Multi-mode intelligent compliance pre-auditing method and system

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a multi-modal intelligent compliance pre-auditing method and system, and the method comprises the following steps: receiving a qualification package uploaded by a supplier, and automatically completing the decompression operation; accurately marking the decompressed page as five types of a subject license, a financial report, industry qualification, a product catalog and environmental protection and safety evaluation by using a layout perception model, and executing fingerprint hash de-duplication processing on the homologous file to generate a unique document identifier DocID; the method has the beneficial effects that the authenticity verification, financial analysis and directory matching of all qualifications of a single supplier can be completed within about 30 seconds by adopting a one-package type automatic sorting link and parallel API aggregation verification; and compared with the traditional manual item-by-item check which merges the process of several hours, the whole check period is shortened by more than 90%, and the manpower of compliance specialists is greatly released.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to a multi-modal intelligent compliance pre-audit method and system. BACKGROUND

[0002] The existing supplier access process relies on manual checking of a large number of scanned documents and PDF qualification packages, including certificate authenticity comparison, validity period check, business scope matching with bidding subject, and financial stability analysis. Manual methods not only take a long time, have large subjective errors, but also have difficulty in timely discovering hidden risks such as certificate forgery, financial report embellishment, or product catalog inconsistency with bidding needs.

[0003] Although there are certificate recognition tools based on OCR and independent financial indicator calculation software on the market, there is a lack of an intelligent system that can integrate multi-modal analysis, regulation correction, financial semantic reasoning, and catalog matching in the same workflow, resulting in fragmented data and broken review chain, and unable to form an explainable and traceable automated compliance pre-audit mechanism. SUMMARY

[0004] The present application aims to provide a multi-modal intelligent compliance pre-audit method and system to solve the problems raised in the background.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solution: a multi-modal intelligent compliance pre-audit method, comprising the following steps:

[0006] A) receiving a qualification compressed package uploaded by a supplier and automatically completing the decompression operation;

[0007] B) using a layout perception model to accurately label the decompressed pages into five categories: main certificate, financial report, industry qualification, product catalog, and environmental safety evaluation, and performing fingerprint hash deduplication processing on homologous files to generate a unique document identifier DocID;

[0008] C) performing optical character recognition on the DocID labeled as the main certificate, calling at least one national-level regulatory interface to compare the license authenticity, status, and validity period, and generating a four-level risk label according to the preset rules;

[0009] D) using a large language model to extract the business scope, license items, and registered capital key fields from the business license text, matching the business scope vector with the bidding material catalog semantic ontology library, and outputting out-of-scope or qualification gap prompt information;

[0010] E) using a search enhancement-reasoning chain workflow to analyze the financial report for the past three years, calculating the current ratio, asset-liability ratio, shareholder equity return rate, operating cash net, and accounts receivable turnover rate, and generating a financial stability index combined with industry quantiles;

[0011] F) The product catalog is first processed by OCR, and then compared with the bid bill of materials vector after multi-modal embedding, to obtain the specification matching degree and missing items;

[0012] G) Comprehensive subject license risk markers, financial stability index, catalog matching degree, and seal credibility, according to the configurable weight, calculate the comprehensive score of 0-100, output the access, supplement the certificate or reject conclusion based on the score threshold, and generate a list of non-compliance items containing regulations.

[0013] Preferably, the layout perception model in step B uses LayoutLMv3, and the fingerprint hash deduplication uses both fast perceptual hash FP-Hash and SHA-256 algorithms to directly mark and skip the subsequent verification process for the identified duplicate files.

[0014] Preferably, the national regulatory interface called in step C at least covers the national enterprise credit information public system interface, the housing and urban-rural construction department qualification interface, and the national drug administration license interface. The operating abnormity, suspension, and punishment records returned by the interface are normalized in a unified field to facilitate subsequent accurate analysis and risk marking.

[0015] Preferably, the search enhancement-reasoning chain in step E specifically includes:

[0016] a) Provide accurate definition basis for financial indicator calculation by searching IFRS / GAAP financial definition fragments through vector search;

[0017] b) Call a large language model to accurately locate the cell coordinates of the financial table, ensuring the accuracy of financial data extraction;

[0018] c) Automatically calculate five financial indicators with the help of built-in functions in the model, and output chain explanation to facilitate auditors to understand the calculation process and basis.

[0019] Preferably, it also includes the step of optimizing and improving the audit process and results: collect comprehensive logs of "model determination and manual review differences", and through offline fine-tuning of the large language model entity extractor and comprehensive score weight, according to the accumulated data and feedback within the running period, the audit accuracy is gradually optimized and improved with the running period, and the performance of the multi-modal intelligent compliance pre-audit method is continuously improved.

[0020] A multi-modal intelligent compliance pre-audit method and system, comprising:

[0021] File receiving and decompression module: used for receiving the qualification compressed package uploaded by the supplier, and automatically completing the decompression operation to provide basic files for subsequent processing;

[0022] File classification and identification module: adopt layout perception model to label the decompressed pages into five categories: main body license, financial report, industry qualification, product catalog and environmental safety evaluation, and perform fingerprint hashing deduplication on homologous files to generate unique document identification DocID, realizing accurate classification and unique identification of files;

[0023] License verification module: optical character recognition is performed on the DocID labeled as main body license, at least one national regulatory interface is called to compare the license authenticity, status and validity period, and a four-level risk label of red, orange, yellow and green is formed according to the preset rules to complete the preliminary judgment of compliance of the license;

[0024] Business scope matching module: use a large language model to extract the business scope, licensed items and registered capital key fields from the business license text, match the business scope vector with the semantic ontology library of the bidding material catalog, output the out-of-scope or qualification gap prompt, and assist in determining whether the supplier's business qualification meets the requirements;

[0025] Financial analysis module: use search-enhanced-reasoning chain workflow to analyze the financial report of the past three years, calculate the current ratio, asset-liability ratio, shareholder equity return rate, operating cash net and accounts receivable turnover rate, and generate a financial stability index combined with industry quantiles to comprehensively assess the financial condition of the supplier;

[0026] Product catalog comparison module: after the product catalog is processed by OCR and multi-modal embedding, it is compared with the bidding material list vector to obtain the specification matching degree and missing items, ensuring that the products provided by the supplier meet the bidding requirements;

[0027] Comprehensive scoring and decision-making module: combine the main body license risk label, financial stability index, catalog matching degree and chapter print credibility, calculate a comprehensive score of 0-100 according to the configurable weight, output the admission, supplementary certificate or rejection conclusion based on the scoring threshold, and generate a list of non-compliance items containing regulations for the final decision.

[0028] Preferably, the layout perception model in the file classification and identification module is LayoutLMv3, and the fingerprint hashing deduplication simultaneously uses fast perceptual hashing FP-Hash and SHA-256 algorithm to directly mark and skip the subsequent verification process for duplicate files, improving the efficiency of file processing.

[0029] Preferably, the license verification module calls at least one national regulatory interface including the national enterprise credit information public system interface, the housing and urban-rural construction department qualification interface and the national drug administration license interface, and performs unified field normalization processing on the returned business abnormality, revocation and punishment records to accurately analyze and mark the license risk.

[0030] Preferably, the semantic matching of the business scope matching module adopts a double-encoder architecture, embeds the business scope text into a 768-dimensional vector space, and performs cosine matching with pre-constructed tender material ontology library node vectors; when the maximum matching degree is less than 0.35, an out-of-scope warning is triggered, improving the accuracy of business scope matching.

[0031] Preferably, the following optimization module is further included:

[0032] The weight configuration module: the weight configuration file of the comprehensive score adopts YAML format, allowing compliance personnel to adjust the weight of each dimension online; the system refreshes the historical pass rate in real time after adjustment and records the version number for tracing, enhancing the flexibility and manageability of the system;

[0033] The non-conformance list generation module: the generated non-conformance list is automatically written by a large language model according to the comprehensive score results and the regulation knowledge base, the output format includes a three-part structure of "problem description - regulation reference - rectification suggestion", and automatically inserts the original text page screenshot positioning, improving the standardization and readability of the non-conformance list;

[0034] The decision execution and feedback module: the admission, supplement or rejection conclusion is written back to the SRM system through the JSON interface and triggers the BPMN process: the comprehensive score is greater than or equal to 80, automatically admitted, 60-79 generates a supplement work order, and less than 60 is automatically rejected and the supplier archive is sealed, realizing the automatic execution of the decision;

[0035] The model optimization module: the "model determination and manual review difference" is collected, and the entity extractor of the large language model and the comprehensive score weight are fine-tuned offline, so that the audit accuracy is gradually optimized and improved with the running period, and the performance of the system is continuously improved.

[0036] Compared with the prior art, the beneficial effects of the present application are:

[0037] The multi-modal intelligent compliance pre-audit method and system proposed by the present application adopts a "one-bag" automatic sorting link and parallel API aggregation verification, which can complete the authenticity verification, financial analysis and directory matching of all qualifications of a single supplier within about 30 seconds. Compared with the traditional manual item-by-item inspection process of several hours, the overall audit cycle is shortened by more than 90%, greatly releasing the manpower of compliance officers. Through the layout-semantic double model, Ontology vector matching and CLIP multi-modal comparison, the system can not only accurately identify fake business licenses, out-of-scope qualifications, but also find hidden risks such as product directory and tender material mismatch, financial whitewashing, etc. The comprehensive score engine aggregates the four-dimensional indicators of certificates, finance, products and seals, and the average risk omission rate is reduced by about 70%. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The present application is a method flowchart. DETAILED DESCRIPTION

[0039] In order to make the purposes, technical solutions of the present application clear, complete and the advantages more clear and apparent, the embodiments of the present application are further described in detail below with reference to the drawings. It should be understood that the specific embodiments described herein are part of the embodiments of the present application, rather than all the embodiments, and are only used to explain the embodiments of the present application, and do not limit the embodiments of the present application, and all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0040] Embodiment one, the present application provides a technical solution: a multi-modal intelligent compliance pre-audit method, comprising the following steps:

[0041] Step 1: Intelligent sorting and deduplication of qualification files

[0042] 1-A decompression and preliminary classification

[0043] After the system receives the ZIP package, the file daemon process is called, and all PDF / pictures are pulled into the object storage; then the LayoutLMv3 layout analyzer is started, and the page header, chapter, two-dimensional code, red seal contour and other features of each page document are extracted. The LLM classifier based on few-shot Prompt quickly maps the page to five labels of "certificate, finance, industry qualification, product catalog, environmental safety evaluation" and outputs the page-file relationship table.

[0044] 1-B file objectification and fingerprint deduplication

[0045] According to the label, the page is reassembled into a logical document object (DocID), and the fast perceptual hash (FP-Hash) and SHA-256 are calculated for each object. The system queries the fingerprint library: if the DocID already exists and the version number is consistent, directly mark "duplicate=true" and skip the subsequent verification; if the SHA is the same but the timestamp is updated, it is considered as "new version", and the old version is retained for difference comparison.

[0046] 1-C metadata database

[0047] DocID, label, upload time, and supplier ID are written into MongoDB; if the file contains seal or signature images, the field needs_seal_check=1 is added for subsequent seal comparison routing. This table becomes the main index for subsequent steps, and the whole process is connected by DocID string.

[0048] Step 2: Certificate and qualification authenticity and timeliness double check

[0049] 2-AOCR+field alignment

[0050] PaddleOCR is performed on the DocID labeled "license", and the fields of unified social credit code, legal person name, registered capital, business period, etc. are extracted; LLMPrompt will correct the OCR disorder, line breaks, and fill in the missing fields with the word "to be filled in".

[0051] 2-B Multi-source interface aggregation query

[0052] The unified social credit code is used as the primary key, and the system requests the national enterprise credit information public system, the Ministry of Housing and Construction qualification database, and the drug administration bureau license database in parallel; the intermediate layer performs interface return standardization, and the unified fields are status, expiry_date, and penalty_flag. If the interface times out, record status = unknown and write to the retry queue.

[0053] 2-C Rule determination and red, orange, yellow, green marking

[0054] The rule engine reads the returned fields: status!= active or penalty_flag = true, which means red card; valid period ≤ 6 months, which means orange card; data anomaly but can be manually supplemented, which means yellow card, and the rest is green card. The marking result is written back to the DocStatus table for comprehensive scoring.

[0055] 2-D Asynchronous review and write-back

[0056] For licenses marked as red / orange / yellow, the system automatically triggers an SRM work order: rejection or license supplement. If the compliance officer confirms that there is no error, you can check "review passed" on the UI, and the system calls the API to write back status_override = pass to refresh the routing of the next link.

[0057] Step 3: Align the scope of business with the bidding directory semantics

[0058] 3-A Business scope entity extraction

[0059] The system calls the self-developed 13B parameter LLM to perform NER on the business license text, extracts the business_scope array and license_items sub-list, and analyzes the registered capital and legal information, and writes them into the SupplierProfile index.

[0060] 3-B Ontology vector matching

[0061] business_scope is encoded by Sentence-Transformer to generate a 768-dimensional vector. The system reads the "Tender Material Ontology Tree" node vector and performs cosine matching; if the maximum matching degree <0.35, trigger "out-of-scope warning". For the gray area with a matching degree of 0.35-0.6, generate "operable but requires additional license" prompts by LLM.

[0062] 3-C Gap and Out-of-Scope Reporting

[0063] For all gap nodes, the system lists the necessary qualification clauses and penalty clauses for the tender, generating a "gap list"; out-of-scope operations refer to business penalties cases (retrieve regulatory knowledge base) to output risk scores. The results are pushed to the ScopeCheck table.

[0064] Step 4: Intelligent Evaluation of Financial Soundness

[0065] 4-A Financial Statement Analysis

[0066] DocID labeled "Finance" is input into RAG-Agent: vector search IFRS field definition → LLM parses balance sheet, cash flow statement, and profit statement, capturing core subject values. Analyze the link and reference fragments to ensure traceability.

[0067] 4-B Core Indicator Calculation and Industry Calibration

[0068] Tool-Call calculates five major indicators: current ratio, asset-liability ratio, ROE, operating cash net, and accounts receivable turnover; then queries the local industry benchmark table to calculate quantile deviation and generate a "soundness index". LLM outputs a "financial diagnosis report" explaining the abnormal subject sources one by one.

[0069] 4-C Risk Radar and Interpretive Visualization

[0070] Map the five indicators to a radar chart, and SHAP algorithm identifies the top two positive / negative factors. Users can click on any indicator to expand the "chain explanation + improvement suggestions" generated by LLM, forming a transparent financial analysis experience.

[0071] Step 5: Product Catalog Matching and Comprehensive Scoring Output

[0072] 5-A Product Catalog Vectorization

[0073] Perform OCR+CLIP encoding on the line description, specifications, brand, and pictures in the catalog, and merge them into a unified embedding; the tender material list has been pre-encoded and stored. The system calculates the top-k cosine matching and outputs the matching degree matrix and the "missing specifications" list.

[0074] 5-B Comprehensive Scoring Engine

[0075] Engine reads four dimensions: license_pass_rate, financial_index, catalog_match_score, seal_trust_score, calculates a comprehensive score according to YAML weights. trigger configuration allows compliance officers to adjust weights online and preview pass rate changes in real time.

[0076] 5-C Result Generation and Process Triggering

[0077] LLM calls the regulatory knowledge base and scoring results to generate a "Non-Compliance List & Rejection / Supplement Opinion" PDF: Each issue references specific regulatory provisions and screenshot positioning page numbers. The system also writes JSON back to the SRM, triggering BPMN:

[0078] Comprehensive score ≥ 80 automatically admitted;

[0079] 60-79 initiate the supplement process;

[0080] <60 Reject and seal the supplier's information.

[0081] Background tasks write all "model judgment vs manual review" differences to the training set for weekly offline fine-tuning of NER and scoring weights to continuously iterate audit accuracy.

[0082] In example two, on the basis of example one, a multi-modal intelligent compliance pre-audit method system is proposed, which includes:

[0083] 1. Intelligent sorting and deduplication of qualification documents

[0084] This layer is responsible for splitting the ZIP / PDF "qualification package" uploaded by the supplier once into structured document objects. The system uses LayoutLMv3 combined with layout geometric features to first locate the cover, chapter mark, and page number, and then perform "license, financial statements, industry qualifications, product catalog, environmental / safety evaluation" five categories of tags according to the semantics of each page.

[0085] After sorting, trigger quick hash deduplication: Calculate FP-Hash for the original scanned copy and record the version number; If the same file is repeated across packages or the historical version has been stored, the system returns "duplicate file number" to avoid repeated audits. All metadata (file ID, category, upload time, supplier ID) are written to the MongoDB document library to provide a unified reference for subsequent links.

[0086] When parsing encounters photos, seals, or signature-intensive pages, the system will store the scanned copy as an object in the object storage and mark it with a "visual verification required" tag in the metadata, reserving an entry for the subsequent seal consistency detection module to achieve file-image dual-track processing.

[0087] 2. Certificate and qualification authenticity and validity verification layer

[0088] This layer focuses on the subject certificates such as "business license, administrative license, industry qualification certificate", and completes the authenticity verification and validity period check. The process entry pulls the OCR text from the "certificate" file queue of the previous layer, then calls multiple data sources such as the national enterprise credit information public system, the housing and urban-rural development department qualification API, and the drug administration bureau license API for field-level comparison.

[0089] The LLM-Prompt template developed by the system can automatically correct OCR errors and missing fields, ensuring the completeness of the comparison request parameters; the suspension, abnormal operation, and penalty records in the return results are normalized and mapped to "red, orange, yellow, green" four-level signs and update the certificate status table.

[0090] If the API returns "non-real-time searchable" or times out, the system switches to asynchronous polling mode and displays the "pending review" status on the front end; after manual review, the status can be updated through the embedded signing interface, and the credibility factor is written back, forming a "human-machine co-review" closed loop.

[0091] 3. Business scope and bidding catalog semantic alignment layer

[0092] After the subject certificate passes the authenticity check, the system calls the self-developed Chinese large model (13B parameters, industry finance and tax corpus fine-tuning) to perform named entity recognition on the full text of the business license, extracting key fields such as business scope, license items, registered capital, and legal representative.

[0093] For the extracted business scope, the system uses the Dual-Encoder retrieval framework to perform cosine matching between its vector and the "bidding material level tree + industry whitelist Schema", outputting the matching score, gap list, and "out-of-scope warning" flag; for high-risk gaps, corresponding legal provision citations and penalty case links are generated to facilitate secondary verification by compliance personnel.

[0094] The results of this layer are not only written into the supplier profile, but also pushed to the subsequent product catalog matching layer; when the subsequent layers find catalog-material conflicts, they can write back the "second-level verification failed" status, driving the business scope re-audit, and providing chain consistency assurance for the entire process.

[0095] 4. Intelligent evaluation of financial soundness

[0096] The system sends the PDF audit report of the past three years into the RAG-Agent workflow: vector search IF-RS / GAAP definition fragments → LLM parses financial tables → Tool-Call calculates five indicators (current ratio, asset-liability ratio, ROE, operating cash flow net, accounts receivable turnover).

[0097] To avoid external financial model dependence, the application preinstalls industry benchmark tables and quantile statistics in the local knowledge base. The LLM first retrieves "industry benchmark value = X" in the reasoning stage, then compares the real indicators of the enterprise to generate "deviation explanation"; all intermediate links (thinking track, reference benchmark) are recorded in Audit-Trail, which is convenient for financial personnel to review.

[0098] The five indicators are weighted to obtain "robust index 0-100", and the SHAP risk radar is output: highlight the financial sub-items that contribute most to the overall score, and give improvement suggestions such as "compress accounts receivable" and "reduce liabilities", realizing explainable quantitative scoring.

[0099] 5. Product catalog matching and comprehensive scoring output layer

[0100] This layer receives the supplier product catalog after OCR analysis, generates a CLIP multi-modal vector for each material description; then compares it with the bid material vector library to calculate the "specification-brand-origin" three-dimensional matching degree, and generate a difference list.

[0101] The system summarizes the license qualification rate, robust index, catalog matching degree and chapter stamp credibility output by the previous layer in four dimensions, calculates the comprehensive score through configurable weights; the scoring rules and thresholds are written in YAML, which can be online tuned by the compliance department.

[0102] LLM automatically generates **《Non-conformance List and Rejection Opinion》** according to the scoring results and defect fields: quote regulations or procurement system chapters one by one to explain the reasons for rejection or certificate supplement requirements. The final result is output in the form of JSON data interface + watermarked PDF report, and pushed to the SRM process:

[0103] Green (≥80) → Admitted;

[0104] Yellow (60-79) → Certificate supplement work order;

[0105] Red (<60) → Automatic rejection.

[0106] Although embodiments of the application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, alternatives and variations can be made thereto without departing from the principles and spirit of the application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-modal intelligent compliance pre-screening method, characterized in that: Comprising the following steps: A) Receiving the qualification compressed package uploaded by the supplier and automatically completing the decompression operation; B) Using a layout perception model to accurately label the decompressed pages into five categories: main body license, financial report, industry qualification, product catalog, and environmental safety evaluation, and performing fingerprint hash deduplication processing on homologous files to generate a unique document identifier DocID; C) For DocID labeled as main body license, perform optical character recognition, call at least one national regulatory interface to compare the license authenticity, status and validity period, and generate a four-level risk label of red, orange, yellow and green according to the preset rules; D) Extract the business scope, license items and registered capital key fields from the business license text using a large language model, match the business scope vector with the bidding material catalog semantic ontology library, and output out-of-scope or qualification gap prompt information; E) Analyze the financial report for the past three years using a search enhancement-reasoning chain workflow, calculate the current ratio, asset-liability ratio, shareholder equity return rate, operating cash net and accounts receivable turnover rate, and generate a financial stability index combined with industry quantiles; F) First, perform OCR processing on the product catalog, then embed it in multiple modalities, and compare it with the bidding material list vector to obtain the specification matching degree and missing items; G) Comprehensive main body license risk label, financial stability index, catalog matching degree and chapter print credibility, calculate the comprehensive score of 0-100 according to the configurable weight, output the admission, certificate supplement or rejection conclusion based on the score threshold, and generate a list of non-compliance items containing regulations.

2. The multi-modal intelligent compliance pre-screening method of claim 1, wherein: The layout perception model in step B uses LayoutLMv3, and the fingerprint hash deduplication simultaneously uses fast perceptual hash FP-Hash and SHA-256 algorithm to directly label and skip the subsequent verification process for duplicate files identified.

3. The multi-modal intelligent compliance pre-screening method of claim 2, wherein: The national regulatory interface called in step C includes at least the national enterprise credit information public system interface, the Ministry of Housing and Urban-Rural Development qualification interface and the State Drug Administration license interface, and performs unified field normalization processing on the operation abnormality, suspension and punishment records returned by the interface to facilitate subsequent accurate analysis and risk labeling.

4. The multi-modal intelligent compliance pre-screening method of claim 3, wherein: The search enhancement-reasoning chain in step E specifically includes: a) Through vector search IFRS / GAAP financial definition fragments, provide accurate definition basis for financial indicator calculation; b) Call a large language model to accurately locate the financial table cell coordinates to ensure the accuracy of financial data extraction; c) Use built-in functions in the model to automatically calculate five financial indicators and output chain explanations to facilitate auditors to understand the calculation process and basis.

5. The multi-modal intelligent compliance pre-screening method of claim 4, wherein: It also includes steps to optimize and improve the audit process and results: Collect comprehensive logs on "model determination and manual review differences", fine-tune the large language model entity extractor and comprehensive score weight offline, and based on the data and feedback accumulated during the running period, the audit accuracy is gradually optimized and improved with the running period, and the performance of the multi-modal intelligent compliance pre-audit method is continuously improved.

6. A system for use in the method of multi-modal intelligent compliance pre-screening as claimed in claim 5, characterized in that: It includes: File receiving and decompression module: used to receive the qualification compressed package uploaded by the supplier and automatically complete the decompression operation, providing basic files for subsequent processing; File classification and identification module: The decompressed page is labeled as five categories of main certificate, financial report, industry qualification, product catalog and environmental safety evaluation by using a layout perception model. Meanwhile, the same source files are executed fingerprint hash deduplication to generate a unique document identification DocID, realizing accurate classification and unique identification of files. Certificate verification module: The DocID labeled as main certificate is subjected to optical character recognition, at least one national regulatory interface is called to compare the license authenticity, status and validity period, and a four-level risk label of red, orange, yellow and green is formed according to the preset rules to complete the preliminary judgment of the compliance of the certificate. Business scope matching module: The business scope, license items and registered capital key fields are extracted from the business license text by using a large language model. The business scope vector is matched with the semantic ontology library of the bidding material catalog to output out-of-scope or qualification gap prompts, assisting in determining whether the supplier's business qualifications meet the requirements. Financial analysis module: The financial report of the past three years is analyzed by using the search-enhanced-reasoning chain workflow. The current ratio, asset-liability ratio, shareholder equity return rate, operating cash net and accounts receivable turnover rate are calculated. Combined with the industry quantile, the financial stability index is generated to comprehensively evaluate the financial condition of the supplier. Product catalog comparison module: After the product catalog is subjected to OCR and multi-modal embedding, it is compared with the bidding material list vector to obtain the specification matching degree and missing items, ensuring that the products provided by the supplier meet the bidding requirements. Comprehensive scoring and decision-making module: The main certificate risk label, financial stability index, catalog matching degree and chapter print credibility are combined to calculate a comprehensive score of 0-100 according to the configurable weight. Based on the scoring threshold, the access, supplementary certificate or rejection conclusion is output, and a list of non-compliance items containing regulations is generated to provide the basis for the final decision.

7. The system of claim 6, wherein: The layout perception model in the file classification and identification module is LayoutLMv3. The fingerprint hash deduplication simultaneously uses the fast perceptual hash FP-Hash and SHA-256 algorithm to directly mark and skip the subsequent verification process for duplicate files, improving the efficiency of file processing.

8. A system as in claim 7, wherein: The certificate verification module calls at least one national regulatory interface, including the national enterprise credit information public system interface, the housing and urban-rural development department qualification interface and the national drug administration license interface. The operating abnormity, revocation and punishment records returned are normalized in a unified field to facilitate accurate analysis and labeling of certificate risks.

9. A system as in claim 8, wherein: The semantic matching of the business scope matching module uses a double-encoder architecture to embed the business scope text into a 768-dimensional vector space and perform cosine matching with the pre-constructed bidding material ontology library node vector. When the maximum matching degree is less than 0.35, an out-of-scope warning is triggered to improve the accuracy of business scope matching.

10. The system of claim 9, wherein: The following optimization modules are also included: Weight configuration module: The weight configuration file of the comprehensive score adopts YAML format, allowing compliance personnel to adjust the weight of each dimension online. The system refreshes the historical pass rate in real time after adjustment and records the version number for traceability, enhancing the flexibility and manageability of the system. The non-compliance item list generation module: the generated non-compliance item list is automatically written by the large language model according to the comprehensive score results and the regulation knowledge base, the output format includes the three-section structure of "problem description-regulation reference-reform suggestion", and the original text page number screenshot is automatically inserted to improve the standardization and readability of the non-compliance item list; The decision execution and feedback module: the access, supplementary certificate or rejection conclusion is written back to the SRM system through the JSON interface and triggers the BPMN process: the comprehensive score is greater than or equal to 80 to automatically access, 60-79 to generate a supplementary certificate work order, and less than 60 to automatically reject and seal the supplier archive, realizing the automatic execution of the decision; The model optimization module: the difference between the model judgment and the artificial review is collected, the entity extractor and the comprehensive score weight of the large language model are fine-tuned offline, the audit accuracy is gradually optimized and improved with the running cycle, and the performance of the system is continuously improved.

Images