Material application automatic filling and auditing method and system based on multi-agent cooperation

Abstract

The application relates to the technical field of enterprise intelligent information management, and discloses a material application automatic filling and auditing method and system based on multi-agent cooperation, which comprises the following steps: a user submits a material application request through text, voice or a Web form; a system generates a structured intention according to the material application request and automatically generates a material application form by calling a retrieval enhancement RAG model; then, a comprehensive verification result is obtained through four-dimensional parallel verification of format, inventory, budget and policy; a rule-machine learning dual-track auditing model is used for synchronous scoring; when conflicts occur, a three-level risk decision is started, including low-risk automatic single passing, medium-risk limited-time sampling and high-risk forced manual review; the review result is used for incremental updating of a model and a rule base online; finally, an approval report is generated and written into a blockchain. Through the RAG enhancement technology and the multi-agent cooperation mechanism, the application realizes full-process, dynamic and intelligent management and control effects from intention understanding to final auditing of the material application.

Description

Technical Field

[0001] This invention relates to the field of enterprise intelligent information management technology, and in particular to a method and system for automatic material application filling and review based on multi-agent collaboration. Background Technology

[0002] The existing automatic material application filling and approval scheme relies heavily on a strictly defined static rule base, lacking a true understanding of the semantics of material application requests. When business rules change, the rule base must be manually updated through manual intervention, lacking dynamic adaptation and evolution capabilities, resulting in high maintenance costs and slow response to changes.

[0003] Existing automated material application and review schemes often lack an effective knowledge connection mechanism with the latest material specifications, tax codes, or compliance regulations. Therefore, their review basis may be outdated, and it cannot ensure that the application content always complies with the latest standards and policies, thus posing a compliance risk.

[0004] Existing attempts to automate the automatic filling and review of material applications mostly target isolated links in the process. There is a lack of a unified, end-to-end intelligent collaborative framework that connects natural language application, structured generation, multi-dimensional verification, and intelligent review. Data interaction between functional modules is not smooth, making it difficult to achieve deep integration of multi-angle information and collaborative decision-making, resulting in bottlenecks in the efficiency of application and review.

[0005] Existing automated material application and review systems generally lack the ability to continuously learn from historical review results and human feedback. They cannot automatically optimize their review rules and verification logic, nor can they provide users with application guidance strategies. As a result, they cannot achieve intelligent evolution of system performance over time and with the accumulation of cases. Summary of the Invention

[0006] The purpose of this invention is to propose a method and system for automatic material application filling and review based on multi-agent collaboration in order to solve the problems in the prior art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: an automatic material application filling and review method based on multi-agent collaboration, comprising the following steps:

[0008] Step S1: The user interface receives the material request and generates standardized request data.

[0009] Step S2: Parse the material request and generate a structured intent description;

[0010] Step S3: Based on the structured intent description, generate a pre-filled material requisition form using a retrieval-enhanced RAG model;

[0011] Step S4: Initiate multi-dimensional parallel verification of the pre-filled material requisition form and output the comprehensive verification result;

[0012] Step S5: Input the comprehensive verification results into the rule-machine learning dual-track review model to generate rule decision results;

[0013] Step S6: The rule decision result is adjudicated and manually reviewed through a preset collaborative adjudication strategy, and the material requisition form that has been adjudicated is written into the ERP interface.

[0014] Step S7: Based on the results of the adjudication and manual review, the parameters of the dual-track audit model are incrementally updated using an online learning algorithm;

[0015] Step S8: Encapsulate the material requisition form, the corresponding approval result, and the decision-making process into an approval report and write it into the blockchain storage.

[0016] A multi-agent collaborative automatic material requisition filling and approval system, comprising:

[0017] The user interface module is used to receive material requisition requests through the user interaction interface and generate standardized request data.

[0018] A dialogue management intelligent agent is used to parse the material application request and generate a structured intent description;

[0019] A form generation agent is used to generate pre-filled material requisition forms based on a retrieval-enhanced RAG model, according to a structured intent description.

[0020] A multi-dimensional verification agent is used to initiate multi-dimensional parallel verification of pre-filled material requisition forms and output comprehensive verification results.

[0021] The model review agent is used to input the comprehensive verification results into the rule-machine learning dual-track review model to generate rule decision results.

[0022] The strategy adjudication module is used to adjudicate and manually review the rule decision results through a preset collaborative adjudication strategy, and write the material requisition form that has been adjudicated into the ERP interface.

[0023] The model self-optimization module is used to incrementally update the parameters of the dual-track review model based on the results of adjudication and manual review through an online learning algorithm;

[0024] The report generation module is used to encapsulate material requisition forms, corresponding approval results, and decision-making processes into approval reports and write them to blockchain storage.

[0025] The beneficial effects of the technical solution provided by this invention include at least the following:

[0026] This invention utilizes a technical approach of BERT-BiLSTM-CRF semantic understanding + RAG retrieval enhancement to automatically generate pre-filled application forms based on user-input natural text, voice, or web forms. Compared to the traditional manual data entry mode, field accuracy is improved, and the filling time is reduced from an average of 10 minutes to 30 seconds, significantly reducing the application cancellation rate.

[0027] This invention adopts a distributed collaborative architecture of four types of intelligent agents: dialogue management, form generation, multi-dimensional verification, and model review. Each intelligent agent can be independently scaled horizontally and upgraded in a gray-scale manner. The approval process is compressed from multiple serial steps to multiple parallel intelligent agents, which greatly improves the overall throughput. Moreover, whenever a new review rule is released, only the corresponding intelligent agent needs to be updated, which effectively reduces the system maintenance cost.

[0028] This invention innovatively proposes a rule-machine learning dual-track audit model and introduces a conflict detection-confidence weighting-three-level threshold adjudication mechanism. The rule track ensures that the decision-making process is 100% auditable, while the machine learning track is trained using 12 months of historical data to reduce the false negative rate of decision-making. When the two conflict, manual sampling is automatically triggered, keeping the high-risk misjudgment rate within 1%, thus achieving controllable intelligent audit decision-making.

[0029] The system of this invention utilizes the review results in real time through an online learning engine, and adopts adaptive gradient descent and rule pruning and regeneration techniques to perform minute-level incremental updates on the dual-track review model. After the system goes online, as the sample size increases, the automatic approval rate of material applications gradually increases, and the amount of manual sampling decreases accordingly. Moreover, the entire process is free of downtime and rollback, truly realizing the intelligent evolution of system performance over time and with the accumulation of cases. Attached Figure Description

[0030] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 A flowchart of the method provided in an embodiment of the present invention;

[0032] Figure 2 This is a system structure diagram provided for an embodiment of the present invention. Detailed Implementation

[0033] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the automatic material application filling and review method and system based on multi-agent collaboration proposed in this invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0034] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0035] The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0036] The following description, in conjunction with the accompanying drawings, details the specific solution of the automatic material application filling and review method and system based on multi-agent collaboration provided by this invention.

[0037] Please see Figure 1 The diagram illustrates a flowchart of an automatic material requisition filling and review method based on multi-agent collaboration, according to an embodiment of the present invention. The method includes the following steps:

[0038] Step S1: The user interface receives the material request and generates standardized request data.

[0039] Step S1 further includes the following sub-steps:

[0040] S1-1, Users can send material requisition requests through the user interface by filling out a web form, inputting voice, or inputting natural text.

[0041] S1-2, perform format recognition and preprocessing on the request to generate standardized request data. Preprocessing includes form field extraction, speech-to-text conversion and text cleaning.

[0042] S1-3: For standardized request data, attach a user identity token and a timestamp, and route the standardized request data to the dialogue management agent through the API gateway.

[0043] It should be noted that form field extraction: based on DOM parsing and XPath / CSS selectors, it automatically locates each control in the web form, extracts key-value pairs and maps them to standard field names, outputs structured data, and supports multilingual tag recognition and data type conversion.

[0044] Speech-to-text: MFCC features are extracted using 16kHz sampling and decoded in real time using a DeepSpeech2 streaming ASR model deployed in a private enterprise environment. The result is timestamped text with text confidence scores.

[0045] Text cleaning: Filter special symbols, HTML tags and sensitive words in the text using regular expressions, perform text segmentation, error correction and normalization using NLTK, and output the cleaned text.

[0046] User identity token: Includes user account ID, real name, department, role, data permission scope, and a 15-minute expiration timestamp. The gateway verifies the user's permissions each time they log in. It supports refreshing and synchronizing the revocation list to ensure that the application interface can only be called by legitimate and authorized users.

[0047] API Gateway: It adopts Kong as the unified entry point, configures mTLS two-way certificate authentication and a 100req / s·key rate limiting policy, automatically routes standardized request data to the corresponding dialogue management agent according to the URL path, supports canary release and A / B testing, and integrates the Zipkin plugin to realize full-link tracing.

[0048] Step S2: Parse the material request and generate a structured intent description;

[0049] Step S2 further includes the following sub-steps:

[0050] S2-1 uses the BERT-BiLSTM-CRF model to perform word segmentation, entity recognition and dependency parsing on standardized request data to generate candidate entities, which include the name, quantity, demand date and requesting department of the material.

[0051] S2-2, semantic disambiguation of candidate entities is performed through the domain dictionary, and standardized mapping is performed through the ontology library to obtain standardized entities. Standardized entities include standard material codes, units of measurement and business attributes.

[0052] S2-3 uses a pre-defined intent classification model to map the intent categories of standardized entities, obtain the intent categories of standardized request data, and output the confidence level. The intent categories include new requests, supplementary requests, urgent requests, and regular requests.

[0053] S2-4 encapsulates the standardized entity, intent category, and confidence level into a structured intent description in JSON format, and adds a unique request identifier and generation timestamp.

[0054] It should be noted that the BERT-BiLSTM-CRF model uses BERT-Base-Chinese as the underlying layer to obtain dynamic character vectors of the context, then connects to a bidirectional LSTM to capture long-distance dependencies, and finally the CRF layer ensures the validity of BIO annotations through the state transition matrix, outputting the globally optimal entity sequence.

[0055] Domain dictionary: Usually maintained jointly by the purchasing, warehousing and process departments, it covers standard material aliases, abbreviations, colloquial expressions and stop words, and is updated monthly incrementally through multi-modal matching of AC automatic machine.

[0056] Ontology library: Built on OWL 2 DL language, it includes core fields such as "material", "unit of measurement", "business attribute" and "department", as well as attribute fields such as "material code", "budget item" and "environmental protection label", and has mutual exclusion, inheritance and constraint rules.

[0057] The pre-built intent classification model is designed based on the lightweight BERT-Tiny+TextCNN scheme. This scheme concatenates three types of convolutional kernels (3×3, 4×4, and 5×5) at the top layer of BERT to extract local key phrase features. The output includes four categories: new application, supplementary application, urgent application, and regular application. The training set used by the model consists of historical application titles and text annotations. It is trained using mixed precision and R-drop regularization.

[0058] Step S3: Based on the structured intent description, generate a pre-filled material requisition form using a retrieval-enhanced RAG model;

[0059] Step S3 further includes the following sub-steps:

[0060] S3-1 extracts material code, required quantity, required date, applying department and intent category from the structured intent description, and concatenates them into a query vector using a BERT-based structured field encoding method;

[0061] S3-2, using a retrieval tool based on the RAG model, performs an approximate nearest neighbor search in the de-identified historical material requisition database and ERP master data index according to the query vector to obtain candidate records. The candidate records include multiple similar requisition records and corresponding metadata fields.

[0062] S3-3 uses a cross-attention mechanism to weight and fuse the query vector and candidate records to generate a draft application form. Then, based on the enterprise's preset required fields and business constraint rules, the draft application form is further modified to fill in missing values, validate numerical ranges, and standardize the format, forming a pre-filled material application form with an attached source citation identifier.

[0063] It should be noted that the structured field encoding method based on the BERT model involves first inputting the key fields (material code, required quantity, required date, applying department, and intent category) into BERT, and then concatenating them into vectors in the order of "material code-required quantity-required date-applying department-intent category". A learnable "application hint" is added to the beginning of the concatenated vector segment to form a fixed 768-dimensional query vector. This method can preserve the order and business weight between fields and avoid information dilution caused by ordinary text concatenation.

[0064] The RAG model-based retrieval system uses HNSW (Approximate Nearest Neighbor Search) to return the top-K records most similar to the current application within milliseconds. These records, along with metadata such as inventory, budget, and suppliers, are sent back to the generator. This approach utilizes real internal enterprise data to complete the fields while preventing data leakage, thus achieving secure retrieval within a closed domain.

[0065] Cross-attention mechanism: The weights of the query vector and candidate records are calculated through cross-attention to obtain a weighted "memory vector". Then, it is aligned with the field template to automatically fill in missing items and verify the range of values ​​and unify the format, generating a business application form that can be submitted directly, which significantly reduces the workload of manual data entry.

[0066] Step S4: Initiate multi-dimensional parallel verification of the pre-filled material requisition form and output the comprehensive verification result;

[0067] Step S4 further includes the following sub-steps:

[0068] S4-1 performs compliance checks on the required fields, data types, numerical ranges, and format specifications of pre-filled material requisition forms, calculates field-level compliance scores and error lists, and outputs format verification results;

[0069] S4-2 calculates inventory gaps, inventory satisfaction indicators, and expected stockout risk levels by calling the ERP inventory interface in real time, based on material codes and required quantities, and outputs inventory verification results.

[0070] S4-3, by reading the current budget amount and the amount already used by the requesting department, gives the proportion of the requested amount used in this material request and the warning indicator, and outputs the budget verification result;

[0071] S4-4 matches the pre-filled material requisition form with the preset procurement policy rule set, checks whether it violates the centralized procurement policy, energy conservation and environmental protection policy or supplier access policy, and outputs the policy verification result.

[0072] S4-5 uses a dynamic weight allocation method based on historical misjudgment rate inversion to weight and fuse the format verification results, inventory verification results, budget verification results, and policy verification results to obtain a comprehensive verification result and output a four-dimensional verification report.

[0073] It should be noted that the ERP inventory interface uses the material code as the primary key and adopts a RESTful style to return the current inventory, in-transit quantity, frozen quantity, and safety stock threshold in real time, providing a quantitative basis for subsequent approvals.

[0074] The centralized procurement policy requires that once the annual cumulative amount of a single type of material exceeds the limit, it must be included in the group's unified bidding. Departments are prohibited from placing orders on their own. Its version is synchronized with the group's procurement department quarterly and supports gray-scale adjustments to ensure that local companies can adapt to the latest catalog in a timely manner.

[0075] The energy conservation and environmental protection policy includes a built-in list of prohibited substances and a national energy efficiency database. During verification, the material code is compared first, and then the environmental protection indicator field in the product specification is read. If no environmental protection certificate is provided or the energy efficiency level does not meet the requirements, a violation mark is triggered. Violation records can be used for annual ESG audits.

[0076] The supplier access policy requires that partners must be on the group's qualified list and not be blacklisted or frozen. The policy verification uses the suggested supplier code in the material requisition form as the key to call the master data MDM service to verify the validity of its business license, ISO certificate and blacklist records. If the certificate is expired or the supplier is blacklisted, a violation mark is triggered and three qualified suppliers in the same category are automatically recommended for the user to replace them, thereby reducing the risk of subsequent contracts.

[0077] The dynamic weight allocation method based on historical misjudgment rate inversion: that is, to count the number of misjudgments in each dimension over the past twelve months, calculate the "false negative rate" (FNR) of format, inventory, budget and policy and use it as dynamic weights, which is automatically updated once a quarter. When merging, a weighted geometric average is used to ensure that the high-confidence dimension dominates the results and continuously improve the accuracy and adaptability of the comprehensive verification.

[0078] Step S5: Input the comprehensive verification results into the rule-machine learning dual-track review model to generate rule decision results;

[0079] Step S5 further includes the following sub-steps:

[0080] S5-1, in the rule engine track of the dual-track audit model, extracts the four-dimensional tag sequences of format, inventory, budget and policy from the comprehensive verification results through the 4D-RCM method and matches them with configurable rule chains to generate rule decision labels;

[0081] S5-2, in the machine learning track of the dual-track review model, the vectorized sequence of each dimension of the comprehensive verification result is input into the pre-trained gradient boosting tree model to obtain the risk probability interval. The risk probability interval is further mapped to ML decision labels, including pass, warning or rejection, through the probability mapping table.

[0082] S5-3, unify the encoding of rule decision labels and ML decision labels to generate dual-track decision vectors, and assign confidence levels to them respectively through business presets to obtain rule decision vectors and ML decision vectors;

[0083] S5-4: Compare the dual-track decision vectors using a conflict detector. If they match, adopt the result directly as the rule-based decision. If they conflict, proceed to step S6 and mark the cause of the conflict.

[0084] It should be noted that the 4D-RCM method, or four-dimensional rule chain matching method, arranges the four-dimensional verification tags of format, inventory, budget, and policy into a 4-bit tag sequence in a fixed order. It then matches these tags sequentially through a pre-configured decision chain containing condition → action → priority triples, achieving zero-black-box, auditable rule reasoning. An example is given below:

[0085] ① Extract verification flags (0 = normal, 1 = abnormal):

[0086] For example: Format dimension: 0, Inventory dimension: 1, Budget dimension: 1, Policy dimension: 0, then the synthesized 4-bit tag sequence is: 0110;

[0087] ② Match according to the configurable rule chain in sequence:

[0088]

[0089] For example, if the label sequence is 0110, then the output rule decision label is: Reject - Multiple Risks (Budget Gap / Insufficient Inventory).

[0090] Pre-trained gradient boosting tree model: This model uses LightGBM as the framework and uses closed applications from the past year of the enterprise as training samples (positive and negative samples in a ratio of 1:3). It outputs the risk probability in the range of 0 to 1 for each of the four-dimensional labeled sequences. The model is updated incrementally at regular intervals. The inference phase takes an average of 8ms on CPU, which meets the requirements of high concurrency and real-time scoring.

[0091] Here is an example of mapping risk probability intervals to ML decision labels and attaching confidence levels using a probability mapping table:

[0092] Risk probability range ML decision labels Confidence 0.00~0.30 pass 0.90 0.30~0.45 Warning - Low 0.75 0.45~0.70 Warning - High 0.65 0.70~0.85 Reject - Low 0.80 0.85~1.00 Rejection - High 0.92

[0093] Step S6: Adjudicate the rule decision result through a preset collaborative adjudication strategy and conduct manual review, and write the material requisition form that passes the adjudication into the ERP interface;

[0094] Among them, in step S6, the following sub-steps are further included:

[0095] S6-1: Read the decision label, confidence level, and conflict reason in the rule decision result, and calculate the comprehensive risk score R through a preset scoring function. The comprehensive risk score R is between 0 and 1;

[0096] S6-2: Preset a low-risk threshold T1 and a high-risk threshold T2. The initial values of T1 and T2 are preset as T1 = 0.3 and T2 = 0.7 respectively, and they can be dynamically updated through the FAR curve of the historical false positive rate in the recent 12 months;

[0097] S6-3: Compare the comprehensive risk score R with the preset low-risk threshold T1 and high-risk threshold T2 through a three-level threshold comparator, and output the adjudication grade, including: output the "low-risk" grade when R ≤ T1, output the "medium-risk" grade when T1 < R < T2, and output the "high-risk" grade when R ≥ T2;

[0098] S6-4: When the adjudication grade is "low-risk", encapsulate the requisition number, material details, and approval mark of the material requisition form and write them into the ERP interface through the API gateway;

[0099] S6- When the adjudication grade is "medium-risk", encapsulate the requisition number, material details, approval mark, and budget freeze condition of the material requisition form, and request a time-limited manual random inspection. If the inspection is qualified, write it into the ERP interface manually. If the inspection is unqualified, convert it to the "high-risk" grade;

[0100] S6-6: When the adjudication grade is "high-risk", automatically lock the requisition number of the material requisition form and push a manual review request to the review work pool. The manual review request includes the risk reason, decision basis, and recommended measures.

[0101] It should be noted that the FAR curve of the historical false positive rate takes the decision results of the rule-machine learning dual-track decision model in the past 12 months as samples, and divides the comprehensive risk score R (0, 1) into 100 equal-width intervals, and statistically calculate the ratio of the number of application forms that "are automatically passed but finally rejected manually" in each interval to the total number of application forms (i.e., the FAR value). This curve usually shows a monotonic increase. Take the R values corresponding to FAR = 5% and FAR = 1% on the curve as the new T1 and T2 thresholds to replace the default T1 = 0.3 and T2 = 0.7, so as to reduce the probability of repeated mistakes. The threshold adjustment needs to be confirmed and take effect by the enterprise procurement director, which not only makes the threshold adapt to the business changes, but also retains the manual fallback right.

[0102] Step S7: Based on the results of the adjudication and manual review, the parameters of the dual-track audit model are incrementally updated using an online learning algorithm;

[0103] Step S7 further includes the following sub-steps:

[0104] S7-1 Extract the comprehensive verification vector, decision label and confidence level corresponding to the material requisition forms that have passed the adjudication, passed after manual review and rejected after manual review, and encapsulate them as training samples, and input the training samples into the online learning engine;

[0105] S7-2, in the online learning engine, the machine learning track of the rule-machine learning dual-track audit model is updated by using the adaptive gradient descent algorithm, adjusting its leaf node weights and split gain to obtain the updated machine learning model file;

[0106] S7-3, in the online learning engine, updates the rule engine track of the rule-machine learning dual-track audit model through rule pruning and regeneration methods, deactivates rules with declining support and supplements new rules through frequent item mining methods, and obtains the updated rule base version number;

[0107] S7-4 writes the updated machine learning model file and rule base version number back to the dual-track audit model, and records the update time and training sample size.

[0108] It should be noted that the online learning engine is an independent microservice engine deployed within Kubernetes. It captures the decision result data stream in real time through Kafka, caches the latest samples in a sliding window, and automatically triggers training and completes incremental updates to the dual-track decision model when the window is full.

[0109] Rule pruning and regeneration method: Count the number of hits for each rule in the past three months and automatically deactivate rules with fewer than 0.1% of the total number of hits. Then, use FP-Growth to mine high-frequency hit rules on new samples to generate a candidate rule set. After manual review, the rules are merged into the rule base. The rule base version number is automatically incremented by SemVer. This method eliminates outdated rules and adds new business scenarios, ensuring that the rule engine evolves in sync with the machine learning track.

[0110] Step S8: Encapsulate the material requisition form, corresponding approval results, and decision-making process into an approval report and write it into the blockchain storage;

[0111] Step S8 further includes the following sub-steps:

[0112] S8-1 encapsulates the material requisition form, corresponding approval results, and decision-making process into a JSON structured approval report, and uses the system SM2 private key issued by the enterprise-level CA to sign the report digest with a private key;

[0113] S8-2 encapsulates the approval report, private key signature, and timestamp into a transaction payload, and calls the blockchain SDK interface to upload the transaction payload to the blockchain mempool;

[0114] S8-3 If the transaction payload fails to be uploaded to the blockchain, a retry is triggered. If the retry fails after three attempts, an error code is generated, the corresponding approval report is encrypted and cold-stored, and an alarm notification is sent to the relevant management personnel.

[0115] It should be noted that the system SM2 private key issued by the enterprise-level CA is uniformly issued by the company's internal certificate center, is deployed only within the encryption machine and cannot be exported, ensuring the authenticity and legal validity of the approval data and meeting the compliance requirements of the national cryptographic algorithm.

[0116] Blockchain SDK Interface: The blockchain business layer packages the signed approval report, signature, and timestamp all at once, calls the write function provided by the SDK to convert the data format into a transaction structure recognized on the chain, completes node addressing, transaction numbering, caching, and retries, and can send the approval report into the blockchain mempool within seconds, realizing rapid connection between business and evidence storage, and returning the transaction hash for subsequent querying.

[0117] Please see Figure 2 This diagram illustrates the system architecture of an automatic material requisition filling and review system based on multi-agent collaboration, according to an embodiment of the present invention. The system includes:

[0118] The user interface module is used to receive material requisition requests through the user interaction interface and generate standardized request data.

[0119] A dialogue management intelligent agent is used to parse material requisition requests and generate structured intent descriptions;

[0120] A form generation agent is used to generate pre-filled material requisition forms based on a retrieval-enhanced RAG model, according to a structured intent description.

[0121] A multi-dimensional verification agent is used to initiate multi-dimensional parallel verification of pre-filled material requisition forms and output comprehensive verification results.

[0122] The model review agent is used to input the comprehensive verification results into the rule-machine learning dual-track review model to generate rule decision results.

[0123] The strategy adjudication module is used to adjudicate and manually review the rule decision results through preset collaborative adjudication strategies, and write the material requisition forms that have been adjudicated into the ERP interface.

[0124] The model self-optimization module is used to incrementally update the parameters of the dual-track review model based on the results of adjudication and manual review through an online learning algorithm;

[0125] The report generation module is used to encapsulate material requisition forms, corresponding approval results, and decision-making processes into approval reports and write them to blockchain storage.

[0126] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for automatic material requisition filling and approval based on multi-agent collaboration, characterized in that, The method includes: Step S1: The user interface receives the material request and generates standardized request data. Step S2: Parse the material request and generate a structured intent description; Step S3: Based on the structured intent description, generate a pre-filled material requisition form using a retrieval-enhanced RAG model; Step S4: Initiate multi-dimensional parallel verification of the pre-filled material requisition form and output the comprehensive verification result; Step S5: Input the comprehensive verification results into the rule-machine learning dual-track review model to generate rule decision results; Step S6: The rule decision result is adjudicated and manually reviewed through a preset collaborative adjudication strategy, and the material requisition form that has been adjudicated is written into the ERP interface. Step S7: Based on the results of the adjudication and manual review, the parameters of the dual-track audit model are incrementally updated using an online learning algorithm; Step S8: Encapsulate the material requisition form, corresponding approval results, and decision-making process into an approval report and write it into the blockchain storage; Step S4 further includes the following sub-steps: S4-1 performs compliance checks on the required fields, data types, numerical ranges, and format specifications of pre-filled material requisition forms, calculates field-level compliance scores and error lists, and outputs format verification results; S4-2 calculates inventory gaps, inventory satisfaction indicators, and expected stockout risk levels by calling the ERP inventory interface in real time, based on material codes and required quantities, and outputs inventory verification results. S4-3, by reading the current budget amount and the amount already used by the requesting department, gives the proportion of the requested amount used in this material request and the warning indicator, and outputs the budget verification result; S4-4 matches the pre-filled material requisition form with the preset procurement policy rule set, checks whether it violates the centralized procurement policy, energy conservation and environmental protection policy or supplier access policy, and outputs the policy verification result. S4-5, The format verification result, inventory verification result, budget verification result and policy verification result are weighted and fused by a dynamic weight allocation method based on historical misjudgment rate inversion to obtain a comprehensive verification result and output a four-dimensional verification report; The dynamic weight allocation method based on historical misjudgment rate inversion: that is, to count the number of misjudgments in each dimension over the past twelve months, calculate the "false negative rate" (FNR) of format, inventory, budget and policy and use it as dynamic weights, which are automatically updated once a quarter. When merging, a weighted geometric average is used to ensure that the high-confidence dimension dominates the results and continuously improve the accuracy and adaptability of the comprehensive verification. Step S5 further includes the following sub-steps: S5-1, in the rule engine track of the dual-track audit model, uses the 4D-RCM method to extract the format, inventory, budget and policy tag sequences, match the rule chain and generate rule decision tags; 4D-RCM method: also known as four-dimensional rule chain matching method, which arranges the four-dimensional verification tags of format, inventory, budget and policy into a 4-bit tag sequence in a fixed order, and matches them sequentially through a pre-configured decision chain containing condition → action → priority triplet, to achieve zero black box and auditable rule reasoning; S5-2, in the machine learning track of the dual-track review model, after vectorizing the labeled sequences of each dimension, input them into the pre-trained gradient boosting tree model to obtain a risk probability interval, and generate ML decision labels of pass, warning or rejection through a mapping table; S5-3, encode the rule decision label and the ML decision label uniformly to generate a dual-track decision vector, and assign confidence levels respectively through business presets to obtain a rule decision vector and an ML decision vector; S5-4, compare the dual-track decision vectors through a conflict detector. If they are consistent, directly adopt it as the rule decision result. If there is a conflict, transfer to step S6 and mark the reason for the conflict; Among them, in step S6, the following sub-steps are further included: S6-1, read the decision label, confidence level and reason for the conflict in the rule decision result, and calculate the comprehensive risk score R through a preset scoring function. The comprehensive risk score R is between 0 and 1; S6-2, preset a low risk threshold T1 and a high risk threshold T2. The initial values of T1 and T2 are preset as T1 = 0.3 and T2 = 0.7 respectively, and they can be dynamically updated through the FAR curve of the historical misjudgment rate in the recent 12 months; S6-3, compare the comprehensive risk score R with the preset low risk threshold T1 and high risk threshold T2 through a three-level threshold comparator, and output the adjudication level, including: output the "low risk" level when R ≤ T1, output the "medium risk" level when T1 < R < T2, and output the "high risk" level when R ≥ T2; S6-4, when the adjudication level is "low risk", encapsulate the application number, material details and approval mark of this material requisition form and write them into the ERP interface by the API gateway; S6-5, when the adjudication level is "medium risk", encapsulate the application number, material details, approval mark and budget freeze condition of this material requisition form, and request a time-limited manual random inspection. If the random inspection is qualified, manually write it into the ERP interface. If the random inspection is unqualified, convert it to the "high risk" level; S6-6, when the adjudication level is "high risk", automatically lock the application number of this material requisition form and push a manual review request to the review work pool. The manual review request includes the risk reason, decision basis and recommended measures; Among them, in step S7, the following sub-steps are further included: S7-1, extract the comprehensive verification vector, decision label and confidence level corresponding to the material requisition forms that have passed the adjudication, passed after manual review and been rejected after manual review, encapsulate them as training samples, and input the training samples into the online learning engine; S7-2, in the online learning engine, update the machine learning track of the rule-machine learning dual-track review model through the adaptive gradient descent algorithm, adjust its leaf node weights and split gains, and obtain an updated machine learning model file; S7-3, in the online learning engine, update the rule engine track of the rule-machine learning dual-track review model through the rule pruning and regeneration method, deactivate the rules with decreasing support and supplement new rules through the frequent item mining method, and obtain an updated rule library version number; S7-4, Write the updated machine learning model file and rule base version number back to the dual-track audit model, and record the update time and training sample size.

2. The method for automatic filling and review of material requests based on multi-agent collaboration as described in claim 1, characterized in that: Step S1 further includes the following sub-steps: S1-1, Users can send material requisition requests through the user interface by filling out a web form, inputting voice, or inputting natural text. S1-2, perform format recognition and preprocessing on the request to generate standardized request data. The preprocessing includes form field extraction, speech-to-text conversion and text cleaning. S1-3, Attach a user identity token and a timestamp to the standardized request data, and route the standardized request data to the dialogue management agent through the API gateway.

3. The method for automatic filling and review of material requests based on multi-agent collaboration as described in claim 1, characterized in that: Step S2 further includes the following sub-steps: S2-1 uses the BERT-BiLSTM-CRF model to perform word segmentation, entity recognition and dependency parsing on standardized request data to extract material name, quantity, demand date and candidate entities of the requesting department; S2-2 utilizes a domain dictionary to perform semantic disambiguation on candidate entities and extracts standardized entities of standard material codes, units of measurement, and business attributes through ontology-based standardized mapping. S2-3: Use a pre-built intent classification model to map the intent of standardized entities and output the intent categories of new, supplementary, urgent and regular applications and their confidence levels. S2-4 encapsulates the standardized entity, intent category, and confidence level into a structured intent description in JSON format, and adds a unique request identifier and generation timestamp.

4. The method for automatic filling and review of material requests based on multi-agent collaboration as described in claim 1, characterized in that: Step S3 further includes the following sub-steps: S3-1 extracts material code, required quantity, required date, applying department and intent category from the structured intent description, and concatenates them into a query vector using a BERT-based structured field encoding method; S3-2, using the RAG retrieval tool to perform a nearest neighbor search in the de-identified historical application database and ERP master data index based on the query vector, to obtain similar application records and corresponding metadata; S3-3 uses a cross-attention mechanism to fuse query vectors and candidate records, generates a draft application form, and completes the completion, verification, and format standardization based on the enterprise's required fields and business rules, forming a pre-filled application form with source identifier.

5. The method for automatic filling and review of material requests based on multi-agent collaboration as described in claim 1, characterized in that: Step S8 further includes the following sub-steps: S8-1 encapsulates the material requisition form, corresponding approval results, and decision-making process into a JSON structured approval report, and uses the system SM2 private key issued by the enterprise-level CA to sign the report digest with a private key; S8-2 encapsulates the approval report, private key signature, and timestamp into a transaction payload, and calls the blockchain SDK interface to upload the transaction payload to the blockchain mempool; S8-3 If the transaction payload fails to be uploaded to the blockchain, a retry is triggered. If the retry fails after three attempts, an error code is generated, the corresponding approval report is encrypted and cold-stored, and an alarm notification is sent to the relevant management personnel.

6. A multi-agent collaborative automatic material requisition filling and review system for the method of claim 1, characterized in that, The system includes: The user interface module is used to receive material requisition requests through the user interaction interface and generate standardized request data. A dialogue management intelligent agent is used to parse the material application request and generate a structured intent description; A form generation agent is used to generate pre-filled material requisition forms based on a retrieval-enhanced RAG model, according to a structured intent description. A multi-dimensional verification agent is used to initiate multi-dimensional parallel verification of pre-filled material requisition forms and output comprehensive verification results. The model review agent is used to input the comprehensive verification results into the rule-machine learning dual-track review model to generate rule decision results. The strategy adjudication module is used to adjudicate and manually review the rule decision results through a preset collaborative adjudication strategy, and write the material requisition form that has been adjudicated into the ERP interface. The model self-optimization module is used to incrementally update the parameters of the dual-track review model based on the results of adjudication and manual review through an online learning algorithm; The report generation module is used to encapsulate material requisition forms, corresponding approval results, and decision-making processes into approval reports and write them to blockchain storage.

Images