algorithms are used for extraction and conversion.

[0029] The core approach in this processing step is to utilize deep learning models to analyze the visual and structural information of documents. This intelligently distinguishes different components of the document and extracts content in an optimized manner. This step plays a role in data standardization within the overall solution, transforming complex original bidding documents of various formats into a clean, organized, and structured data stream, providing high-quality input for subsequent retrieval enhancement and multi-agent review processes.

[0030] Furthermore, a multi-level cleaning and standardization strategy is employed to preprocess the text in the bidding documents based on the parsing results, resulting in standardized text. This process aims to eliminate interference from format differences and irrelevant information in subsequent analysis.

[0031] Specific actions may include: unifying encoding and standardizing format to ensure that all text has a consistent encoding format; using rule-based and machine learning methods to identify and remove watermarks, page numbers, advertisements, and other content that is not related to compliance review; and correcting spelling errors or formatting issues in the text.

[0032] Understandably, this preprocessing stage employs multi-level filtering and calibration techniques to ensure the purity and consistency of the input text. It removes noisy data and standardizes the text format, providing a foundation for subsequent vectorized representation and semantic understanding. Within the overall solution, this stage avoids model misjudgments caused by raw data quality issues, improves the stability and reliability of all subsequent intelligent analysis modules, and ensures accurate review.

[0033] Finally, the standardized text is segmented, and the results of the segmentation are classified to obtain multiple text fragments in the bidding documents.

[0034] Furthermore, to ensure that the inherent semantic connections are not disrupted while segmenting the text, this embodiment employs a semantic integrity detection algorithm to guarantee that each segmented text fragment is a relatively independent and complete unit semantically. After segmentation, each fragment undergoes preliminary content classification, for example, being labeled as categories such as "qualification requirements," "technical specifications," or "review criteria," facilitating subsequent precise retrieval and review.

[0035] The core approach to text slicing and segmentation in step S101 is to employ an intelligent segmentation algorithm that considers both physical and semantic boundaries, ensuring that each text segment is both of appropriate size and semantically complete. Decomposing extremely long documents, which are difficult to process directly, into a series of independently analyzable semantic units reduces the complexity of subsequent processing and creates conditions for parallel review and accurate problem localization, which is a technical prerequisite for achieving efficient system operation.

[0036] Step S102: Based on the text fragment, the retrieval enhancement generation mechanism retrieves multiple relevant compliance bases that match the text fragment from the preset legal knowledge base and industry rule base.

[0037] After completing the structured parsing of the text, the core objective of this step is to identify relevant compliance judgment criteria from laws, regulations, and industry standards for each text fragment to be reviewed. This application embodiment employs a search-enhanced generation mechanism. This mechanism combines external, authoritative knowledge sources with internal analysis processes to ensure that review judgments are based on verifiable evidence, thereby significantly improving the accuracy and credibility of the system.

[0038] In one specific embodiment, the execution process of this mechanism includes: S1. Using a pre-trained language model matched to the bidding document domain, the text fragments are vectorized to obtain text vector representations. To better capture the professional terminology and semantics of the bidding domain, this embodiment prioritizes using domain-adapted pre-trained language models, such as models further pre-trained on legal and bidding domain texts (e.g., Legal-BERT, Finance-BERT, etc.). These models can encode the deep semantic information of text fragments into high-dimensional mathematical vectors, i.e., text vector representations.

[0039] This vectorization process utilizes domain-adaptive deep learning models to convert natural language text into mathematical representations that machines can understand and compute. By understanding the meaning of the clauses under review, it becomes possible to find legal grounds that are superficially dissimilar but highly related in content, potentially uncovering deeper, implicit compliance issues.

[0040] S2 employs a multi-path recall-fusion ranking retrieval strategy, performing semantic similarity searches based on text vector representations within a pre-defined regulatory knowledge base and industry rule base to obtain relevant compliance evidence. This strategy combines the advantages of different retrieval methods to ensure the comprehensiveness and accuracy of the search results. Specifically, this strategy includes: A BERT-based dense search engine is employed to obtain the first set of target fragments by calculating the similarity between the text vector representation and the query vectors of entries in a pre-defined regulatory knowledge base and industry rule base. The dense search engine excels at understanding semantic similarity and can identify regulatory clauses that are semantically similar to the text fragments to be reviewed but may use completely different wording, thereby uncovering implicit compliance issues.

[0041] Based on preset compliance monitoring points, sparse keyword retrieval is performed from a preset regulatory knowledge base and industry rule base to obtain a second set of target fragments. Among them, sparse retrieval (such as the BM25 algorithm) focuses on precise keyword matching and can serve as an effective supplement to dense retrieval, ensuring that clauses containing specific legal terms or key indicators are not missed.

[0042] Furthermore, by learning a ranking algorithm, the first and second target fragment sets are fused and ranked based on multi-dimensional reference factors, and relevant compliance evidence is output based on the ranking results. After recalling the two target fragment sets, a learned ranking algorithm is used to intelligently fuse and re-rank them.

[0043] The algorithm comprehensively considers multiple reference factors, including semantic relevance, keyword matching, and text importance (e.g., the level of laws and regulations, the mandatory nature of clauses), and finally outputs a list of compliance evidence sorted from high to low relevance.

[0044] In this embodiment, multi-path recall and fusion ranking combine the depth of semantic retrieval with the precision of keyword retrieval, and use a machine learning model to intelligently select and rank the search results. This ensures that not only relevant legal basis can be found, but also the most relevant and authoritative basis, avoiding misjudgments caused by incomplete information or incorrect priority judgment, and greatly improving the accuracy and efficiency of subsequent review processes.

[0045] Finally, through prompt word engineering, the text fragments and relevant compliance evidence are integrated to form input information for analysis by the multi-agent collaborative review system. This step organizes the original text fragments to be reviewed and the most relevant compliance evidence retrieved according to a preset template format, forming a structured and information-rich prompt. This prompt will serve as input for the next step of multi-agent review, guiding the agents to make judgments within a context with clear evidence.

[0046] Step S103: Through a multi-agent collaborative review system, the text fragments are reviewed in parallel based on relevant compliance criteria to generate compliance review results.

[0047] In this embodiment, this step does not rely on a single, generalized large language model for judgment, but constructs a collaborative working system composed of multiple highly specialized agents. Each agent acts as a virtual domain expert, receiving input information from the integrated compliance evidence generated in step S102, reviewing text fragments in parallel from their respective professional perspectives, and ultimately forming a unified review result through a collaborative mechanism. This architecture, through specialized division of labor and cross-validation, achieves a balance between the depth and breadth of complex review tasks.

[0048] In one specific embodiment, the multi-agent collaborative review system includes: Intelligent Interaction and Semantic Understanding Agent: This agent acts as the system's "first point of contact" and information hub, responsible for the deep processing and distribution of input information. Through a large language model, it performs deep semantic analysis, key information extraction, and background context construction and standardization on text fragments, outputting standardized text fragments. Specifically, it is used for: Deep semantic understanding: Utilize large models to analyze the deep semantics of bidding documents and understand the true intentions behind the bidding purpose, technical requirements, and commercial terms.

[0049] Information extraction and structuring: Accurately extract key entities and information from unstructured text, such as purchaser information, project name, budget amount, key technical parameters, qualification requirements list, scoring criteria details, etc., and transform them into structured data.

[0050] Contextualization: Provides the agents with a summary of background information and contextual information required for subsequent special review, ensuring that each agent works within a unified cognitive framework.

[0051] The core approach of this intelligent agent is to leverage the natural language understanding capabilities of a large language model to deeply process the original text, transforming it from a simple sequence of words into knowledge units containing rich semantic and structured information. It acts as a task parsing and distribution center in the overall solution. Through preprocessing, subsequent specialized intelligent agents can directly handle standardized and easily understood task inputs, greatly improving the operational efficiency and accuracy of the entire collaborative system.

[0052] Rule-based intelligent agent: This agent focuses on reviewing the compliance of text content based on retrieved compliance criteria. In a more specific embodiment, the rule-based intelligent agent can be further subdivided into: The clause compliance review agent focuses on identifying implicit, biased, exclusive, or ambiguous clauses in the text, as well as logical contradictions between clauses. For example, leveraging the semantic understanding and logical reasoning capabilities of a large language model, it detects unreasonable anti-competitive statements such as "specifying a specific brand" or "requiring specific qualifications unrelated to the actual needs of the project"; identifies unquantified and vague standards such as "having extensive relevant experience" or "well-known brand products"; and compares clauses before and after in the document for contradictions, such as whether technical standards match acceptance criteria.

[0053] Quantitative Indicator Review Agent: This sub-agent focuses on determining whether the values ​​of standardized text fragments comply with corresponding mandatory regulations. It extracts all quantifiable indicators from the text, such as bid bond amount, performance bond ratio, and defect liability period (warranty period), and automatically compares them with a built-in regulatory knowledge base and industry standards to quickly identify whether they exceed the legally mandated upper and lower limits. Simultaneously, it also reviews whether the thresholds set for a company's registered capital, financial status, and similar project performance are unreasonably high or discriminatory.

[0054] In this embodiment, the rule-based intelligent agent combines the reasoning capabilities of a large model with a structured regulatory knowledge base to perform in-depth and fine-grained compliance assessments. The overall solution identifies the most time-consuming and error-prone aspects of manual review from both qualitative and quantitative perspectives. This ensures the comprehensiveness and depth of the review, enabling the discovery of hidden violations that are difficult for humans to detect.

[0055] Fact-checking agent: This agent uses a trusted database to perform factual consistency checks, correlation checks, and market price checks on standardized text fragments. It enhances the objectivity and credibility of the review results, specifically including: Factual consistency verification: By connecting to external official databases, information such as the status of enterprise qualification certificates, the validity period of safety production licenses, and administrative penalty records mentioned in documents is verified in real time. Its technical implementation may include: using a Named Entity Recognition (NER) model to extract structured factual elements (such as company names, certificate numbers, etc.) from standardized text fragments, vectorizing these factual elements, and constructing a factual consistency graph; subsequently, by executing a graph reasoning contradiction detection algorithm on the factual consistency graph, logical conflicts in the facts within the standardized text fragments are identified.

[0056] Related party relationship review: By querying databases such as enterprise registration information, we analyze whether there are any related party relationships such as controlling stakes or management among different bidders, and alert users to the risks.

[0057] Market price review: When necessary, access to the market price database can be used to make a preliminary judgment on the reasonableness of the budget, maximum price, etc. of key products or services set in the bidding documents.

[0058] In this embodiment, the fact-checking agent connects to an external database in real time via an API interface for cross-validation. The overall solution provides objective data support for the review conclusions. This expands the review process beyond the text itself to verify real-world facts, fundamentally enhancing the reliability and authority of the review results and effectively preventing compliance judgments based on false information.

[0059] Decision-making agent: This agent collaboratively evaluates the review results output by the rule-making agent and the fact-checking agent, and outputs the final compliance review result. Its workflow includes: First, the review results output by the rule-making agent and the fact-checking agent are compared and verified to identify conflicts or inconsistencies. For example, the rule-making agent may consider a certain technical parameter setting reasonable, but the fact-checking agent finds that the parameter points to a single supplier. In this case, the decision-making agent needs to identify this conflict.

[0060] Furthermore, based on a pre-defined decision-making logic or weighted algorithm, a comprehensive decision is made to obtain the compliance review result of the text fragment. This decision-making logic can set priorities and judgment rules according to factors such as risk level and problem nature, ultimately forming a unified and clear review conclusion, including problem description, risk level, and judgment basis.

[0061] In an optional embodiment, to continuously improve the performance of the large language model in the multi-agent collaborative review system, this method also provides a model training and optimization step. Specifically, firstly, a high-quality labeled dataset is constructed to perform fine-grained annotations on various compliance issues in historical bidding documents, including information such as issue type, issue description, and original text location. Then, when fine-tuning the large language model on which the multi-agent collaborative review system is based, an improved loss function is used. The mathematical expression of this improved loss function is:

[0062] Where t is the summation iteration variable, m+n represents the total length of the model output sequence, and P(xt|x1...xt-1) represents the conditional probability. This is an indicator function, where T represents the repetitive template text content in the bidding documents. This design allows the model to ignore recurring, formulaic text with no review value (such as general statements and standard format instructions) in a large number of documents during training. This allows the learning resources to be more focused on understanding and analyzing substantive and changeable review content, significantly improving the efficiency and effectiveness of model fine-tuning.

[0063] Finally, the solution also provides a results generation and feedback management module, which presents the results of multi-agent collaborative review to users in a clear and interpretable format and establishes a closed-loop optimization mechanism. Review result reports typically include a risk overview, detailed problem descriptions (with original text location), and specific regulatory recommendations and amendments. Simultaneously, the system provides a feedback management mechanism, allowing expert users to evaluate and revise the review results. This expert feedback information is recorded by the system and used for the next round of model iteration and knowledge base updates, thus forming a continuously learning and evolving adaptive optimization closed loop.

[0064] Through steps S101 to S103 above, compared with the prior art, this application provides a method for reviewing the compliance of bidding documents. By constructing a collaborative working architecture composed of multiple professional intelligent agents, including intelligent interaction, clause compliance, quantitative indicators, and real-time data verification, and combining RAG technology to achieve accurate retrieval and deep integration of regulatory knowledge, this method solves the problems of poor flexibility of traditional rule systems, insufficient accuracy and logical consistency of single deep learning models in professional domain knowledge, and low efficiency in processing extremely long and complex bidding documents in the prior art. Through the professional division of labor and collaborative decision-making of each intelligent agent, it can efficiently and comprehensively identify various compliance risks in the documents, such as bias, exclusivity, and logical contradictions, and generate an interpretable review report containing problem location, judgment basis, and modification suggestions. While significantly improving the accuracy and efficiency of review, it also has the ability to continuously learn and quickly adapt to changes in regulations. It can provide a high-precision and high-efficiency intelligent solution for the field of bidding supervision.

[0065] Secondly, this application also provides a system for reviewing the compliance of bidding documents. Figure 2 This is a structural block diagram of a bidding document compliance review system according to an embodiment of this application, such as... Figure 2 As shown, the system includes: an acquisition module 20, a retrieval enhancement module 21, and an examination module 22, wherein: The acquisition module 20 is used to acquire the bidding documents to be reviewed and perform structured parsing to obtain multiple text fragments in the bidding documents; The retrieval enhancement module 21 is used to retrieve and obtain multiple relevant compliance bases corresponding to the text fragments from the preset legal knowledge base and industry rule base through the retrieval enhancement generation mechanism. The review module 22 is used to conduct parallel collaborative review of text fragments based on relevant compliance criteria through a multi-agent collaborative review system, and generate compliance review results.

[0066] Figure 3 This is an architecture diagram of a bidding document compliance review system according to an embodiment of this application, such as... Figure 3As shown, the uploaded bidding documents first enter the document parsing and preprocessing module for text extraction, cleaning, standardization, and vectorization. The processed data is then fed into a vector database and knowledge base, and enhanced with RAG retrieval technology to enrich contextual information. Further, the multi-agent collaborative review module includes an agent orchestrator responsible for coordinating the parallel work of multiple specialized agents. Additionally, collaborative review is conducted through multiple agents, including intelligent interaction and semantic understanding agents, clause compliance review agents, quantitative indicator review agents, and real-time data verification agents. After review, the result generation and feedback management module generates a review report. This module also features a comment feedback function; the feedback information is used for continuous optimization of the knowledge base and review module, as well as model optimization iterations, forming a closed-loop system of continuous learning and improvement.

[0067] Compared to existing technologies, the above system effectively handles extremely long and complex bidding documents through structured parsing and knowledge enhancement techniques; it utilizes a multi-agent collaborative architecture to achieve multi-dimensional and professional parallel review of text content, improving the depth and breadth of the review; and it combines interpretable result generation and feedback optimization mechanisms to solve the technical problems of low efficiency, insufficient accuracy, and inconsistent standards in traditional review methods, achieving efficient, accurate, and interpretable intelligent review of the compliance of bidding documents.

[0068] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0069] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method of tender document compliance review, characterized by, The method includes: Obtain the tender documents to be reviewed and perform structured parsing to obtain multiple text fragments in the tender documents; Based on the text fragment, a retrieval enhancement generation mechanism is used to retrieve multiple relevant compliance bases that match the text fragment from a preset legal knowledge base and industry rule base; The text fragments are reviewed in parallel using a multi-agent collaborative review system based on the relevant compliance criteria, generating compliance review results.

2. The method of claim 1, wherein, By employing a retrieval-enhanced generation mechanism, multiple relevant compliance bases matching the text fragment are retrieved from a pre-defined regulatory knowledge base and industry rule base, including: The text fragment is vectorized using a pre-trained language model that matches the domain of the bidding documents to obtain a text vector representation of the text fragment; By employing a multi-path recall-fusion ranking retrieval strategy, semantic similarity retrieval is performed on the preset regulatory knowledge base and the industry rule base based on the text vector representation to obtain the relevant compliance basis.

3. The method of claim 2, wherein, Using a multi-path recall-fusion ranking retrieval strategy, semantic similarity retrieval is performed based on the text vector representation in the preset regulatory knowledge base and the industry rule base to obtain the relevant compliance evidence, including: A dense retrieval tool based on BERT is used to obtain the first target fragment by calculating the similarity between the text vector representation and the query vectors in the preset legal knowledge base and industry rule base. Based on preset compliance monitoring points, sparse keyword retrieval is performed from preset legal knowledge base and industry rule base to obtain the second target fragment; By learning the ranking algorithm, the first target fragment and the second target fragment are fused and ranked according to multi-dimensional reference factors, and the relevant compliance basis is output according to the ranking result. The multi-dimensional reference factors include: semantic relevance, keyword matching degree, and text importance.

4. The method according to claim 1, characterized in that, Through prompt word engineering, the text fragments and relevant compliance evidence are integrated to form input information for analysis by the multi-agent collaborative review system, which includes: The intelligent interaction and semantic understanding agent is used to perform deep semantic analysis, key information extraction, background context construction and standardization processing on the text fragments through a large language model, and output standardized text fragments. A rule-based intelligent agent is used to perform format compliance review on the standardized text fragments based on the relevant compliance criteria using a large language model. The fact-checking intelligent agent is used to perform factual consistency review, correlation review, and market price review on the standardized text fragments through a trusted database; The decision-making agent is used to collaboratively judge the review results output by the rule-making agent and the fact-verification agent, and output the compliance review result.

5. The method according to claim 4, characterized in that, The rule-based intelligent agent includes: The clause compliance review intelligent agent is used to determine whether the standardized text fragment contains biased clauses, exclusive clauses, and ambiguous clauses, as well as whether there are contradictory clauses. A quantitative indicator review agent is used to determine whether the values ​​of the standardized text fragments conform to the corresponding mandatory specifications. The fact-checking agent is used to perform factual consistency checks on the standardized text fragments, including: Structured fact elements are extracted from the standardized text fragments using a named entity recognition model, the fact elements are vectorized, and a fact consistency graph is constructed based on the vectorization results. By performing a contradiction detection algorithm based on graph reasoning on the fact consistency graph, factual logical conflicts in the standardized text fragments are identified.

6. The method according to claim 4, characterized in that, The decision-making agent is used to collaboratively judge the review results output by multiple rule agents and fact-checking agents in the multi-agent collaborative review system, including: The review results output by the rule-making agent and the fact-verifying agent are compared and verified to identify conflicts or inconsistencies in the results; Furthermore, based on preset decision logic and weighted algorithm, a comprehensive decision is made according to the review results output by the rule-based intelligent agent and the fact-verification intelligent agent to obtain the compliance review result of the text fragment.

7. The method according to claim 4, characterized in that, The method further includes: In the multi-agent collaborative review system, a fine-tuning method is used to train and optimize the large language model running in the system, wherein the fine-tuning method includes: Construct a labeled dataset to perform fine-grained annotation of compliance issues in the bidding documents; The large language model is fine-tuned using an improved loss function, wherein the calculation logic of the improved loss function is configured as follows: For each current output unit in the model output sequence, calculate the negative logarithm of the conditional probability of generating the current output unit given the historical output sequence; The system determines whether the current output unit belongs to the text content of a preset template text set based on the indicated conditions. When calculating the total loss of the model output sequence, the negative logarithms of the conditional probabilities corresponding to the current output units that do not belong to the preset template text set are summed.

8. The method according to claim 1, characterized in that, The method further includes: After generating the compliance review results, obtain feedback from external experts on the non-compliance points in the compliance review results, wherein the feedback information includes suggestions for correction or annotation information on the non-compliance points; Based on the feedback information, the preset knowledge base, rule base, or model parameters in the multi-agent collaborative review system are updated in a targeted manner to achieve continuous optimization of the multi-agent collaborative review system.

9. The method according to claim 1, characterized in that, The bidding documents to be reviewed are obtained and structuredly parsed to obtain multiple text fragments in the bidding documents, including: The bidding document is processed by a pre-trained document layout analysis model to obtain regional analysis results, and each content region in the bidding document is processed by a parser corresponding to the regional analysis results to obtain parsing results. A multi-level cleaning and standardization strategy is adopted to preprocess the text in the bidding documents based on the parsing results to obtain standardized text; The standardized text is segmented, and the results of the segmentation are classified to obtain multiple text fragments in the bidding document.

10. A system for reviewing the compliance of bidding documents, characterized in that, The system includes: an acquisition module, a retrieval enhancement module, and an examination module, wherein: The acquisition module is used to acquire the bidding documents to be reviewed and perform structured parsing on them to obtain multiple text fragments in the bidding documents; The retrieval enhancement module is used to retrieve multiple relevant compliance bases that match the text fragment from a preset legal knowledge base and industry rule base through a retrieval enhancement generation mechanism, based on the text fragment. The review module is used to perform parallel collaborative review of the text fragments based on the relevant compliance criteria through a multi-agent collaborative review system, and generate compliance review results.

Images