Multi-mode intelligent auditing method for supply chain purchase bid invitation project

Abstract

The invention provides a supply chain purchase bid invitation project multi-mode intelligent auditing method, and belongs to the technical field of electronic purchase. Comprising the following steps: defining a review rule system comprising rule contents, review points and review logic; in the bidding stage, bidding files and quotation information uploaded by suppliers are received, and after bid opening, bid invitation files, bidding files of entered suppliers and corresponding review rule systems are input into a multi-mode large model server; performing unstructured data analysis on the bidding document and the bidding document by using a multi-modal large model in combination with an OCR (Optical Character Recognition) technology, and identifying and matching response contents in the bidding document through intention; pushing the extracted content to a corresponding business auditing agent, and performing judgment according to an auditing point and auditing logic; and summarizing the review results, and generating an intelligent review report containing the number of conformity items, the number of non-conformity items, conclusions of the review points and judgment bases. The examination efficiency and accuracy are improved, and human errors and compliance risks are reduced.

Description

Technical Field

[0001] This application belongs to the field of electronic procurement technology, specifically involving a multimodal intelligent auditing method for supply chain procurement bidding projects. Background Technology

[0002] Projects in large-scale infrastructure, energy, transportation and other fields are becoming increasingly large in scale. Enterprise procurement bidding documents are often thousands of pages long, involving complex content from multiple dimensions such as technology, commerce, qualifications, and pricing. The traditional method of relying on manual review of each item is no longer able to meet the requirements of efficient, accurate and compliant review.

[0003] Review experts are prone to fatigue under high-intensity work, leading to frequent omissions and misjudgments. Furthermore, subjective differences in the interpretation of clauses among different experts affect the fairness and consistency of the evaluation process. In addition, tender documents may contain implicit exclusive or biased clauses, and bid rigging and collusion are highly concealed, making them difficult to fully identify using only human experience, thus continuously increasing compliance risks. Summary of the Invention

[0004] To address at least one of the technical problems existing in the background art, this application provides a multimodal intelligent review method for supply chain procurement bidding projects. By integrating multimodal large model, OCR and intent recognition technologies, it realizes automated and intelligent review of enterprise procurement bidding documents, improves review efficiency and accuracy, and reduces human error and compliance risks.

[0005] The technical solution adopted in this application is as follows: The first aspect of this application provides a multimodal intelligent auditing method for supply chain procurement bidding projects, including: Define a review rule system that includes rule content, review points, and review logic; During the bidding stage, the system receives the bid documents and price information uploaded by suppliers. After the bid opening, the system inputs the bidding documents, the bid documents of shortlisted suppliers, and the corresponding review rules into the multimodal large model server. The unstructured data of the tender documents and bid documents are parsed using a multimodal large model combined with OCR technology. The review requirements are extracted from the tender documents according to the review rules, and the response content in the bid documents is matched by intent recognition. The extracted content is pushed to the corresponding business review AI agent for evaluation based on review points and review logic; The review results are summarized to generate an intelligent review report that includes the number of compliant items, the number of non-compliant items, the conclusions of each review point, and the basis for the judgment.

[0006] According to one embodiment of this application, the definition includes a review rule system comprising rule content, review points, and review logic, specifically as follows: During the project initiation phase, after the bidding party creates the procurement project and uploads the bidding documents, it configures a review rule system applicable to the project based on industry standards, laws and regulations, and corporate internal control requirements. The review rule system includes multiple structured review rules. Each review rule includes a rule number, rule name, scope of application, rule content text, review points to be checked, corresponding judgment logic expression, and an identifier indicating whether it is a veto item.

[0007] According to one embodiment of this application, during the bidding stage, receiving the bid documents and price information uploaded by the suppliers, and after the bid opening, inputting the bidding documents, the bid documents of the shortlisted suppliers, and the corresponding review rule system into the multimodal large model server, specifically: Before the bid submission deadline, the system will open an interface for suppliers to upload encrypted bid documents and price quotes. During the bid opening process, the system decrypts and verifies the authenticity and format compliance of all bid documents. For bid documents that pass the preliminary verification, they are packaged with the corresponding bidding documents and the configured review rule system. Using automated scripts or triggering mechanisms, the packaged dataset is pushed to the multimodal large model server after the bidding is completed; Each bid document is analyzed based on a built-in algorithm, and pre-defined information is extracted and compared with the requirements in the bidding documents.

[0008] According to one embodiment of this application, the unstructured data parsing of the tender documents and the bid documents is performed using a multimodal large model combined with OCR technology, and the review requirements are extracted from the tender documents according to the review rules. Specifically, the response content in the bid documents is matched through intent recognition. The uploaded tender documents and bid documents are formatted. For scanned documents, image-type PDFs, or documents containing charts, OCR technology is used to convert them into editable text streams while preserving the original text and image layout information. The multimodal large model is used to jointly analyze the semantics of the converted text and images to identify the chapter structure, table content, technical parameters, qualification requirements, service commitments and quotation details in the document; Based on the configured review rules system, the review requirements related to each review point are extracted from the tender documents to form a structured review requirements list; In the tender documents, an intent recognition algorithm based on semantic understanding is used to locate the response paragraphs or data items that correspond to the above review requirements, thereby achieving content alignment across documents; During the intent recognition process, a contextual attention mechanism is used to determine whether the bid content is substantially responded to, partially responded to, or not responded to, and the matching confidence level is marked. Output structured data pairs, including a mapping table of "tender requirements - bid response", key field extraction results, and original document location coordinates.

[0009] According to one embodiment of this application, the review requirements list includes eligibility criteria, technical specifications, performance capabilities, response time, and pricing rules.

[0010] According to one embodiment of this application, the step of pushing the extracted content to the corresponding business review intelligent agent and judging it based on review points and review logic is as follows: The structured data pairs of "Bidding Requirements - Bid Responses" are distributed to the corresponding business review agents according to the categories of review rules. The business review agents include prequalification agents, business compliance agents, technical compliance agents, and price reasonableness agents. Based on the intelligent agents for each business review, the preset review points, and the formalized evaluation logic, the content of the bid response is logically judged. Verify the authenticity, completeness, consistency, and compliance of the response content; Output the evaluation results for each review point, including three statuses: "passed", "failed", and "requires manual review".

[0011] According to one embodiment of this application, it also includes: A horizontal comparative analysis of the intelligent review results of multiple shortlisted suppliers is conducted, and a structured comparison matrix is ​​generated using a multimodal large model to compare suppliers in terms of technical parameter response, service commitment level, price composition details, and qualification conditions. The comparison matrix is ​​presented in the form of a visual chart, marking the strengths, weaknesses and differences of each supplier, and automatically generating a summary of evaluation recommendations.

[0012] A second aspect of this application provides a multimodal intelligent auditing device for supply chain procurement bidding projects, comprising: The rule system definition module is suitable for defining a review rule system that includes rule content, review points, and review logic; The information processing module is suitable for receiving the tender documents and quotation information uploaded by suppliers during the bidding stage. After the bid opening, it inputs the tender documents, the tender documents of shortlisted suppliers and the corresponding review rules system into the multimodal large model server. The parsing module is suitable for using a multimodal large model combined with OCR technology to parse unstructured data from the tender documents and the bid documents, extracting review requirements from the tender documents according to review rules, and matching the response content in the bid documents through intent recognition; The judgment module is suitable for pushing the extracted content to the corresponding business review intelligent agent and judging it based on the review points and review logic; The results output module is suitable for summarizing review results and generating an intelligent review report that includes the number of compliant items, the number of non-compliant items, the conclusions of each review point, and the basis for the judgment.

[0013] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the multimodal intelligent auditing method for supply chain procurement bidding projects as described in any of the first aspects above.

[0014] This application also provides a non-volatile computer storage medium storing computer-executable instructions thereon, which, when executed by a processor, implement the multimodal intelligent auditing method for supply chain procurement bidding projects as described in any of the embodiments of the first aspect above.

[0015] Due to the adoption of the above technical solution, the beneficial effects achieved by this application are as follows: The multimodal intelligent review method for supply chain procurement bidding projects provided in the first aspect of this application achieves unified modeling and flexible configuration of bidding review standards by systematically constructing a structured review rule system. This ensures that the review basis is traceable and reusable, avoiding inconsistencies in judgment caused by misunderstandings in traditional manual review. During the bidding stage, the method receives bid documents and price information uploaded by suppliers and automatically inputs the bidding documents, shortlisted suppliers' bid documents, and corresponding review rule system into the multimodal large model server after the bid opening. This achieves automated triggering and data integration of the review process, reducing human intervention and operational delays. By utilizing the multimodal large model combined with OCR technology for deep analysis of unstructured documents, the method can accurately identify text, tables, and graphic layout information in scanned documents and image-type PDFs. It also extracts specific review requirements from the bidding documents through semantic understanding and then uses intent recognition technology to locate matching requirements in the bid documents. The response content achieves semantic alignment and substantive response judgment across documents, overcoming the limitations of traditional keyword matching, which is prone to misjudgment and omission. Subsequently, the parsed structured data of "bidding requirements - bid response" is pushed to the corresponding business review agents (such as qualification review, technical compliance, business compliance, and price reasonableness). Each agent performs automated evaluation based on preset review points and formal logic rules, improving the objectivity, consistency, and logical rigor of the review. Finally, the evaluation results of each agent are summarized to generate an intelligent review report containing the number of compliant items, the number of non-compliant items, the conclusions of each review point, and detailed judgment basis. This not only significantly shortens the preparation time for expert evaluation but also provides a clear and traceable chain of evidence for subsequent manual review. Overall, it significantly improves the efficiency, accuracy, compliance, and transparency of enterprise procurement in reviewing supplier bids, effectively reducing the risks of bid rigging, collusion, and qualification fraud, and promoting the intelligent and standardized development of the bidding process. Attached Figure Description

[0016] 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 A flowchart illustrating the multimodal intelligent review method for supply chain procurement bidding projects provided in this application embodiment; Figure 2 A schematic diagram of the structure of the multimodal intelligent auditing device for supply chain procurement bidding projects provided in this application embodiment; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0017] Figure label: 110. Rule system definition module; 120. Information processing module; 130. Parsing module; 140. Judgment module; 150. Result output module; 810, Processor; 820, Communication interface; 830, Memory; 840, Communication bus. Detailed Implementation

[0018] To more clearly illustrate the overall concept of this application, a detailed explanation is provided below with reference to the accompanying drawings.

[0019] Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application may also be implemented in other ways different from those described herein. Therefore, the scope of protection of this application is not limited to the specific embodiments disclosed below. It should be noted that, unless otherwise specified, the embodiments of this application and the features thereof can be combined with each other.

[0020] In this application, unless otherwise expressly specified and limited, the "above" or "below" of the second feature can mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediate medium. In the description of this specification, references to terms such as "an embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in one or more embodiments or examples.

[0021] like Figure 1 As shown, the first aspect of this application provides a multimodal intelligent auditing method for supply chain procurement bidding projects, including: Step 100: Define the review rule system, which includes rule content, review points, and review logic.

[0022] Step 200: During the bidding stage, receive the bid documents and quotation information uploaded by suppliers. After the bid opening, input the bidding documents, the bid documents of shortlisted suppliers and the corresponding review rules into the multimodal large model server.

[0023] Step 300: Use a multimodal large model combined with OCR technology to parse unstructured data from the tender documents and bid documents, extract review requirements from the tender documents according to the review rules, and match the response content in the bid documents through intent recognition.

[0024] Step 400: Push the extracted content to the corresponding business review AI agent for evaluation based on review points and review logic.

[0025] Step 500: Summarize the review results and generate an intelligent review report containing the number of compliant items, the number of non-compliant items, the conclusions of each review point, and the basis for the judgment.

[0026] In step 100, a systematic modeling of bidding review standards is achieved by pre-defining a structured review rule system that includes rule content, review points, and review logic. This rule system is configured based on laws and regulations, industry standards, and corporate internal control requirements. It supports flexible use of templates or custom additions according to project type, ensuring that the review basis is consistent, traceable, and reusable. This avoids the subjectivity and inconsistency in judgment caused by differences in expert understanding in traditional manual reviews, and improves the standardization and compliance level of review standards.

[0027] In step 200, a secure interface is opened during the bidding stage to receive encrypted bid documents and pricing information uploaded by suppliers, ensuring data integrity and confidentiality. After the bid opening, the bidding documents, shortlisted suppliers' bid documents, and the configured review rule system are automatically packaged and pushed to the multimodal large model server. This realizes the automated triggering of the review process and the efficient integration of multi-source data, reduces manual intervention and operation delays, improves system response speed and process continuity, and lays a data foundation for subsequent intelligent analysis.

[0028] In step 300, a multimodal large model combined with OCR technology is used to jointly parse unstructured information such as text, images, and tables in bidding and tender documents. This can accurately identify key content in scanned documents and mixed text and image documents, while preserving contextual semantics and layout information. Based on this, the review requirements corresponding to each review point are automatically extracted from the bidding documents, and the response paragraphs or data items are accurately located in the tender documents through a semantically driven intent recognition algorithm. This achieves cross-document semantic alignment and substantive response judgment, overcoming the technical defects of traditional keyword matching methods that are prone to misjudgment and omission, and significantly improving the accuracy and intelligence level of information extraction.

[0029] In step 400, the structured data pairs of "Bidding Requirements - Bid Responses" generated in step 300 are distributed to the corresponding business review agents (such as prequalification agents, technical compliance agents, business compliance agents, and price reasonableness agents) according to their categories. Each agent performs automated evaluation based on preset review points and formal evaluation logic (such as logical expressions, threshold judgments, database comparisons, etc.) to efficiently verify the authenticity, completeness, consistency, and compliance of the response content. This mechanism realizes professional division of labor in the review, enhances the logical rigor and domain adaptability of the evaluation process, significantly reduces the risk of human error, and improves the objectivity and reliability of the review results.

[0030] In step 500, the evaluation results of all business review agents are summarized to generate a structured intelligent review report. The report clearly lists the conclusions, number of compliant items, number of non-compliant items, judgment basis, cited original texts, and evidence chain index for each review point. Some embodiments can also generate a horizontal comparison matrix between suppliers to help experts quickly grasp the differences. This report not only provides comprehensive, transparent, and traceable decision support for bid evaluation experts, significantly shortening bid preparation time, but also enhances the auditability and credibility of the entire review process. Overall, it achieves high efficiency, intelligence, and standardization in enterprise procurement bidding review, effectively preventing compliance risks such as bid rigging and qualification fraud, and promoting the intelligent transformation and upgrading of bidding management.

[0031] The multimodal intelligent review method for supply chain procurement bidding projects provided in the first aspect of this application achieves unified modeling and flexible configuration of bidding review standards by systematically constructing a structured review rule system. This ensures that the review basis is traceable and reusable, avoiding inconsistencies in judgment caused by misunderstandings in traditional manual review. During the bidding stage, the method receives bid documents and price information uploaded by suppliers and automatically inputs the bidding documents, shortlisted suppliers' bid documents, and corresponding review rule system into the multimodal large model server after the bid opening. This achieves automated triggering and data integration of the review process, reducing human intervention and operational delays. By utilizing the multimodal large model combined with OCR technology for deep analysis of unstructured documents, the method can accurately identify text, tables, and graphic layout information in scanned documents and image-type PDFs. It also extracts specific review requirements from the bidding documents through semantic understanding and then uses intent recognition technology to locate matching requirements in the bid documents. The response content achieves semantic alignment and substantive response judgment across documents, overcoming the limitations of traditional keyword matching, which is prone to misjudgment and omission. Subsequently, the parsed structured data of "bidding requirements - bid response" is pushed to the corresponding business review agents (such as qualification review, technical compliance, business compliance, and price reasonableness). Each agent performs automated evaluation based on preset review points and formal logic rules, improving the objectivity, consistency, and logical rigor of the review. Finally, the evaluation results of each agent are summarized to generate an intelligent review report containing the number of compliant items, the number of non-compliant items, the conclusions of each review point, and detailed judgment basis. This not only significantly shortens the preparation time for expert evaluation but also provides a clear and traceable chain of evidence for subsequent manual review. Overall, it significantly improves the efficiency, accuracy, compliance, and transparency of enterprise procurement in reviewing supplier bids, effectively reducing the risks of bid rigging, collusion, and qualification fraud, and promoting the intelligent and standardized development of the bidding process.

[0032] In some embodiments of this application, an examination rule system is defined, which includes rule content, examination points, and examination logic, specifically as follows: During the project initiation phase, after the bidding party creates the procurement project and uploads the bidding documents, it configures a review rule system applicable to the project based on industry standards, laws and regulations, and corporate internal control requirements. The review rule system includes multiple structured review rules. Each review rule contains a rule number, rule name, scope of application, rule content text, review points to be checked, corresponding judgment logic expression, and an identifier indicating whether it is a veto item.

[0033] During the project initiation phase of a corporate procurement project, once the bidding party completes the creation of the procurement project and uploads the formal bidding documents in the electronic procurement system, the system automatically guides the user to the "Review Rule Configuration" stage. In this stage, the bidding party or project management personnel, based on the project's procurement type (e.g., goods, services, engineering), budget size, industry attributes, and organizational management requirements, and in conjunction with relevant national laws and regulations (e.g., the "Tendering and Bidding Law" and the "Government Procurement Law"), industry technical standards, and the company's internal compliance and risk control policies, select a suitable rule set from a pre-set review rule template library, or manually add and modify review rules using a visual editing interface, thereby constructing a structured review rule system highly compatible with the current project.

[0034] This review rule system is organized using a standardized data model. Each review rule is defined by structured fields, including: rule number (for unique identification and traceability), rule name (a concise description of the rule's subject), scope of application (clearly defining which procurement categories or project types the rule applies to), rule content text (original clause description, usually derived from tender documents or policy documents), review points to be verified (breaking down the rule into specific executable inspection items, such as "the supplier must possess a valid safety production license"), corresponding judgment logic expression (converting the review points into machine-executable judgment logic, such as "certificate type = safety production license AND validity period ≥ bid opening date"), and an identifier indicating whether it is a veto item (marking whether failure to pass this review point directly invalidates the bid). The judgment logic expression supports simple conditional judgments, numerical comparisons, and regular expression matching, and can also integrate external database query interfaces (such as enterprise credit information disclosure systems and qualification certificate verification platforms) to achieve dynamic verification.

[0035] The above rule configuration process supports multi-person collaboration, version management, and approval workflows, ensuring the accuracy and compliance of rule settings. After configuration, the system can perform syntax verification and conflict detection on the rule system (such as multiple rules setting contradictory logic for the same review point), and use it as a "knowledge benchmark" for automated evaluation of multimodal large models in subsequent review processes.

[0036] By transforming ambiguous textual clauses into structured, actionable review points and logical expressions, the problem of inconsistent rule interpretation in traditional manual review is resolved, ensuring that all bidders are evaluated under a unified standard and enhancing the fairness and transparency of the review process. The structured rule system provides clear evaluation criteria for multimodal large-scale models and business review intelligent agents, enabling machines to automatically identify, extract, and verify bid responses, serving as a prerequisite and core support for achieving end-to-end intelligent review. Through a template-based configuration mechanism, existing rule sets can be quickly reused for similar projects, reducing repetitive work; at the same time, flexible customization is supported to meet the personalized review needs of different projects, balancing standardization and flexibility. Clearly marking "veto items" helps the system automatically identify major violations, promptly intercepting unqualified bids and reducing legal and audit risks; the complete rule traceability mechanism also facilitates subsequent auditing and liability determination.

[0037] In some embodiments of this application, during the bidding stage, the tender documents and price information uploaded by suppliers are received. After the bid opening, the tender documents, the tender documents of shortlisted suppliers, and the corresponding review rule system are input into the multimodal large model server, specifically as follows: Before the bid submission deadline, the system will open an interface for suppliers to upload encrypted bid documents and price quotes. During the bid opening process, the system decrypts and verifies the authenticity and format compliance of all bid documents. For bid documents that pass the preliminary verification, they are packaged with the corresponding bidding documents and the configured review rule system. Using automated scripts or triggering mechanisms, the packaged dataset is pushed to the multimodal large model server after the bidding is completed; Each bid document is analyzed based on a built-in algorithm, and pre-defined information is extracted and compared with the requirements in the bidding documents.

[0038] During the bidding phase of enterprise procurement projects, the system provides a secure data upload interface to all pre-qualified suppliers through the electronic procurement platform. This interface supports standard file formats (such as PDF, DOCX, XLSX, etc.) and employs high-strength encrypted transmission protocols (such as HTTPS / TLS) and file-level encryption technology (such as AES-256) to ensure the confidentiality and integrity of bid documents and pricing information during transmission and storage, preventing information leakage or tampering. After suppliers complete file uploads before the bid deadline, the system automatically records the upload timestamp and generates a digital fingerprint (such as a SHA-256 hash value) as the basis for subsequent verification of document consistency.

[0039] The bid opening process is automatically triggered at a preset system time or manually initiated by authorized personnel. During the bid opening, the system first decrypts all bid documents and performs preliminary verification based on preset rules: this includes file integrity verification (comparing hash values), format compliance checks (whether it is an allowed file type, whether there is any corruption), digital signature verification (confirming the bidder's identity is genuine and valid), and structured extraction and logical verification of pricing information (such as whether the total price matches the itemized sums). For bid documents that fail verification, the system automatically marks them as abnormal and notifies relevant personnel for processing; only "qualified" bid documents that pass the preliminary verification are included in the subsequent review process.

[0040] Subsequently, the system automatically packages each shortlisted supplier's bid document, its corresponding original tender document, and the structured review rules system configured during the project initiation phase into a single "review task data package." Each data package contains complete contextual information, ensuring that the multimodal large model can accurately understand the review background and evaluation criteria during analysis.

[0041] Following this, the system pushes the data packet to the multimodal large model server immediately after the bid opening, via automated scripts or event-driven mechanisms (such as message queue triggers or API calls). This process requires no manual intervention, achieving seamless integration and efficient startup of the review process. Upon receiving the data packet, the multimodal large model server immediately invokes its built-in document parsing algorithm (integrating OCR, natural language processing, and image recognition technologies) to conduct in-depth analysis of the bidding documents and tender documents, automatically extracting key information (such as technical parameters, qualification certificates, service commitments, and price details), and initially establishing a mapping relationship between "bidding requirements and tender responses," laying the data foundation for subsequent accurate review.

[0042] By employing encrypted upload and digital signature mechanisms, the confidentiality of the bidding process and the authenticity of bidders' identities are ensured, preventing risks such as leaks and impersonation, and enhancing the credibility of the bidding process. From bid receipt and opening verification to task assignment, the entire process is automated, reducing human intervention and operational delays, significantly improving the efficiency of review initiation, and ensuring timely commencement of bid evaluation. The system integrates bidding documents, bid documents, and review rules into a unified package, avoiding information gaps or mismatches, ensuring multimodal model analysis within a unified context, and improving the accuracy of intelligent review. Based on an interface-based and script-driven design, the system can simultaneously handle the upload and parsing of a large number of bid documents, making it suitable for large projects or multi-supplier competitive scenarios, and possessing excellent scalability and stability.

[0043] In some embodiments of this application, a multimodal large model combined with OCR technology is used to parse unstructured data from tender documents and bid documents, and review requirements are extracted from the tender documents according to review rules. Intent recognition is then used to match the response content in the bid documents. Specifically: The uploaded tender documents and bid documents are formatted. For scanned documents, image-type PDFs, or documents containing charts, OCR technology is used to convert them into editable text streams while preserving the original text and image layout information. The multimodal large model is used to jointly analyze the semantics of the converted text and images to identify the chapter structure, table content, technical parameters, qualification requirements, service commitments and quotation details in the document; Based on the configured review rules system, the review requirements related to each review point are extracted from the tender documents to form a structured review requirements list; In the tender documents, an intent recognition algorithm based on semantic understanding is used to locate the response paragraphs or data items that correspond to the above review requirements, thereby achieving content alignment across documents; During the intent recognition process, a contextual attention mechanism is used to determine whether the bid content is substantially responded to, partially responded to, or not responded to, and the matching confidence level is marked. Output structured data pairs, including a mapping table of "tender requirements - bid response", key field extraction results, and original document location coordinates.

[0044] First, the system intelligently recognizes the format of uploaded tender documents and bid documents to determine their document type (e.g., plain text PDF, image PDF, scanned document, Word document, Excel spreadsheet, etc.). For non-editable documents containing images or scanned content, the system automatically uses high-precision OCR (Optical Character Recognition) technology to convert the text information in the image into an editable and searchable text stream. During this process, the system not only extracts the text content but also retains the original document's layout information (e.g., paragraph positions, table borders, heading levels, page number coordinates, etc.), constructing a dual data structure of "text + layout." This ensures that subsequent parsing accurately restores the document's semantic structure and avoids misinterpretation of information due to format loss.

[0045] Subsequently, a multimodal large-scale model is used to jointly analyze the semantics of the transformed text and images. This large-scale model has the ability to process multimodal information such as text, images, and tables simultaneously, and can deeply understand the document content. For example, it can identify the chapter structure of the document (such as "Chapter 3 Technical Requirements" and "Section 5 Business Terms"), automatically segment functional areas; perform structured parsing of complex tables, extract row and column data and understand their semantics (such as equipment model, quantity, and unit price in "Itemized Quotation"); identify key information fields such as technical parameters (such as "maximum output power ≥ 500kW"), qualification requirements (such as "must have ISO9001 quality management system certification"), service commitments (such as "warranty period of not less than 3 years"), and quotation details, and store them in a structured manner.

[0046] Building upon this foundation, the system, in conjunction with its pre-configured structured review rules, automatically extracts review requirements directly relevant to each review point from the tender documents. For example, regarding the rule "Suppliers must provide performance data for similar projects in the past three years," the system locates the "Performance Requirements" section in the tender documents, extracts the specific clause "Completed at least two similar projects with a contract value of no less than 5 million yuan in the past three years," and generates a corresponding "Review Requirements List." This list is organized in a structured format, including fields such as review point number, requirement description, data type, and source location, serving as the benchmark for subsequent matching.

[0047] Next, in the tender documents, the system uses a deep learning-based intent recognition algorithm to find responses that correspond to the aforementioned review requirements. This algorithm not only relies on keyword matching but also analyzes the contextual semantics using semantic understanding models (such as BERT and RoBERTa) to determine whether a passage expresses the intent of "we have completed similar projects" or "we possess the relevant qualifications." For example, in the tender documents, "Our company undertook the XX power plant project in 2022, with a contract amount of 6.8 million yuan" is identified as a response to the requirement of "similar project performance." The system also incorporates a contextual attention mechanism to analyze the logical relationships between the preceding and following text, determine whether the response is complete, truthful, and unambiguous, and assess the matching confidence level (e.g., high, medium, low) for subsequent decision-making regarding whether manual review is necessary.

[0048] Ultimately, the system outputs structured data pairs, mainly including: a mapping table of "tender requirements - bid response" (clearly showing the content and location of the bid response corresponding to each tender requirement), key field extraction results (such as extracted certificate numbers, technical parameter values, quotation amounts, etc.), and original document location coordinates (such as "page 45 of the tender document, row 2 of Table 3"), forming a machine-readable and human-readable intermediate dataset for review.

[0049] By fusing OCR with a multimodal large-scale model, the system achieves high-precision digitization and semantic understanding of complex documents such as scanned documents and image-based PDFs, solving the technical challenge of traditional systems being unable to handle paper tender documents or mixed text and image documents. Based on intent recognition and contextual attention mechanisms, the system can accurately identify substantive responses to bidding requirements in tender documents, avoiding "false responses" or "missed responses" caused by keyword matching, and significantly improving review accuracy. It unifies and structures key information scattered across text, tables, and images, providing high-quality data input for subsequent automated evaluation and supporting precise logical judgment.

[0050] In some embodiments of this application, the review requirements list includes eligibility criteria, technical specifications, performance capabilities, response time, and pricing rules.

[0051] Qualification Requirements: The system extracts relevant clauses from sections such as "Bidder Qualification Requirements" and "Qualification Requirements" in the bidding documents to form qualification review items. Examples include "Suppliers must possess a valid business license," "Project manager must hold a registered construction engineer certificate," and "No major violations in the past three years." A multimodal large-scale model combined with OCR recognizes qualification documents in the bidding documents (such as scanned copies of business licenses and certificates) and determines whether they meet the requirements through intent recognition. It can also connect to external databases (such as the National Enterprise Credit Information Publicity System) for authenticity verification.

[0052] Technical Specifications: The system identifies technical parameters, performance requirements, and standards / specifications related to products or services in tender documents, such as "server CPU frequency ≥ 2.5GHz," "software system must support national cryptographic algorithms," and "equipment protection level IP68." It extracts key values ​​and comparison relationships through semantic parsing and locates response parameters in the technical solution section of the tender documents, determining whether they meet logical conditions such as "≥," "≤," and "=", and supports automatic identification and labeling of technical deviations.

[0053] Performance Capability: The system extracts requirements for suppliers' historical performance, financial status, staffing, and after-sales service system, such as "the total contract amount for similar projects in the past three years shall not be less than 10 million yuan" and "after-sales service outlets must be established in the project location." By analyzing the performance tables, financial statements, personnel resumes, and other information in the tender documents, and combining this with the context to understand the completeness and credibility of their performance commitments, the system assesses the supplier's actual execution capability.

[0054] Response Time: The system identifies time-related clauses, including delivery period, construction period, response time, and warranty period, such as "delivery to be completed within 60 days of contract signing" and "fault response time not exceeding 2 hours." Specific time nodes or time ranges are extracted through semantic understanding, and the corresponding response content is searched in the "Service Commitment" or "Implementation Plan" of the tender documents to determine whether it meets or exceeds the tender requirements.

[0055] Price Composition Rules: The system extracts price review rules from the bidding documents' pricing requirements, such as "the price must not exceed the maximum price limit," "the total of itemized prices must equal the total price," and "a detailed cost composition analysis must be provided." By parsing the price list and price descriptions in the bid documents, the system automatically verifies the compliance, consistency, and reasonableness of the prices, identifying potential risks such as abnormally low prices or unreasonable price breakdowns.

[0056] By categorizing review requirements into five core dimensions—qualification, technology, performance, timeliness, and price—the system ensures comprehensive coverage of all review elements, encompassing key risk points in enterprise procurement and bidding, and enhancing the systematic and complete nature of the review process. This checklist provides clear review inputs for pre-qualification agents, technical compliance agents, business compliance agents, and price reasonableness agents, enabling "categorized distribution and professional evaluation" and improving the accuracy and logic of automated reviews. Transforming tender terms described in natural language into structured, computable review items allows large-scale models and rule engines to automatically identify, extract, and judge them, serving as a crucial bridge to achieving intelligent review throughout the entire process.

[0057] In some embodiments of this application, the extracted content is pushed to the corresponding business review intelligent agent, and judged based on review points and review logic, specifically as follows: The structured data pairs of "Bidding Requirements - Bid Responses" are distributed to the corresponding business review agents according to the categories of the review rules. The business review agents include prequalification agents, business compliance agents, technical compliance agents, and price reasonableness agents. Based on the intelligent agents for each business review, the preset review points, and the formalized evaluation logic, the content of the bid response is logically judged. Verify the authenticity, completeness, consistency, and compliance of the response content; Output the evaluation results for each review point, including three statuses: "passed", "failed", and "requires manual review".

[0058] After parsing the unstructured data of the bidding documents and tender documents and generating structured data pairs of "bidding requirements - tender responses", the system automatically distributes the data pairs to the corresponding business review agents according to the category attributes of each review rule, realizing professional and modular automated evaluation. These agents are lightweight AI proxies built on domain-specific knowledge and rule engines, each performing its own function and working together to complete a comprehensive review.

[0059] The specific distribution and evaluation process is as follows: Distribution Mechanism: The system routes the "Bidding Requirements - Bid Response" data pairs to the corresponding agents based on the category tags of the review rules (such as "Qualification," "Technical," "Business," and "Price"). For example: Pre-qualification intelligent agent: Accepts review items involving business licenses, qualification certificates, personnel qualifications, and no adverse records; Technology compliance agent: processes technical response content such as technical parameters, performance indicators, standard compliance, and solution feasibility; Business Compliance Intelligent Agent: Responsible for reviewing business terms such as performance capability, delivery time, after-sales service commitment, and response to contract terms; Quotation Reasonableness Intelligent Agent: Specialized in handling financial-related reviews such as quotation amount, price composition, price limit compliance, and consistency of item totals.

[0060] Automated evaluation: After receiving data, each business review agent performs logical judgments based on preset review points and formally expressed evaluation logic. The evaluation logic exists in the form of executable expressions or rule sets, for example: Qualification category: "Certificate validity period ≥ Bid opening date AND Certificate status = Valid"; Technical category: "The CPU frequency provided by the bidder must be ≥ 2.5GHz as required by the tender." Commercial category: "Committed delivery period ≤ 60 days as required by the tender"; Price category: "Total bid price ≤ maximum price AND total of items = total price".

[0061] The agent invokes the built-in rule engine or script parser to execute this logic and performs cross-validation in conjunction with external data sources (such as enterprise credit platforms and product databases).

[0062] Multi-dimensional verification: Based on logical judgment, the intelligent agent further verifies the authenticity, completeness, consistency, and compliance of the response content. Authenticity: The certificate image recognized by OCR is compared with the official database, or the digital signature is verified; Completeness: Check for any missing key response items (such as failure to provide performance reports for the past three years); Consistency: Verify that the information within the tender documents is consistent (e.g., whether the equipment models described in the technical solution are consistent with the quotation list). Compliance: Determine whether there is a violation of laws, regulations, or corporate internal control policies (such as responses to exclusivity clauses).

[0063] Output evaluation results: Each agent outputs a clear evaluation status for the review point it is responsible for, including: "Passed": The response content fully meets the requirements; "Not Passed": There are substantial deviations or non-compliance items, which may trigger the veto mechanism; "Requires manual review": Situations with low confidence, ambiguous semantics, or involving complex judgments (such as the merits of technical solutions) will be automatically marked by the system and pushed to the manual review stage.

[0064] The assessment results are accompanied by detailed explanations, including the basis for the judgment, the cited clauses, evidence snippets, and the location of the original documents, forming a traceable review log.

[0065] By constructing business review agents dedicated to multiple domains, complex review tasks are broken down into specialized modules, improving the accuracy and professional depth of judgment and avoiding the problem of insufficient judgment capability of "general models" in specific domains. Each agent processes different types of review tasks in parallel, significantly shortening the overall review time and reducing reliance on human experts, making it particularly suitable for large-scale, multi-supplier bidding. All suppliers are evaluated under the same set of formalized logical rules, avoiding subjective bias caused by human factors and ensuring a fair, impartial, and repeatable review process.

[0066] In some embodiments of this application, it also includes: A horizontal comparative analysis of the intelligent review results of multiple shortlisted suppliers is conducted, and a structured comparison matrix is ​​generated using a multimodal large model to compare suppliers in terms of technical parameter response, service commitment level, price composition details, and qualification conditions. The comparison matrix is ​​presented in the form of a visual chart, marking the strengths, weaknesses and differences of each supplier, and automatically generating a summary of evaluation suggestions.

[0067] After completing the intelligent review of each shortlisted supplier and generating an independent review report, the system further activates the horizontal comparative analysis module. Leveraging the semantic understanding and data integration capabilities of the multimodal large-scale model, it conducts a comprehensive cross-subject, multi-dimensional comparison of the review results of multiple suppliers. This process goes beyond simple data listing; instead, through deep semantic analysis and structured modeling, it generates a comprehensive, objective, and highly readable structured comparison matrix.

[0068] The specific implementation process is as follows: Data normalization and alignment: The system first standardizes the review results of each supplier to ensure consistency in units, format, and semantics across different sources. For example, "Delivery period: 60 days" and "Construction period: 2 months" are standardized to "days"; "3-year warranty" and "36-month warranty period" are uniformly expressed. Simultaneously, based on the "Tender Requirements - Tender Response" mapping table, the system extracts the responses from each supplier at the same review points, forming a horizontally comparable dataset.

[0069] Construct a structured comparison matrix: Using review dimensions as columns (e.g., technical parameters, service commitments, price structure, qualification requirements) and suppliers as rows, create a two-dimensional comparison table. The system automatically populates the content of each cell, including: Technical parameter response: Compare the equipment performance, technical solutions, and functional implementation provided by each supplier, and indicate whether they are better than, equal to, or lower than the bidding requirements; Service commitment level: Compare soft indicators such as after-sales service response time, warranty period, training plan, and localization support; Price breakdown details: Analyze the quotation structure, compare the total price, itemized unit prices, and cost composition for reasonableness, and identify abnormally low prices or hidden costs; Strengths and weaknesses of qualifications: Summarize the company's qualifications, project performance, staffing, industry certifications and other hard conditions to assess its overall strength.

[0070] Intelligent Strengths and Weaknesses Identification and Labeling: A multimodal large-scale model, combined with pre-set scoring logic and an industry knowledge base, automatically identifies each supplier's strengths (e.g., "lowest price," "technical parameters comprehensively superior to bidding requirements," "possessing national-level qualifications") and weaknesses (e.g., "lacking a key certificate," "delivery time exceeding requirements," "opaque pricing structure"), and highlights them in the comparison matrix using color, icons, or labels (e.g., green "✓" indicates strengths, red "✓" indicates weaknesses). (This indicates a risk).

[0071] Visualized chart presentation: The system transforms the structured comparison matrix into various visual chart formats, including: Radar chart: Displays the distribution of each supplier's overall capabilities across dimensions such as technology, price, service, and qualifications; Bar charts / column charts: visually compare key indicators (such as price and delivery time); Heatmap: The intensity of color reflects the level of response; Difference Annotation View: Directly annotate differences in the table for easy and quick location of discrepancies.

[0072] Automatically Generate Evaluation Recommendation Summary: Based on the comparative analysis results, the system utilizes the natural language generation (NLG) capabilities of the large model to automatically generate a clearly structured and linguistically standardized evaluation recommendation summary. The summary includes: overall comparative conclusions, a comprehensive analysis of the advantages and disadvantages of each supplier, explanations of key differences, potential risk warnings (e.g., a supplier's unusually low price warrants attention regarding performance risks), and preliminary recommendations (e.g., "It is recommended to prioritize suppliers A and B"), providing efficient and objective decision support for evaluation experts.

[0073] By automatically generating structured comparison matrices and visual charts, the system helps evaluation experts quickly grasp the differences between suppliers, avoiding the need for manual comparison of numerous documents and significantly shortening evaluation preparation time. Based on unified standards and data-driven approaches, the system conducts horizontal comparisons, reducing the influence of human bias and subjective judgment, ensuring a fairer, more impartial, and quantifiable evaluation process. Through multi-dimensional and granular comparative analysis, it clearly presents the strengths and weaknesses of each supplier in terms of technology, service, price, and qualifications, helping experts identify bids that are "superficially compliant but actually inadequate" or "low-priced but high-risk."

[0074] like Figure 2 As shown, a second aspect of this application provides a multimodal intelligent auditing device for supply chain procurement bidding projects, comprising: The rule system definition module 110 is suitable for defining a review rule system that includes rule content, review points, and review logic; The information processing module 120 is suitable for receiving the tender documents and quotation information uploaded by suppliers during the bidding stage, and after the bid opening, inputting the tender documents, the tender documents of shortlisted suppliers and the corresponding review rule system into the multimodal large model server. The parsing module 130 is suitable for parsing unstructured data from tender documents and bid documents using a multimodal large model combined with OCR technology, extracting review requirements from the tender documents according to review rules, and matching the response content in the bid documents through intent recognition. The judgment module 140 is suitable for pushing the extracted content to the corresponding business review intelligent agent and judging it based on the review points and review logic; The results output module 150 is suitable for summarizing review results and generating an intelligent review report that includes the number of compliant items, the number of non-compliant items, the conclusions of each review point, and the basis for the judgment.

[0075] The multimodal intelligent auditing device for supply chain procurement bidding projects provided in the second aspect of this application can realize the multimodal intelligent auditing method for supply chain procurement bidding projects in any of the embodiments of the first aspect above. Therefore, it can achieve any of the technical effects in the multimodal intelligent auditing method for supply chain procurement bidding projects above, which will not be elaborated here.

[0076] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the multimodal intelligent auditing method for supply chain procurement bidding projects described in any of the first aspects above.

[0077] Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3 As shown, the electronic device may include: a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute the multimodal intelligent auditing method for supply chain procurement bidding projects in any embodiment of the first aspect described above. This method includes: Step 100: Define the review rule system, which includes rule content, review points, and review logic.

[0078] Step 200: During the bidding stage, receive the bid documents and quotation information uploaded by suppliers. After the bid opening, input the bidding documents, the bid documents of shortlisted suppliers and the corresponding review rules into the multimodal large model server.

[0079] Step 300: Use a multimodal large model combined with OCR technology to parse unstructured data from the tender documents and bid documents, extract review requirements from the tender documents according to the review rules, and match the response content in the bid documents through intent recognition.

[0080] Step 400: Push the extracted content to the corresponding business review AI agent for evaluation based on review points and review logic.

[0081] Step 500: Summarize the review results and generate an intelligent review report containing the number of compliant items, the number of non-compliant items, the conclusions of each review point, and the basis for the judgment.

[0082] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0083] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is able to execute the multimodal intelligent auditing method for supply chain procurement bidding projects provided by the above methods, the method comprising: Step 100: Define the review rule system, which includes rule content, review points, and review logic.

[0084] Step 200: During the bidding stage, receive the bid documents and quotation information uploaded by suppliers. After the bid opening, input the bidding documents, the bid documents of shortlisted suppliers and the corresponding review rules into the multimodal large model server.

[0085] Step 300: Use a multimodal large model combined with OCR technology to parse unstructured data from the tender documents and bid documents, extract review requirements from the tender documents according to the review rules, and match the response content in the bid documents through intent recognition.

[0086] Step 400: Push the extracted content to the corresponding business review AI agent for evaluation based on review points and review logic.

[0087] Step 500: Summarize the review results and generate an intelligent review report containing the number of compliant items, the number of non-compliant items, the conclusions of each review point, and the basis for the judgment.

[0088] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multimodal intelligent auditing method for supply chain procurement bidding projects provided by the methods described above, the method comprising: Step 100: Define the review rule system, which includes rule content, review points, and review logic.

[0089] Step 200: During the bidding stage, receive the bid documents and quotation information uploaded by suppliers. After the bid opening, input the bidding documents, the bid documents of shortlisted suppliers and the corresponding review rules into the multimodal large model server.

[0090] Step 300: Use a multimodal large model combined with OCR technology to parse unstructured data from the tender documents and bid documents, extract review requirements from the tender documents according to the review rules, and match the response content in the bid documents through intent recognition.

[0091] Step 400: Push the extracted content to the corresponding business review AI agent for evaluation based on review points and review logic.

[0092] Step 500: Summarize the review results and generate an intelligent review report containing the number of compliant items, the number of non-compliant items, the conclusions of each review point, and the basis for the judgment.

[0093] Finally, the present invention also provides a non-volatile computer storage medium storing computer-executable instructions thereon, which, when executed by a processor, implement the multimodal intelligent auditing method for supply chain procurement bidding projects provided by the above methods, the method comprising: Step 100: Define the review rule system, which includes rule content, review points, and review logic.

[0094] Step 200: During the bidding stage, receive the bid documents and quotation information uploaded by suppliers. After the bid opening, input the bidding documents, the bid documents of shortlisted suppliers and the corresponding review rules into the multimodal large model server.

[0095] Step 300: Use a multimodal large model combined with OCR technology to parse unstructured data from the tender documents and bid documents, extract review requirements from the tender documents according to the review rules, and match the response content in the bid documents through intent recognition.

[0096] Step 400: Push the extracted content to the corresponding business review AI agent for evaluation based on review points and review logic.

[0097] Step 500: Summarize the review results and generate an intelligent review report containing the number of compliant items, the number of non-compliant items, the conclusions of each review point, and the basis for the judgment.

[0098] For any parts not mentioned in this application, existing technologies may be used or referenced.

[0099] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0100] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A multimodal intelligent review method for supply chain procurement bidding projects, characterized in that, include: Define a review rule system that includes rule content, review points, and review logic; During the bidding stage, the system receives the bid documents and price information uploaded by suppliers. After the bid opening, the system inputs the bidding documents, the bid documents of shortlisted suppliers, and the corresponding review rules into the multimodal large model server. The unstructured data of the tender documents and bid documents are parsed using a multimodal large model combined with OCR technology. The review requirements are extracted from the tender documents according to the review rules, and the response content in the bid documents is matched by intent recognition. The extracted content is pushed to the corresponding business review AI agent for evaluation based on review points and review logic; The review results are summarized to generate an intelligent review report that includes the number of compliant items, the number of non-compliant items, the conclusions of each review point, and the basis for the judgment.

2. The multimodal intelligent auditing method for supply chain procurement bidding projects according to claim 1, characterized in that, The definition includes a review rule system comprising rule content, review points, and review logic, specifically: During the project initiation phase, after the bidding party creates the procurement project and uploads the bidding documents, it configures a review rule system applicable to the project based on industry standards, laws and regulations, and corporate internal control requirements. The review rule system includes multiple structured review rules. Each review rule includes a rule number, rule name, scope of application, rule content text, review points to be checked, corresponding judgment logic expression, and an identifier indicating whether it is a veto item.

3. The multimodal intelligent auditing method for supply chain procurement bidding projects according to claim 1, characterized in that, During the bidding stage, the system receives bid documents and price information uploaded by suppliers. After the bid opening, the system inputs the bidding documents, shortlisted suppliers' bid documents, and corresponding review rules into the multimodal large model server. Specifically: Before the bid submission deadline, the system will open an interface for suppliers to upload encrypted bid documents and price quotes. During the bid opening process, the system decrypts and verifies the authenticity and format compliance of all bid documents. For bid documents that pass the preliminary verification, they are packaged with the corresponding bidding documents and the configured review rule system. Using automated scripts or triggering mechanisms, the packaged dataset is pushed to the multimodal large model server after the bidding is completed; Each bid document is analyzed based on a built-in algorithm, and pre-defined information is extracted and compared with the requirements in the bidding documents.

4. The multimodal intelligent auditing method for supply chain procurement bidding projects according to claim 1, characterized in that, The process involves using a multimodal large model combined with OCR technology to parse unstructured data from the tender documents and bid documents, extracting review requirements from the tender documents according to review rules, and matching the response content in the bid documents through intent recognition. Specifically: The uploaded tender documents and bid documents are formatted. For scanned documents, image-type PDFs, or documents containing charts, OCR technology is used to convert them into editable text streams while preserving the original text and image layout information. The multimodal large model is used to jointly analyze the semantics of the converted text and images to identify the chapter structure, table content, technical parameters, qualification requirements, service commitments and quotation details in the document; Based on the configured review rules system, the review requirements related to each review point are extracted from the tender documents to form a structured review requirements list; In the tender documents, an intent recognition algorithm based on semantic understanding is used to locate the response paragraphs or data items that correspond to the above review requirements, thereby achieving content alignment across documents; During the intent recognition process, a contextual attention mechanism is used to determine whether the bid content is substantially responded to, partially responded to, or not responded to, and the matching confidence level is marked. Output structured data pairs, including a mapping table of "tender requirements - bid response", key field extraction results, and original document location coordinates.

5. The multimodal intelligent auditing method for supply chain procurement bidding projects according to claim 4, characterized in that, The review requirements list includes eligibility criteria, technical specifications, performance capabilities, response time, and pricing rules.

6. The multimodal intelligent auditing method for supply chain procurement bidding projects according to claim 4, characterized in that, The extracted content is pushed to the corresponding business review AI, and evaluated based on review points and review logic, specifically as follows: The structured data pairs of "Bidding Requirements - Bid Responses" are distributed to the corresponding business review agents according to the categories of review rules. The business review agents include prequalification agents, business compliance agents, technical compliance agents, and price reasonableness agents. Based on the intelligent agents for each business review, the preset review points, and the formalized evaluation logic, the content of the bid response is logically judged. Verify the authenticity, completeness, consistency, and compliance of the response content; Output the evaluation results for the review points, including three statuses: "passed", "failed", and "requires manual review".

7. The multimodal intelligent auditing method for supply chain procurement bidding projects according to any one of claims 1 to 6, characterized in that, Also includes: A horizontal comparative analysis of the intelligent review results of multiple shortlisted suppliers is conducted, and a structured comparison matrix is ​​generated using a multimodal large model to compare suppliers in terms of technical parameter response, service commitment level, price composition details, and qualification conditions. The comparison matrix is ​​presented in the form of a visual chart, marking the strengths, weaknesses and differences of each supplier, and automatically generating a summary of evaluation recommendations.

8. A multimodal intelligent auditing device for supply chain procurement bidding projects, characterized in that, include: The rule system definition module is suitable for defining a review rule system that includes rule content, review points, and review logic; The information processing module is suitable for receiving the tender documents and quotation information uploaded by suppliers during the bidding stage. After the bid opening, it inputs the tender documents, the tender documents of shortlisted suppliers and the corresponding review rules system into the multimodal large model server. The parsing module is suitable for using a multimodal large model combined with OCR technology to parse unstructured data from the tender documents and the bid documents, extracting review requirements from the tender documents according to review rules, and matching the response content in the bid documents through intent recognition; The judgment module is suitable for pushing the extracted content to the corresponding business review intelligent agent and judging it based on the review points and review logic; The results output module is suitable for summarizing review results and generating an intelligent review report that includes the number of compliant items, the number of non-compliant items, the conclusions of each review point, and the basis for the judgment.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the multimodal intelligent auditing method for supply chain procurement bidding projects as described in any one of claims 1 to 7.

10. A non-volatile computer storage medium storing computer-executable instructions thereon, characterized in that, When the computer-executable instructions are executed by the processor, they implement the multimodal intelligent auditing method for supply chain procurement bidding projects as described in any one of claims 1 to 7.

Images