A decision-making method based on fine-tuning of structured inference chains and multi-agent collaboration

Abstract

This invention relates to a decision-making method based on structured reasoning chain fine-tuning and multi-agent collaboration, comprising: generating a corresponding three-dimensional structured link based on each user dialogue sample in a cold-start dialogue dataset; pairing each user dialogue sample with the three-dimensional structured link to construct a supervised training sample set; using the supervised training sample set to perform supervised fine-tuning on a base large language model to obtain a first model; using a pre-built public knowledge base, generating weighted keyword queries and semantic vector queries based on the current user dialogue and the three-dimensional structured link generated in real time by the first model, performing BM25 keyword recall and HNSW vector semantic recall, and generating professional context based on the two recall results; generating a decision support report based on the current user dialogue, the three-dimensional structured link, and the professional context; quantitatively evaluating the decision support report; and reprocessing if the evaluation result does not reach the evaluation threshold until the preset conditions are met to obtain the final decision support report.

Description

Technical Field

[0001] This invention relates to the fields of computer data processing and artificial intelligence technology, specifically to a decision-making method based on structured reasoning chain fine-tuning and multi-agent collaboration. Background Technology

[0002] With the application of large language models and intelligent agent technology in scenarios such as platform business management, compliance review, and business judgment, automatically retrieving professional information based on user questions and generating decision support reports has become an important way to improve the efficiency of processing complex business information. However, platform business questions are usually characterized by non-standard expressions, context spanning multiple rounds, multiple participating entities, overlapping constraints, and scattered professional basis. Existing platform business processing systems and general large language models are unable to reliably complete the continuous processing from question understanding to report output.

[0003] The existing technology has the following main technical defects: (1) Insufficient parsing of user intent and business constraints: Traditional platform business systems mostly rely on keyword matching or fixed rules, while general large language models mostly generate answers directly based on prompt words. Both lack a unified structured representation of question type, target result, key sub-question, participating subject, analysis object and constraint conditions, which can easily lead to deviation in main question identification, omission of constraints and loss of information in multi-turn dialogues; (2) Insufficient training and evaluation of specialized models for complex scenarios: Under the condition of insufficient real historical question and answer data, the general large language model lacks supervised samples for platform business scenarios, making it difficult to stably output intermediate links with complete and parsable fields; at the same time, existing solutions usually only evaluate the final text, lacking a quantifiable comparison mechanism for structural integrity, field accuracy and the degree of conformity of data requirements description. (3) Insufficient accuracy and consistency of source in professional data retrieval: Simple keyword retrieval is difficult to recall data with different words but related semantics, while simple vector retrieval is prone to missing precise terms such as legal names, subject names and professional terms; when the original documents, document-level source information and fragment-level retrieval index are managed separately, it is also easy to produce problems where the retrieval results cannot accurately correspond to the original source. (4) The report content is disconnected from the professional basis: Existing generative systems often allow the model to memorize and supplement external facts, legal rules, research conclusions or case content based on its own parameters. They lack mandatory association with actual usage segments and their document identifiers, page numbers and source addresses. They also cannot reliably mark the content to be verified when there is insufficient data, making it difficult to trace the basis of the report. (5) Lack of collaborative feedback in each processing stage: Existing solutions usually treat intent understanding, data retrieval and report generation as independent sequential stages. Identification errors or insufficient recall in the preceding stages will be directly transmitted to the final report. When the report does not meet the requirements, the system is difficult to return to the corresponding stage for reprocessing based on the reason for the loss, and it is also impossible to form a quality optimization closed loop with limited number of times control and manual review exit.

[0004] Therefore, how to enable the model to stably generate a structured intermediate link that can represent the problem objectives, business scenarios and data requirements under the condition of insufficient historical data, and how to collaboratively complete the retrieval of traceable public information, the generation of structured reports, objective evaluation and targeted feedback based on the intermediate link, is a technical problem that urgently needs to be solved in the business decision support process of complex platforms. Summary of the Invention

[0005] Based on the above analysis, the embodiments of the present invention aim to provide a decision-making method and system based on structured reasoning chain fine-tuning and multi-agent collaboration, in order to solve the technical problems of insufficient intent parsing, unstable model training, insufficient retrieval accuracy and consistency with sources, disconnect between report content and basis, and lack of closed-loop feedback in existing methods.

[0006] The objective of this invention is mainly achieved through the following technical solutions: This invention provides a decision-making method based on structured inference chain fine-tuning and multi-agent collaboration, comprising the following steps: Step S1: Generate a corresponding 3D structured link based on each user dialogue sample in the cold start dialogue dataset; pair each user dialogue sample with the corresponding 3D structured link to construct a supervised training sample set; Step S2: The base large language model is fine-tuned under supervision using the supervised training sample set to obtain the first model; wherein, the first model is used to generate the corresponding three-dimensional structured link in real time based on the user dialogue of the input natural language. Step S3: Using a pre-built public knowledge base, based on the current user dialogue and the three-dimensional structured link generated in real time by the first model, weighted keyword query and semantic vector query are generated respectively. BM25 keyword recall and HNSW vector semantic recall are executed in parallel. Based on the two recall results, the corresponding professional context is generated. Step S4: Based on the current user dialogue, three-dimensional structured link, and professional context, a multi-agent collaborative decision support report is generated; the decision support report is quantitatively evaluated; if the evaluation result does not reach the evaluation threshold, the process returns to steps S2, S3, or S4 for reprocessing until the preset conditions are met to obtain the final decision support report.

[0007] Furthermore, based on each user dialogue sample in the cold start dialogue dataset, a corresponding three-dimensional structured link is generated, including: Based on single-turn or multi-turn dialogues, and utilizing pre-defined link generation template constraints, each user dialogue sample is subjected to question identification, scenario identification, and requirement description identification to obtain the corresponding three-dimensional structured link. Among them, the user dialogue sample includes user questions, supplementary conditions, historical follow-up questions, business constraint information, and expected output targets. The three-dimensional structured link includes problem-dimensional information, scene-dimensional information, and basis-dimensional information; The issue dimension information includes the user's issue type, target result, set of key sub-issues, and issue priority: The scenario dimension information includes scenario type, questioning user role, set of relevant participants, analysis object, and set of constraints. The information provided is a description of the requirements and is used to indicate the direction of information to be focused on in subsequent public information retrieval. It does not indicate specific evidence content, does not limit specific information sources, and does not include internal business data of the platform.

[0008] Furthermore, the preset link generation template constraints include top-level structure constraints, field-level format constraints, content boundary constraints, and serialization format constraints; The top-level structural constraints are used to force the three-dimensional structured link to include three dimensions: problem dimension information, scene dimension information, and basis dimension information, and the order of these dimensions must be fixed. No dimension may be missing, and no undefined additional fields may be added. The field-level format constraints are used to specify the data type and value specifications for each field in the problem dimension information, scenario dimension information, and basis dimension information. The content boundary constraints are used to prohibit the basis dimension information from containing specific evidence content, and to prohibit the self-creation of constraints in the constraint condition set of the scenario dimension information when the user has not explicitly given constraints. The serialization format constraint is used to require the three-dimensional structured link to be serialized into a parsable data format in a fixed field order.

[0009] Furthermore, each user dialogue sample is paired with its corresponding 3D structured link to construct a supervised training sample set, including: Each supervised training sample includes input data and output data; The input data is a single-turn or multi-turn dialogue between users; the output data is a three-dimensional structured link corresponding to the input data. All supervised training samples are divided into a training set and a test set according to a preset ratio; the training set is used to update model parameters during the training of the first model, and the test set is used to evaluate the performance after the first model is trained.

[0010] Furthermore, the base large language model is fine-tuned under supervision using the aforementioned supervised training sample set to obtain a first model, including: Using the samples in the training set as training data, the input data as input to the base big language model, and the output data as the supervised target output of the base big language model, the base big language model learns the mapping relationship from user dialogue in natural language to standardized link data. The output of the supervision target is converted into a tag sequence, which consists of multiple tags formed by sequentially splitting the output of the supervision target; Using the cross-entropy loss function, for each labeled position in the supervised target output, the probability that the model predicts that position is correctly labeled is calculated. The predicted probabilities of all labeled positions are summed and the negative logarithm is taken to obtain the loss value of the training sample. The loss value is used to quantify the difference between the model prediction result and the standard output. Based on the loss value, the parameters of the base large language model are updated through backpropagation. The training is iterated with the goal of minimizing the loss value. After completing a preset number of training rounds, the model parameters are saved to obtain the first model.

[0011] Furthermore, based on the current user dialogue and the three-dimensional structured link generated in real time by the first model, weighted keyword queries and semantic vector queries are generated respectively, including: The problem dimension information, scenario dimension information and basis dimension information are read from the three-dimensional structured link. Chinese word segmentation and stop word filtering are performed on the read field content. According to the preset weight rules, the segmented words are assigned weights to form a weighted keyword query. The non-empty fields in the three-dimensional structured link are concatenated into semantic query text according to a preset order. A fixed retrieval instruction is appended before the semantic query text and then input into a pre-trained text embedding model to generate a semantic vector query. The weighted keyword query is used to perform BM25 keyword retrieval in the public information knowledge base, and the semantic vector query is used to perform HNSW vector semantic retrieval in the public information knowledge base.

[0012] Furthermore, BM25 keyword recall and HNSW vector semantic recall are performed in parallel, and corresponding professional contexts are generated based on the two recall results, including: Based on the requirement description in the three-dimensional structured link and the set of constraints in the scene dimension, the data type is determined and data type filtering conditions are generated. Based on the data type filtering conditions, the BM25 algorithm is used to perform keyword recall on the weighted keyword query to obtain the BM25 keyword recall list; in parallel, based on the data type filtering conditions, the HNSW algorithm is used to perform vector semantic recall on the semantic vector query to obtain the semantic recall list. A reciprocal ranking fusion algorithm is used to merge and sort the BM25 keyword recall list and the semantic recall list. The reciprocal ranking fusion algorithm calculates the fusion score only using the ranking of each segment in the BM25 keyword recall list and the semantic recall list. For segments that appear only in one recall list, only that ranking is used to calculate the score. For segments that appear in both the BM25 keyword recall list and the semantic recall list, the scores corresponding to the rankings in both lists are added together. All segments are then merged and sorted from high to low according to the fusion score. The results after fusion and sorting are sequentially subjected to source integrity verification, text hash deduplication, vector cosine similarity near-duplicate filtering, and control of the number of homologous segments; based on the fusion score, no more than five segments are selected to generate corresponding professional contexts.

[0013] Furthermore, the pre-built public knowledge base is obtained through the following methods: Publicly available information is obtained through automatic downloading, API call extraction, or manual import; the publicly available information includes at least one or more of the following: academic papers, professional textbooks, publicly available professional cases, laws and regulations, and publicly available rule documents. For each piece of publicly available information obtained, a unique document identifier is generated, and the document title, author, publishing organization, information type, publication date, source address, acquisition time, and original file hash value are recorded; when multiple original files have the same hash value, the one with the earliest acquisition time is retained; The publicly available information is processed sequentially to perform document format recognition and text extraction to obtain the document body. The document body is segmented into segments, with each document divided into several document segments, and the text of each segment controlled within a preset word count range; Generate a unique segment identifier for each document segment, and record its document identifier, author, publishing organization, page number, document type and source address to form a standardized segment text; The standardized fragment text is encoded using a pre-trained Chinese text embedding model, and fragment vectors are obtained after L2 normalization. A keyword inverted index of the standardized fragment text and a vector index of the fragment vector are established in the fragment retrieval index. The keyword index and the vector index point to the same fragment through the same fragment identifier.

[0014] Furthermore, based on the current user dialogue, three-dimensional structured links, and professional context, multi-agent collaborative generation of decision support reports is performed, including: Write the target results, key sub-problems, problem priorities, and constraint sets in the three-dimensional structured link into a preset report template; The report generation agent uses the text fragments in the professional context as the sole source of external facts and professional evidence, and generates a decision support report according to the preset report template. When the report generation agent uses the professional fragments in the professional context, it simultaneously outputs the fragment identifier corresponding to the fragment. Based on the fragment identifiers actually used by the report-generating agent, the corresponding document title, author or publishing organization, page number and source address are obtained from the fragment retrieval index and the document source index to form a data source list; the data source list is generated based on the index records and is not freely filled in by the report-generating agent. When the professional context is empty or insufficient to support any key sub-problem, mark the corresponding position of the decision support report as insufficient or pending verification of professional information. The report generating agent shall not use its own parameters to memorize and supplement the professional facts that were not provided. The evaluation feedback agent is invoked to perform a quantitative evaluation of the decision support report, and the evaluation results of the decision support report on each evaluation indicator are obtained.

[0015] Furthermore, if the evaluation result does not reach the evaluation threshold, the process returns to steps S2, S3, or S4 for reprocessing until the preset conditions are met to obtain the final decision support report, including: The quantitative assessment includes the following three assessment indicators: Content completeness rate is used to check whether the decision support report answers the key sub-questions in the three-dimensional structured link item by item and retains the user's explicit constraints; Professional basis coverage is used to check whether the analytical items in the decision support report that require the use of external facts, legal rules, research conclusions or case content are related to the document fragments in the professional context. Citation traceability rate is used to check whether the citations in the decision support report can be traced back to the original document corresponding to the user dialogue through fragment identifiers and document identifiers. The evaluation result is obtained by weighting and summing the content completeness rate, professional basis coverage rate, and citation traceability rate according to preset weights. The decision support report passes the evaluation when the evaluation result is not lower than the preset threshold and the citation traceability rate is 100%. When the content completeness rate does not reach the preset threshold, it is determined that a key sub-problem or constraint is missing, and the processing flow is returned to step S4 to regenerate the decision support report. When the coverage of the professional basis does not reach the preset threshold, it is determined that the required professional basis has not entered the professional context, and the corresponding key sub-question and basis requirement are returned to step S3 for re-retrieval; When the fields in the problem dimension information or scene dimension information are significantly inconsistent with the user's original dialogue, it is determined that the problem target or scene information is incorrectly identified, and the process returns to step S2 to regenerate the three-dimensional structured link. If the evaluation still fails after the preset limit of re-executions is reached, the automatic loop stops and manual review begins.

[0016] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. This invention improves the completeness of complex user intent parsing: by using the question dimension, scenario dimension, and basis dimension to provide a unified structured expression of single-turn or multi-turn user dialogues, it can clearly define the problem objectives, key sub-problems, relevant participants, analysis objects, constraints, and public information requirements, reducing the deviation of the main problem and the omission of constraints caused by relying solely on keywords or free text generation; 2. This invention improves the stability of model output in cold start scenarios: It uses scenario-based cold start dialogue samples and three-dimensional structured links to form supervised data, enabling the base model to learn the mapping relationship from natural language dialogue to fixed field links, and performs quantitative evaluation through structural integrity rate, field accuracy rate and requirement description compliance rate, which facilitates objective verification of fine-tuning effects and control of online output format. 3. This invention balances the accuracy and coverage of professional data retrieval: It employs a dual-path retrieval method combining BM25 (a bag-of-words retrieval model widely used in information retrieval search engines to evaluate the relevance of a document to a given query) keyword retrieval with HNSW (Hierarchical Navigable Small World, an efficient approximate nearest neighbor search algorithm) vector semantic retrieval. Furthermore, it utilizes RRF (Reciprocal Rank Fusion, a ranking fusion algorithm that does not require score normalization, specifically designed to merge multi-path retrieval results, such as the keyword retrieval list from BM25 and the vector retrieval list from HNSW, into a unified, superior ranking list) fusion, data type filtering, and multi-level deduplication. This approach not only preserves the precise matching ability of legal names, subject names, and professional terms, but also retrieves data with synonymous expressions and overall semantic relevance. 4. This invention enhances the traceability of professional evidence: It links the fragment retrieval index, document source index, and original file layer by using the fragment ID (chunk_id) and document ID (doc_id) in the document. The data source in the report is generated based on the actual fragment used. When professional data is insufficient, it clearly outputs missing or unverified marks, reducing the risk of the model fabricating professional facts or generating data whose source cannot be traced back. 5. This invention achieves quantifiable closed-loop optimization of report quality: It conducts a unified evaluation of the report from three aspects: content completeness, coverage of professional basis, and traceability of citations. Based on the reasons for the loss of points, the task is redirected back to the link generation, data retrieval, or report generation stage for reprocessing, so that the error can be located to the specific processing stage and avoids aimless regeneration. 6. This invention improves the efficiency of decision support report generation and system scalability: Through standardized data transfer between multiple agents, a continuous processing flow is formed from user problem analysis and professional information retrieval to report generation, evaluation, and manual review, reducing the workload of manually understanding problems item by item, searching for information across sources, and organizing citations; by expanding scenario samples, updating the public information knowledge base, and configuring evaluation thresholds, it can be adapted to different platform business scenarios.

[0017] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0018] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0019] Figure 1 This is a flowchart of a decision-making method based on structured reasoning chain fine-tuning and multi-agent collaboration in an embodiment of the present invention; Figure 2 For each user dialogue sample in the cold start dialogue dataset in this embodiment of the invention, a corresponding three-dimensional structured link flowchart is generated; Figure 3 This is a flowchart of the online inference process in an embodiment of the present invention; Figure 4 This is a closed-loop diagram of report evaluation and feedback reprocessing in an embodiment of the present invention; Figure 5 This is a flowchart illustrating the core processing steps of the ride-hailing access rule analysis task in this embodiment of the invention. Detailed Implementation

[0020] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0021] This invention is applicable to user intent parsing, public information retrieval, and decision support report generation in complex rule scenarios such as platform business management, compliance review, and business judgment.

[0022] This invention proposes a decision-making method and system based on structured reasoning chain fine-tuning and multi-agent collaboration. Through cold-start data construction, three-dimensional structured link supervision fine-tuning, hybrid retrieval of publicly available information, and a report generation and evaluation closed loop, it forms an integrated processing architecture from user dialogue input to traceable decision support report output. The innovation lies in: (1) Construct a three-dimensional structured link of "problem dimension - scenario dimension - basis dimension" and its supervision and fine-tuning mechanism. Transform unstructured single-turn or multi-turn user dialogues into dialogues based on the problem dimension. Scene dimension and based on The standardized intermediate data objects are composed of dimensions that only describe the direction of the publicly available information to be retrieved. The cold-start dialogue samples and corresponding links are used to form supervised data to perform supervised fine-tuning of the base language model. The fine-tuned first model and the unfine-tuned base language model are evaluated under the same conditions by the three-dimensional structure integrity rate, the accuracy of the question dimension and the scenario dimension fields, and the requirement description compliance rate, thereby improving the stability and parsability of the generated links.

[0023] (2) Construct a traceable hybrid retrieval mechanism driven by three-dimensional structured links. Based on the three-dimensional structured link fields, weighted keyword queries and semantic vector queries are directly generated. BM25 keyword recall and HNSW vector semantic recall are performed in the same fragment retrieval index, respectively. The inverse ranking fusion algorithm is used for unified sorting. Combined with data type filtering, source integrity check, text deduplication, semantic near-duplicate filtering and homologous fragment quantity control, a professional context containing fragment ID (chunk_id), document ID (doc_id), fragment text and source location information is generated to take into account the accurate matching of professional terms, the recall of synonyms, and the backtracking of the original text.

[0024] (3) Construct a multi-agent collaborative mechanism that combines report generation, quantitative evaluation, and targeted feedback. The three-dimensional link generation, public information retrieval, report generation, and report evaluation feedback are configured as mutually collaborative processing agents. The three-dimensional structured link, professional context, report to be evaluated, and evaluation results are used as standardized transmission objects between agents. The report generation agent only uses fragments in the professional context to form external facts and professional evidence. The evaluation feedback agent checks the content completeness rate, professional evidence coverage rate, and citation traceability rate. Based on the reasons for the score loss, it is directed back to the report generation, information retrieval, or link generation stage. After reaching the preset number of feedbacks, it is transferred to manual review, thereby forming a controllable and locationable closed loop for report quality optimization.

[0025] A specific embodiment of the present invention discloses a decision-making method based on structured inference chain fine-tuning and multi-agent collaboration, such as... Figure 1 As shown, it includes the following steps: Step S1: Generate a corresponding 3D structured link based on each user dialogue sample in the cold start dialogue dataset; pair each user dialogue sample with the corresponding 3D structured link to construct a supervised training sample set; Step S2: The base large language model is fine-tuned under supervision using the supervised training sample set to obtain the first model; wherein, the first model is used to generate the corresponding three-dimensional structured link in real time based on the user dialogue of the input natural language. Step S3: Using a pre-built public knowledge base, based on the current user dialogue and the three-dimensional structured link generated in real time by the first model, weighted keyword query and semantic vector query are generated respectively. BM25 keyword recall and HNSW vector semantic recall are executed in parallel. Based on the two recall results, the corresponding professional context is generated. Step S4: Based on the current user dialogue, three-dimensional structured link, and professional context, a multi-agent collaborative decision support report is generated; the decision support report is quantitatively evaluated; if the evaluation result does not reach the evaluation threshold, the process returns to steps S2, S3, or S4 for reprocessing until the preset conditions are met to obtain the final decision support report.

[0026] Step S1 includes steps S11-S13.

[0027] This step addresses the lack of historical question-and-answer data from real business systems, specifically the absence of real user historical question-and-answer data and corresponding labeled samples, which prevents the model from undergoing direct supervised training. It constructs cold-start dialogue data for platform business management, compliance review, and business judgment scenarios, and further extracts a three-dimensional structured link—"question dimension - scenario dimension - basis dimension"—from this cold-start dialogue data. For example, the cold-start dialogue data in this invention mainly covers two high-frequency scenarios: ride-hailing and food delivery services, and the data formats include single-turn and multi-turn dialogues.

[0028] Step S11: Construct a cold start dialogue dataset.

[0029] First, we construct a set of platform business scenarios. This set includes, but is not limited to, ride-hailing and food delivery business scenarios. The ride-hailing business scenarios include sub-scenarios such as pricing adjustments, subsidy strategies, complaint handling, and supply and demand scheduling. The food delivery business scenarios include sub-scenarios such as delivery delays, merchant commissions, subsidy activities, and handling user complaints.

[0030] For each business sub-scenario, a corresponding set of user roles is further pre-defined. User roles include, but are not limited to, platform operations personnel, compliance management personnel, rule review personnel, risk control analysts, and business judgment personnel. These user roles are used to simulate actual business personnel who raise management, analysis, or judgment needs in the platform's business system.

[0031] During the cold start phase, based on the set of business scenarios and user roles, simulated dialogue samples proposed by users in the platform's business system are generated. Each cold start dialogue sample includes at least the user role, application scenario, dialogue type, user question, expected goal, and reference constraint information. Among them, the dialogue type includes single-turn dialogue and multi-turn dialogue; the expected goal is used to characterize the judgment conclusion, risk analysis, solution suggestion, or handling suggestion that the user hopes to obtain; the reference constraint information is used to characterize the business rules, compliance requirements, or analysis boundaries that need to be followed when generating the report.

[0032] For example, 1000 cold start dialogue samples are constructed, mainly covering ride-hailing and food delivery business scenarios; among them, single-turn dialogue samples account for approximately 30%, and multi-turn dialogue samples account for approximately 70%. The above sample quantity, business scenario type, and single / multi-turn dialogue ratio are only examples, and can be flexibly adjusted according to the target platform's business scope and the training requirements of the base large language model in actual applications.

[0033] Step S12: Based on each user dialogue sample in the cold start dialogue dataset, generate the corresponding three-dimensional structured link. For example... Figure 2 As shown.

[0034] Based on each user dialogue sample in the cold start dialogue dataset, a corresponding three-dimensional structured link is generated, including: Based on single-turn or multi-turn dialogues, and utilizing pre-defined link generation template constraints, each user dialogue sample is subjected to question identification, scenario identification, and requirement description identification to obtain the corresponding three-dimensional structured link. Among them, the user dialogue sample includes user questions, supplementary conditions, historical follow-up questions, business constraint information, and expected output targets. The three-dimensional structured link includes problem-dimensional information, scene-dimensional information, and basis-dimensional information; The issue dimension information includes the user's issue type, target result, set of key sub-issues, and issue priority: The scenario dimension information includes scenario type, questioning user role, set of relevant participants, analysis object, and set of constraints. The information provided is a description of the requirements and is used to indicate the direction of information to be focused on in subsequent public information retrieval. It does not indicate specific evidence content, does not limit specific information sources, and does not include internal business data of the platform.

[0035] After constructing the cold start dialogue dataset, the cold start user dialogue samples are subjected to structured parsing to generate three-dimensional structured links corresponding to user input. These three-dimensional structured links are used to convert unstructured user dialogues into trainable, searchable, and verifiable intermediate data objects, providing standardized input for subsequent model training, data retrieval, and report generation.

[0036] Specifically, user single-turn or multi-turn dialogues are used as input. The user dialogue includes user questions, supplementary conditions, historical follow-up questions, business constraints, and expected output targets. A pre-defined link generation template constrains the link generation model, performing question identification, scenario classification, object identification, constraint extraction, and requirement-based identification on the user dialogue to obtain a three-dimensional structured link. . It is the root of the entire data object, like a container that contains the three core dimensions: problem dimension, scenario dimension, and basis dimension, representing a complete breakdown of the user's problem.

[0037] 3D structured link Includes: Problem dimension information Scene dimension information and based on dimensional information As shown below: Formula (1); in, This represents the three-dimensional structured link corresponding to user dialogue D; Indicates the problem dimension information; Represents scene dimension information; This indicates information based on dimensions.

[0038] Through the above structure, an intermediate link with a unified field format is formed before generating the decision support report, so that the subsequent processing has clear problem objectives, business boundaries and basis requirements, thereby reducing the risk of generating loose answers based solely on keyword recall.

[0039] (1) Problem Dimension Information : Problem Dimension Information This is used to characterize the type of problem, goal, and sub-problems that the user actually needs to solve, rather than just extracting keywords from the user input, as shown below: Formula (2); in, This indicates the problem type, used to identify the nature of the user's problem and its possible values ​​include feasibility assessment, risk identification, solution selection, root cause analysis, and implementation recommendations. The system determines that this is not merely a simple "yes / no" question, but a task requiring a comprehensive analysis of potential risks. This represents the target outcome, indicating the type of result the user expects the system to output. Values ​​include determining feasibility, selecting a recommended solution, identifying major risks, formulating handling suggestions, and generating an analysis report. This is the ultimate goal, and all retrieval and generation work revolves around it. This represents a set of key sub-questions, used to break down the user's original question into at least one answerable sub-question. The large goal is broken down into specific, sub-questions that must be answered; the report must address each of these questions individually, otherwise it will be considered "incomplete." This indicates the priority of the question and is used to distinguish between the main question and the subordinate questions. In a multi-turn dialogue, if the user adds content later that changes the original goal, the goal result that was last explicitly stated will be used as the main question.

[0040] (2) Scene dimension information : Scene Dimension Information This is used to characterize the business environment, the main object, and the constraints in which the problem occurs. As shown below: Formula (3); in, This indicates the scenario type, including ride-hailing business scenarios, food delivery business scenarios, and other platform business scenarios, to clarify that this is a problem in a specific business area.

[0041] Indicate the role of the user asking the question. For example, user roles include platform operations personnel, compliance management personnel, risk control personnel, and business analysts. Clearly defining the user role helps determine the focus of the answer. This refers to the set of relevant participants. For example, in the ride-hailing business scenario, the set of relevant participants includes the platform, drivers, and passengers; in the food delivery business scenario, the set of relevant participants includes the platform, delivery personnel, merchants, and users. The analysis object refers to the object of analysis. For example, the analysis object may include admission rules, order dispatch rules, subsidy schemes, commission rules, delivery delay rates, complaint events, etc. The analysis object specifically indicates what the analysis objective is. This field represents a set of constraints, including time, location, budget, etc. If the user does not explicitly specify a constraint, this field is marked as "not specified" and cannot be added manually.

[0042] (3) Based on dimensional information : Based on dimensional information This tool is used to extract publicly available information requirements related to report generation from user conversations. These requirements are linked to the user's question, the context, constraints, and the desired output, providing directional input for subsequent publicly available information retrieval.

[0043] Based on dimensional information This refers only to publicly available information, does not represent specific evidence content, does not limit the specific source of information, and does not include internal platform business data. This step does not perform actual data retrieval; instead, it generates directional descriptions required for subsequent data retrieval.

[0044] Based on dimensional information It serves as a standardized bridge connecting semantic understanding and information retrieval; it translates "what the user wants" into "what direction the system should look for in publicly available information," but it doesn't make decisions for the retrieval system or fabricate evidence for the report generator. This design of "only describing needs, not pre-setting answers" is one of the key cornerstones of the entire technical solution's high stability and scalability.

[0045] Based on dimensional information As shown below: Formula (4); in, This indicates the direction of publicly available information that should be focused on during subsequent data retrieval, based on the requirements description.

[0046] 3D Structured Reasoning Link The data format is instantiated and displayed as follows: " ": { " ": { "Qtype": "Risk Identification", "Gtarget": "Identify risks associated with driver access rule adjustments, assess the feasibility of online training as a supplementary condition, and formulate implementation recommendations." "Ksub": [ What are the potential implementation or compliance risks arising from the existing driver admission rules and proposed adjustments? "Whether online training results are suitable as a supplementary basis for judgment" "What verification, appeal, and manual review measures need to be set up before the rules go live?" What publicly available information should support the key judgments in the report? ], "Qpriority": "Main Question" }, " ": { "Stype": "Ride-hailing service scenario", "Urole": "Compliance Manager", "Aparties": ["Platform", "Driver", "Passenger", "Relevant Management Departments"], "Otarget": "City A's Ride-Hailing Driver Admission Rules and Online Training-Assisted Judgment Mechanism", "Cconstraint": { "time_limit": "Report to be generated within three days", "region": "City A", "analysis_boundary": "Only publicly available information is used; driver personal information and platform internal operational data are not accessed." "decision_boundary": "Online training results are only used as a supplementary judgment condition and do not determine the admission result alone." "data_period": "Undefined" } }, " ": { "Edesc": "It is necessary to retrieve publicly available information related to the access requirements for ride-hailing drivers in City A, the applicable conditions for online training, platform service quality and safety management requirements, appeal and manual review mechanisms, and similar platform governance cases to support risk identification, feasibility assessment, and implementation recommendations." } } } Among them, time_limit is the time constraint; region is the geographical constraint; analysis_boundary is the data source boundary constraint; decision_boundary is the decision power boundary constraint; data_period is not specified and should be clearly marked as missing information to prevent the model from guessing on its own.

[0047] The preset link generation template constraints include top-level structure constraints, field-level format constraints, content boundary constraints, and serialization format constraints; The top-level structural constraints are used to force the three-dimensional structured link to include three dimensions: problem dimension information, scene dimension information, and basis dimension information, and the order of these dimensions must be fixed. No dimension may be missing, and no undefined additional fields may be added. The field-level format constraints are used to specify the data type and value specifications for each field in the problem dimension information, scenario dimension information, and basis dimension information. The content boundary constraints are used to prohibit the basis dimension information from containing specific evidence content, and to prohibit the self-creation of constraints in the constraint condition set of the scenario dimension information when the user has not explicitly given constraints. The serialization format constraint is used to require the three-dimensional structured link to be serialized into a parsable data format in a fixed field order.

[0048] Link generation template constraints are used to ensure that the output of the link generation model is a three-dimensional structured link. It features a unified format, complete fields, and a structure that can be parsed downstream; it is a predefined set of data generation specifications and format requirements. It specifies the format, field hierarchy, data type, and content boundaries that the link generation model must follow.

[0049] (1) Top-level structural constraints: The link generation model is required to include three primary dimensions in a fixed order; no dimension may be missing, and no undefined additional fields may be added.

[0050] (2) Field-level format constraints: Specify the data type and value specification for each field in the three dimensions.

[0051] (3) Content boundary constraints: The content that the link generation model must not generate is clearly specified; According to wei The document must not include specific evidence. Scene dimension constraints If the user does not provide a certain constraint, it must be marked as "not specified" and the model must not be created automatically.

[0052] (4) Serialization format constraints: In training data construction and subsequent online inference, the three-dimensional structured links output by the first model must be serialized in a fixed order into a parsable format (such as JSON format) to ensure that downstream retrieval and generation of decision support reports can be directly read and parsed without additional semantic understanding.

[0053] Step S13: Pair each user dialogue sample with the corresponding three-dimensional structured link to construct a supervised training sample set.

[0054] Each user dialogue sample is paired with its corresponding 3D structured link to construct a supervised training sample set, including: Each supervised training sample includes input data and output data; The input data is a single-turn or multi-turn dialogue between users; the output data is a three-dimensional structured link corresponding to the input data. All supervised training samples are divided into a training set and a test set according to a preset ratio; the training set is used to update model parameters during the training of the first model, and the test set is used to evaluate the performance after the first model is trained.

[0055] Cold-start dialogue samples are paired with corresponding 3D structured links to form supervised training samples. Each supervised training sample includes input data and output data.

[0056] The input data is the context of a single-turn or multi-turn dialogue between the user, denoted as: Formula (5); in, Indicates the first A sample dialogue; Indicates the first The number of dialogue rounds contained in each dialogue sample; This represents the r-th round of dialogue text in the i-th sample. When the sample is a single-turn dialogue, .

[0057] The output data is a 3D structured link corresponding to the input data, as shown below: Formula (6); in, Indicates the first The three-dimensional structured link corresponding to each sample; Represents the problem dimension data; Represents scene-dimensional data; This indicates that the data is based on dimensional data.

[0058] Problem Dimension The elements and problem dimension information in step S12 Maintain consistency in the definitions used to record problem types, target results, sets of key sub-problems, and problem priorities. As shown below: Formula (7); in, Indicates the first The question types of the sample; Indicates the first The target result for each sample; Indicates the first The set of key sub-problems for each sample; Indicates the first Prioritize the issues for each sample.

[0059] Scene dimension The elements are consistent with the scenario dimension definition in step S1, used to record the scenario type, the user role asking the question, the set of relevant participants, the analysis object, and the set of constraints. The field format of the scenario dimension is: Formula (8); in, Indicates the first The scenario type of each sample; Indicates the first The user roles that asked the questions in the sample; Indicates the first The set of relevant participants in each sample; Indicates the first The analysis object of the sample; Indicates the first The set of constraints for each sample. Each element in the problem dimension and scenario dimension corresponds one-to-one with the element of the same name in step S1.

[0060] Based on dimensional information This is used to extract publicly available information requirements descriptions related to report generation from user dialogue content. These descriptions are related to the user's question, the context, constraints, and the desired output, providing directional input for subsequent publicly available information retrieval.

[0061] In this step, the dimension information does not perform actual data retrieval; instead, it generates directional descriptions required for subsequent data retrieval. The dimension is defined as follows: Formula (9); in, Indicates the first The sample is based on a requirement description, which describes the information direction that should be focused on in subsequent public information retrieval.

[0062] Three-dimensional structured link Serialize to target output according to a fixed field order. This results in a supervised dataset, as shown below: Formula (10); in, This represents the dataset consisting of all supervised samples; Indicates the first A sample dialogue; Indicates the total number of samples; Indicates the first The standardized link output corresponding to each dialogue sample.

[0063] The dataset was divided in an 8:2 ratio. Divided into training set and test set The training set is used to update model parameters, while the test set is only used for performance evaluation after training is completed.

[0064] Step S1 aims to construct a supervised training sample set for training a large language model to generate standardized inference links by transforming manually constructed user dialogue samples into a unified format of a three-dimensional structured link consisting of "question dimension - scenario dimension - basis dimension" under cold-start conditions where real historical question-and-answer data is lacking.

[0065] Step S2, specifically.

[0066] This step is used to convert the cold start dialogue samples obtained in step S1 into supervised training data. For example, supervised fine-tuning of the Qwen3.5-9B base large language model is performed, and the fine-tuned model is verified to be able to stably generate a three-dimensional structured link composed of the question dimension, the scene dimension and the basis dimension based on the user input content.

[0067] The base large language model is fine-tuned under supervision using the aforementioned supervised training sample set to obtain a first model, including: Using the samples in the training set as training data, the input data as input to the base big language model, and the output data as the supervised target output of the base big language model, the base big language model learns the mapping relationship from user dialogue in natural language to standardized link data. The output of the supervision target is converted into a tag sequence, which consists of multiple tags formed by sequentially splitting the output of the supervision target; Using the cross-entropy loss function, for each labeled position in the supervised target output, the probability that the model predicts that position is correctly labeled is calculated. The predicted probabilities of all labeled positions are summed and the negative logarithm is taken to obtain the loss value of the training sample. The loss value is used to quantify the difference between the model prediction result and the standard output. Based on the loss value, the parameters of the base large language model are updated through backpropagation. The training is iterated with the goal of minimizing the loss value. After completing a preset number of training rounds, the model parameters are saved to obtain the first model.

[0068] With training set The formatted samples in the dataset were used as training data to perform supervised fine-tuning of the Qwen3.5-9B-based large language model. (Based on user dialogue...) As input to the model, a three-dimensional structured link As a supervised output, it enables the model to learn the mapping relationship from natural language dialogue to standardized link data.

[0069] The first model is obtained after supervised fine-tuning. .

[0070] For the Training samples, target output Convert to a labeled sequence, as shown below: Formula (11); in, Indicates the first The number of tags contained in the target output; Indicates the first The first of the target outputs A tag.

[0071] A token is the smallest basic unit when a large language model processes text.

[0072] A token doesn't necessarily have to be a complete word; it can be: ① A complete word (e.g., "platform" → 1 token); ② a subword or root (for example, "unhappy" may be split into ["un", "happy"] → 2 Tokens); ③ a single character (for example, "网", "约", "车" → each is 1 Token); ④ a punctuation mark (for example, ",", "。" → each is 1 Token); ⑤ a number (for example, "2026" → may be 1 or 4 Tokens, depending on the tokenization method).

[0073] Target output The conversion to a token sequence is completed using a tokenizer, with the following steps: Step 1: The target output , which is a structured JSON string; Step 2: Normalization, unified encoding such as UTF-8, and cleaning of invisible characters; Step 3: Pre-tokenization, rough splitting by spaces, punctuation marks, etc.; Step 4: Subword tokenization, using a trained vocabulary (such as the BPE algorithm) to split words into subwords present in the vocabulary; Step 5: Mapping to numeric IDs, each Token is mapped to a unique numeric ID in the vocabulary. In this application, it generally refers to the Token itself or its corresponding ID for simplified representation; Step 6: Output an ordered Token sequence .

[0074] The model training objective is to minimize the cross-entropy loss of the three-dimensional structured link. The cross-entropy loss function is shown as follows: Equation (11); Wherein, represents the supervised fine-tuning loss; represents the parameters involved in updating in the first model ; represents the number of samples in the training set; represents the conditional generation probability determined by the parameters ; represents the token sequence before the token at the position in the target output of the th sample; the meanings of the remaining symbols are the same as the aforementioned definitions.

[0075] During the training process, the input dialogue and the standard output are simultaneously fed into the model, the difference between the prediction result and the standard output is calculated through the aforementioned cross-entropy loss function to obtain a loss value, and then the parameters are updated according to the loss value After completing the preset number of training rounds, save the model parameters to obtain the first model for online inference. .

[0076] ① Position-by-position comparison: For each token position t in the target output, calculate a probability distribution. This indicates which token it believes is "most appropriate" to generate at that location; ② Comparison with the standard answer: The system generates a probability distribution for this condition and compares it with the actual token at that position in the standard answer. Compare; ③ Calculate cross-entropy: If the model predicts the correct token with a higher probability, the loss contribution at that position will be smaller and closer to 0; conversely, if the model gives a very low probability of the correct token, the loss contribution at that position will be large. ④ Summarize and average: Sum the losses of all locations and all samples, then take the average to obtain the final loss value. .

[0077] test set It does not participate in parameter updates. After training, the first model is used respectively. Inference was performed on the test set using the unsupervised fine-tuned Qwen3.5-9B pedestal large language model.

[0078] Transforming a general-purpose, foundational large language model into a dedicated first-order model through supervised fine-tuning. This enables the generation of high-quality, highly stable data from user dialogue to a standardized three-dimensional chain, providing accurate and parsable "task instructions" for subsequent hybrid retrieval and report generation.

[0079] The trained first model Deployed to an online environment, it receives real user conversations and generates and verifies three-dimensional structured links in real time. .

[0080] During online inference, the system receives raw single-turn or multi-turn dialogue text submitted by the user through the interactive interface and processes it according to the same input format as the training data. The processed dialogue input is then used by the first model. The three-dimensional structured link is directly generated from the first model: Formula (12); in, This indicates user conversations received online; This represents the three-dimensional structured link generated by the first model; , , These represent the problem dimension, scenario dimension, and basis dimension of the model generation, respectively. and The elements contained therein are the same as the problem dimension information in step S12. and scene dimension information Elements maintain a one-to-one correspondence.

[0081] Online reasoning process, such as Figure 3 As shown.

[0082] Online validation only checks the model output and performs structured validation, checking only the following three items: ① Can it be parsed into a specified data format, such as JSON; ② Do the three dimensions of problem dimension, scenario dimension, and basis dimension exist? ③According to the requirements of the maintenance department The format, and Not empty.

[0083] After verification, the system outputs a three-dimensional structured link; the accuracy and conformity of the generated content are evaluated using a test set.

[0084] If verification fails, an error message will be returned or the function will be regenerated (within the preset number of retries).

[0085] Online verification only checks the format, not the accuracy of the content.

[0086] The process involves deploying the trained first model to an online environment, performing real-time inference on real user dialogues, and generating a three-dimensional structured link. The online validation only performs a "threshold check" on format and completeness to ensure that the output can be parsed by downstream systems, while the accuracy of the content is separately checked by offline evaluation and report assessment closed loop, taking into account both the requirements of real-time response and final quality.

[0087] First Model Evaluation vs. Untuned Baseline Large Language Model: Using the test set The first model after supervision and fine-tuning and the untuned base large language model The evaluation was conducted. Two sets of comparison models were set up: The first group consists of the first model after supervised fine-tuning. ; The second group consists of unsupervised fine-tuning base models that generate 3D structured links solely through fixed prompts. .

[0088] Both models use the same test set, input fields (same dialogue format), output format, and generation parameters.

[0089] First Model The output is a standard three-dimensional structured link; the base is a large language model. The output of a standard three-dimensional structured link is required by fixed prompt words.

[0090] This design ensures that the differences between the two sets of models are attributed to "whether or not supervised fine-tuning has been performed", thus objectively demonstrating the technical effects of fine-tuning.

[0091] To verify whether the first model's generated results meet the final output requirements, the following three evaluation metrics are used: 3D structure integrity rate, accuracy of problem dimension and scene dimension fields, and compliance rate with requirement description.

[0092] (1) Three-dimensional structural integrity rate : The 3D structural integrity rate is used to measure whether the model output simultaneously contains the problem dimension, scene dimension, and basis dimension, and whether the required fields in the three dimensions can be parsed normally.

[0093] The problem dimension should include the problem type, target outcome, set of key sub-problems, and problem priority; the scenario dimension should include the scenario type, user role, set of relevant stakeholders, analysis object, and set of constraints; the basis dimension should conform to... The format, and Not empty; the calculation of the three-dimensional structure integrity rate is as follows: Formula (13); in, Indicates the integrity rate of the three-dimensional structure; This indicates the number of test samples whose three-dimensional structure is complete and can be correctly resolved. This represents the total number of samples in the test set.

[0094] For example, the integrity rate of three-dimensional structures If the structural integrity rate is not lower than the three-dimensional structural integrity rate threshold, the first model is deemed to have met the requirements in terms of structural integrity. For example, the three-dimensional structural integrity rate threshold is 0.90; in specific applications, the three-dimensional structural integrity rate threshold can be flexibly adjusted according to specific needs.

[0095] (2) Accuracy of Problem Dimension and Scenario Dimension Fields : The accuracy of the question dimension and scenario dimension fields measures the degree of consistency between the model-generated question type, target result, set of key sub-questions, question priority, scenario type, questioning user role, set of relevant stakeholders, set of analysis objects, and set of constraints and the standard results of the test set. The accuracy of the question dimension and scenario dimension fields is calculated as follows: Formula (14); in, Indicates the accuracy of the problem dimension and scenario dimension fields; This indicates the number of problem-dimensional and scene-dimensional elements whose generated values ​​match the standard values. This represents the total number of question-dimensional elements and scenario-dimensional elements participating in the evaluation of the test set. In one embodiment, When the accuracy of each element in the problem dimension and scenario dimension of the model is not lower than the accuracy threshold of the problem dimension and scenario dimension fields, it is determined that the generation accuracy of each element in the problem dimension and scenario dimension of the model meets the requirements.

[0096] For example, the accuracy threshold for the problem dimension and scenario dimension fields is 0.90; in specific applications, the accuracy threshold for the problem dimension and scenario dimension fields can be flexibly adjusted according to specific needs.

[0097] (3) Based on the requirement description compliance rate : The requirement description compliance rate is used to measure whether the basis dimension generated by the model conforms to the requirements description specified in the publicly available information. For any test sample, the basis dimension of the sample is considered to meet the requirements when the generated result simultaneously meets the following conditions: the data format of the basis dimension is as follows: Not empty; Related to user questions, scenario objects, constraints, or expected output targets; This only describes the areas to focus on when searching for publicly available information. Based on the requirement description's compliance rate, the calculation is as follows: Formula (15); in, This indicates the compliance rate based on the requirements description; This indicates the number of test samples that simultaneously meet the above conditions based on the dimension. This represents the total number of samples in the test set. In one embodiment, If the required description compliance rate is not lower than the threshold, the required description generated by the model is considered to meet the requirements. For example, the required description compliance rate threshold is set to 0.85. In specific applications, the required description compliance rate threshold can be flexibly adjusted according to specific needs.

[0098] The progressive relationship among the three indicators: 3D structure integrity rate, format threshold; Accuracy of question dimension and scenario dimension fields, and content precision; Based on the requirement compliance rate and semantic standardization.

[0099] (4) Calculation of the increase before and after fine-tuning: Calculate the first model separately and the base large language model The scores are based on the three indicators mentioned above: the completeness rate of the three-dimensional structure, the accuracy rate of the problem dimension and scenario dimension fields, and the compliance rate with requirements.

[0100] For any indicator k, its absolute increase is as follows: Formula (16); The relative improvement rate is calculated as follows: Formula (17); in, These correspond to the three-dimensional structure integrity rate, the accuracy rate of the problem dimension and scenario dimension fields, and the compliance rate with the requirement description, respectively. This represents the score of the first model on the k-th metric; This represents the score of the base large language model on the k-th metric; Indicates the absolute increase; Indicates the relative improvement rate; This represents a positive constant to prevent the denominator from being zero.

[0101] The comparison results between the first model and the base large language model are recorded in Table 1.

[0102] Table 1 - Comparison of the first model and the base large language model:

[0103] When the absolute improvement of the first model in all three metrics is greater than 0, and each of the three metrics meets its corresponding threshold, the supervised fine-tuning first model is deemed to be able to reliably convert user dialogue into a three-dimensional structured link conforming to the specified format, and thus the trained first model is obtained. .

[0104] Step S2 is used to perform supervised fine-tuning of the base large language model using the supervised training sample set constructed in step S1, so as to obtain the first model, which enables it to obtain the ability to stably generate three-dimensional structured links that meet the format and content requirements based on the user's natural language dialogue.

[0105] Step S3 involves constructing a public information knowledge base and a hybrid retrieval system, including steps S31-S32.

[0106] This step is used to retrieve professional information related to the current problem from a pre-built public knowledge base based on the user dialogue content and the three-dimensional structured link generated in step S2, and to encapsulate the retrieval results into professional context with source location information for use in subsequent decision support report generation steps.

[0107] The input received in this step As shown below: Formula (18); in, This indicates whether the user is engaging in a single-turn or multi-turn dialogue, and provides user input. After the first model The generated three-dimensional structured link, and .

[0108] This step includes two stages: offline document preprocessing and online hybrid retrieval. The offline document preprocessing stage performs tasks such as obtaining publicly available information, format recognition, text extraction, fragment segmentation, metadata generation, vector generation, and index construction. Online hybrid search phase based on and Generate keyword queries and semantic vector queries, perform BM25 keyword recall and semantic vector recall respectively, and then sort, filter and encapsulate professional context through the inverse ranking fusion algorithm.

[0109] Step S31: Pre-build a knowledge base of publicly available information.

[0110] Pre-built public knowledge bases are obtained through the following methods: Publicly available information is obtained through automatic downloading, API call extraction, or manual import; the publicly available information includes at least one or more of the following: academic papers, professional textbooks, publicly available professional cases, laws and regulations, and publicly available rule documents. For each piece of publicly available information obtained, a unique document identifier is generated, and the document title, author, publishing organization, information type, publication date, source address, acquisition time, and original file hash value are recorded; when multiple original files have the same hash value, the one with the earliest acquisition time is retained; The publicly available information is processed sequentially to perform document format recognition and text extraction to obtain the document body. The document body is segmented into segments, with each document divided into several document segments, and the text of each segment controlled within a preset word count range; Generate a unique segment identifier for each document segment, and record its document identifier, author, publishing organization, page number, document type and source address to form a standardized segment text; The standardized fragment text is encoded using a pre-trained Chinese text embedding model, and fragment vectors are obtained after L2 normalization. A keyword inverted index of the standardized fragment text and a vector index of the fragment vector are established in the fragment retrieval index. The keyword index and the vector index point to the same fragment through the same fragment identifier.

[0111] For example, the preset character count is 800 Chinese characters.

[0112] (1) Obtaining publicly available information: Information related to platform business governance, platform operational decisions, and rule constraints is obtained from publicly accessible and verifiable websites, public data interfaces, or manually imported channels. The obtained information does not include internal platform business data and mainly includes the following four categories: The first category consists of academic papers, specifically gathering research findings, analytical methods, model parameters, applicable conditions, and publicly available experimental data in areas such as platform economy, two-sided markets, pricing mechanisms, incentive mechanisms, competition analysis, and platform governance. This type of material is used to provide research support for problem analysis, method selection, and reasoning in decision support reports. The second category consists of professional textbooks, specifically covering concepts, definitions, analytical frameworks, calculation methods, and applicable prerequisites in fields such as platform economics, industrial organization theory, microeconomics, game theory, and mechanism design. This type of material provides the theoretical foundation for the analytical logic, indicator meanings, and applicable conditions of the methods in decision support reports. The third category consists of publicly available professional case studies, specifically governance cases and outcomes published by industry and research institutions. This type of information is used to provide real-world case support for decision-making reports. The fourth category consists of laws, regulations, and publicly available rules. This involves obtaining relevant legal provisions and publicly available platform rules related to platform governance, and retaining information on applicable subjects, applicable conditions, and their validity status. This type of information is used to identify rule constraints, compliance boundaries, and handling requirements in decision support reports.

[0113] Obtain the above information according to the pre-defined list of publicly available sources. Methods of obtaining this information include: ①Automatic download: For materials available for download on the official website, the system downloads the original file; ② Interface call extraction: For documents that provide public data interfaces, the system calls the corresponding interfaces to obtain the document body and metadata; ③ Manual import: For materials that cannot be downloaded automatically but are allowed to be used publicly, operators will manually import them.

[0114] A unique document identifier is generated for each document obtained. It records the document title, author or publishing organization, document type, publication date, source address, retrieval time, and original file hash value. For example, the original file hash value is SHA-256.

[0115] The original file hash value is used to identify duplicate files with identical content; when two files have the same hash value, only the one with complete source information and earlier acquisition time is retained as the original document, ensuring that there is no redundancy in the knowledge base.

[0116] (2) Document format recognition, text extraction and segmentation: Transform raw documents of various formats into plain text fragments that are well-structured, of moderate length, and semantically complete.

[0117] First, the document format is identified based on the file header characteristics, MIME (Multipurpose Internet Mail Extensions Type) type, and file extension. Then, the parsing program corresponding to that format is called to perform text extraction.

[0118] ① For PDF files containing text layers, extract the text and its page numbers according to the page order; ② For Word documents, extract content according to the order in which headings, body paragraphs, lists, and tables are arranged in the document; ③ For web page files, parse the web page document object model, retain the main text title, main text paragraphs and list content, and delete non-main text nodes such as menus, navigation bars, advertisements and footnote navigation; ④ For plain text files, convert them to UTF-8 encoding before reading.

[0119] ⑤ For scanned PDFs or image documents without a text layer, the system performs OCR text recognition and restores the page numbers and paragraph order based on the page numbers and text area positions returned by the OCR.

[0120] If no text is extracted from a PDF page, or if fewer than 20 valid characters are extracted, the page is identified as the page to be OCR and OCR recognition is performed.

[0121] After text extraction is complete, the system performs text cleaning, removing garbled characters, invisible control characters, consecutive whitespace, and repeated navigation text; while retaining heading levels, paragraphs, page numbers, chapter names, clause numbers, and list order as subsequent segment location information.

[0122] Text segmentation is performed in a fixed order. The core goal of segmentation is to divide long documents into semantically independent, moderately long segments (300-800 words) while preserving their respective themes. Segmentation is performed in the following order: The first step is to divide the entire document into chapter-level text units, using first-level headings, second-level headings, chapter names, or legal clause numbers as boundaries; this maintains semantic integrity. The second step is to further divide each chapter-level text unit into natural paragraphs and attach the current chapter title to the beginning of the main text of each paragraph within that chapter, so that the paragraph can retain its theme even after being separated from the original text. The third step is to merge short paragraphs. For short paragraphs with fewer than 300 Chinese characters after being segmented, merge them sequentially with the next paragraph, provided that the chapter and page numbers are consecutive, until the merged result reaches 300 Chinese characters or has reached the end of the current chapter; to avoid fragmented segments. The fourth step is to segment long paragraphs with more than 800 Chinese characters sequentially using the complete sentence boundaries indicated by periods, question marks, exclamation marks, and semicolons. If a single sentence still exceeds 800 Chinese characters, it is then segmented according to commas and colons to avoid excessively long segments and ensure granularity of retrieval. The fifth step involves overlapping content. Between adjacent segments formed by further splitting the long text, the last 50 to 100 Chinese characters of the previous segment are retained as overlapping content. The actual overlap length is determined based on the principle of not splitting complete sentences. This ensures semantic coherence and prevents the segmentation from cutting off key contextual information.

[0123] After the above sequential processing, the main content of each segment is controlled within the range of 300 to 800 Chinese characters.

[0124] (3) Document fragment standardization and metadata generation: Objective: To generate a unique identifier for each fragment and bind it with complete source metadata.

[0125] Standardize each document fragment. For the first... For each document fragment, the system generates a unique fragment identifier based on the document ID, chapter name, page number, and fragment start character offset. This generates standardized document fragment data, as shown below: Formula (19); in, Indicates a unique identifier for the segment; ; Indicates the author; if there is no explicit author, take a null value. Indicates the issuing organization; Indicates the page number or the start and end page numbers; if there is no fixed page number, take an empty value. This indicates the type of material, such as papers, textbooks, professional data, case studies, or legal rules. Indicates the source address; Fragment metadata only stores "location information" (i.e., how to find this fragment), and does not store "document-level information" such as document title and publication date. This information is uniformly stored in the document source index and associated through doc_id. This avoids data redundancy and inconsistency.

[0126] (4) Document vector generation and retrieval index construction: Objective: To convert fragmented text into semantic vectors and build a dual index in Elasticsearch that supports both keyword and vector retrieval.

[0127] The main text of the excerpt is recorded as The generated corresponding metadata is denoted as We used a BAAI / bge-base-zh-v1.5 pre-trained Chinese text embedding model (768 dimensions) to embed each segment of the main text. Encode it. BAAI / bge-base-zh-v1.5 is an open-source Chinese text embedding model.

[0128] The BAAI / bge-base-zh-v1.5 model uses a Transformer encoding network to convert the entire text into a 768-dimensional dense vector, which is then normalized by L2 to obtain the segment vectors, as shown below: Formula (20); in, This represents a text embedding model (Transformer encoding network). Indicates L2 normalization, Indicates the first The semantic vectors of each segment are 768-dimensional dense vectors. After normalization... .

[0129] Document fragments and online mixed search queries must use the same embedding model, the same vector dimension, and the same normalization method to calculate similarity in the same semantic space.

[0130] Data is stored according to the format "Original Document - Document Source Index - Fragment Retrieval Index". Original PDFs, Word documents, webpage snapshots, or text files are stored in the file storage system for original document verification; document-level information such as document titles, publication dates, file hashes, and the validity status of legal rules is included. The primary key is used to write the document source index to Elasticsearch; each document fragment is written as an independent record to the fragment retrieval index of Elasticsearch, with the record structure... As shown below: Formula (21); In this step, Elasticsearch is used for retrieval rather than for storing the original files. Set as a text field, and create an inverted index after Chinese word segmentation for BM25 keyword retrieval; , , and Set as an exact field for fragment location, document association, and data type filtering; and Simultaneously, a word segmentation field and an exact value field are established for name retrieval and source verification; Used to trace back to the original page number; The keyword index is set to a 768-dimensional dense vector field, using cosine similarity and an HNSW index for semantic nearest neighbor retrieval. Thus, the keyword index and vector index are managed by the same system and identified by the same segment. Point to the same segment to avoid data inconsistencies that can occur when using text databases and vector databases separately.

[0131] Before writing to the index, check , , and Is it complete? Check and Check if at least one exists, and examine Is it a 768-dimensional non-zero vector? Fragments that fail the check will not be included in the fragment retrieval index. When it is necessary to read the document title, publication date, or original document, it depends on... Retrieved from the document source index or the original file layer, and not saved repeatedly in the fragment metadata.

[0132] Elasticsearch is an existing open-source technology. It's a distributed search and analytics engine built on top of Apache Lucene. Released in 2010, it has become one of the most popular search engines. It's a database system specifically designed for the rapid retrieval and analysis of massive amounts of data.

[0133] In this invention, Elasticsearch is used as the core retrieval infrastructure. Specifically, it plays two key roles: ①Document Source Index: Using the document identifier doc_id as the primary key, it stores document-level metadata such as document title, author, publication date, and URL.

[0134] ② Fragment retrieval index: Using the fragment identifier chunk_id as the primary key, it stores the text, metadata, and corresponding 768-dimensional semantic vector of each document fragment.

[0135] In this invention, the keyword index and vector index are managed by the same system. This leverages Elasticsearch's ability to support both types of indexes on a single platform, thus avoiding the data inconsistency issues that arise from maintaining two independent systems. Essentially, it acts as a powerful "database" to support the hybrid retrieval strategy employed in this invention.

[0136] Step S32: Perform online retrieval and query to generate the professional context corresponding to the current user dialogue.

[0137] The online phase is a process of responding to user queries in real time, based on the received user dialogue and the three-dimensional structured link retrieves the most relevant professional segments from the offline constructed knowledge base.

[0138] The structured fields therein are directly used to generate keyword queries and semantic queries without invoking a general large model for query rewriting.

[0139] Based on the current user dialogue and the three-dimensional structured link generated in real time by the first model, a weighted keyword query and a semantic vector query are respectively generated, including: Reading the fields in the question dimension information, scene dimension information and basis dimension information from the three-dimensional structured link, performing Chinese word segmentation and stop word filtering on the read field content, assigning weights to the segmented terms according to preset weight rules, to form a weighted keyword query; Splicing the non-empty fields in the three-dimensional structured link into a semantic query text in a preset order, adding a fixed retrieval instruction before the semantic query text, and inputting the same into a pre-trained text embedding model to generate a semantic vector query; The weighted keyword query is used to perform BM25 keyword recall in the public material knowledge base, and the semantic vector query is used to perform HNSW vector semantic recall in the public material knowledge base.

[0140] (1) Keyword query generation: Read the question dimension key sub-questions , target results and constraint set , the scene dimension in , and , as well as in the basis dimension .

[0141] Specific processing steps: ① Perform Chinese word segmentation and stop word filtering on the above fields, and remove meaningless words such as "de", "le", and "shi"; ② Fully retain the subject name, behavior name, legal name, clause number and numerical conditions, which are the keys for exact matching.

[0142] ③ After setting the weights, a weighted keyword query is formed .

[0143] from , and The weight of the obtained terms is preset to 3. , and The weight of the acquired terms is preset to 2, and the weight of the remaining supplementary terms is preset to 1; when the same term appears repeatedly, the highest weight is used. The processed weighted term set constitutes the keyword query. In practical applications, preset weights can be flexibly set according to specific needs.

[0144] (2) Semantic vector query generation: Semantic query generation consists of three steps: The first step is to concatenate the semantic query text.

[0145] Following a fixed order of "core issue, processing objective, constraints, relevant subjects, target audience, current state, scenario description, and public information requirements," the non-empty content of the corresponding fields in the three-dimensional structured query chain is read, and the non-empty fields are concatenated to form semantic query text. ; Then, supplement the proprietary names and numerical information that were not included in the above fields in the most recent round of user input to ensure that no key information mentioned by the user in the conversation but not entered into the structured link is missed.

[0146] The second step is to attach fixed search instructions.

[0147] exist Prepend the fixed search instruction "Generate a representation for this sentence to retrieve related articles:"; The third step is to generate a normalized query vector.

[0148] for The BAAI / bge-base-zh-v1.5 model was used to obtain the normalized query vector, as shown below: Formula (22); in, This indicates a fixed search instruction, equivalent to "Generate a representation for this sentence to retrieve related articles:". For the BAAI / bge-base-zh-v1.5 model, Indicates text concatenation. For L2 normalization; This represents a 768-dimensional normalized query vector.

[0149] Keyword queries are used to find segments with the same literal words, while query vectors are used to find segments with different words but similar meanings.

[0150] The BM25 keyword recall and HNSW vector semantic recall are performed in parallel. Based on the results of the two recalls, corresponding professional contexts are generated, including: Based on the requirement description in the three-dimensional structured link and the set of constraints in the scene dimension, the data type is determined and data type filtering conditions are generated. Based on the data type filtering conditions, the BM25 algorithm is used to perform keyword recall on the weighted keyword query to obtain the BM25 keyword recall list; in parallel, based on the data type filtering conditions, the HNSW algorithm is used to perform vector semantic recall on the semantic vector query to obtain the semantic recall list. A reciprocal ranking fusion algorithm is used to merge and sort the BM25 keyword recall list and the semantic recall list. The reciprocal ranking fusion algorithm calculates the fusion score only using the ranking of each segment in the BM25 keyword recall list and the semantic recall list. For segments that appear only in one recall list, only that ranking is used to calculate the score. For segments that appear in both the BM25 keyword recall list and the semantic recall list, the scores corresponding to the rankings in both lists are added together. All segments are then merged and sorted from high to low according to the fusion score. The results after fusion and sorting are sequentially subjected to source integrity verification, text hash deduplication, vector cosine similarity near-duplicate filtering, and control of the number of homologous segments; based on the fusion score, no more than five segments are selected to generate corresponding professional contexts.

[0151] Keyword recall and semantic recall: Before performing actual recall, based on and Generate filter conditions based on explicit data types. Laws and regulations, compliance, or specific legal names correspond to legal rules; papers, research, models, or methods correspond to academic papers and textbooks; cases, actions, or judgments correspond to legal cases. When multiple categories are matched, the union of the categories is taken; otherwise, there are no restrictions. This filter operates directly on the Elasticsearch fragment retrieval index, reducing the number of irrelevant data types appearing in candidate results.

[0152] (1) BM25 keyword recall. BM25 (Best Matching 25) is a classic bag-of-words retrieval model used to calculate the relevance score between a document and a query.

[0153] use Search The inverted index records the segment where a term appears and its frequency. BM25 combines the frequency of a term in a segment, its rarity in all segments, and the segment length to calculate a score, as shown below: Formula (23); in, Let be the weight of term t, where t is a term in the query. The number of times term t appears in the main text of the passage. Inverse document frequency, and These are the current segment length and the average segment length, respectively. is the word frequency saturation parameter; b is the length normalization parameter.

[0154] For example, =1.5, b=0.75. The system meets the filtering conditions. From the fragments, the top five are selected in descending order of BM25 score to form a keyword recall list. As shown below: Formula (24); in, This refers to document fragments numbered 1 through 5 in the recall list.

[0155] (2) Vector semantic recall: HNSW Algorithm Principle: HNSW (Hierarchical Navigable Small World) is an approximate nearest neighbor search algorithm. It constructs a multi-layered graph structure: The bottom layer contains all vector nodes, and each node is connected to several neighboring nodes ("small world").

[0156] The upper layer is a "miniature" of the lower layer, with fewer nodes and sparser connections (jumping further).

[0157] The search starts from the top layer to quickly locate the general neighborhood, and then proceeds to more precise locations layer by layer downwards.

[0158] This hierarchical design enables HNSW to perform nearest neighbor searches in high-dimensional spaces in logarithmic time complexity, which is much faster than brute-force traversal.

[0159] Simultaneous use Search The HNSW vector index. HNSW first searches in the graph structure for the HNSW vector index. The 10 nearest candidate vectors are then sorted by cosine similarity, as shown below: Formula (25); The value ranges from [-1, 1], and the larger the value, the closer the semantics are.

[0160] The higher the cosine similarity, the closer the overall semantics of the query and the fragment are.

[0161] HNSW first searches in the graph structure for... The 10 nearest candidate vectors are then sorted by cosine similarity, and the five with the highest similarity are selected to form a semantic recall list. As shown below: Formula (26); in, For semantic recall list The first five recalled document fragments.

[0162] BM25 is responsible for precise matching of technical terms, subject names, and legal names, while vector recall is responsible for synonymous expressions and overall semantic matching. The two outputs are fragment identifiers and rankings respectively, without directly comparing the two types of raw scores.

[0163] (3) RRF (Reciprocal Rank Fusion) fusion sorting and result filtering: BM25 score: Depends on query and document length, typically ranging from 0 to tens.

[0164] Cosine similarity: The range is fixed at [-1, 1].

[0165] The BM25 score and cosine similarity have different numerical ranges and dimensions, so they cannot be directly added or compared. Direct weighted averaging would lead to one side dominating.

[0166] The Reverse Rank Fusion Algorithm (RRF) is a fusion method based on ranking rather than score. The core idea is that the higher the ranking of a fragment in a certain recall path (the smaller the ranking value), the greater its contribution to the total score.

[0167] Merging is performed solely based on the fragment's ranking in both lists. (By...) According to merge and For any candidate fragment, calculate the fusion score. As shown below: Formula (27); in, The candidate segment is the segment that is recalled by any path; Recording fragments In the The ranking in the road list, such as if this segment is ranked 3rd in the BM25 list, then ;, Indicates whether a fragment appears; a value of 1 indicates its presence, and a value of 0 indicates its absence. This is a smoothing constant. To prevent overfitting caused by excessively high scores from top-ranked segments, k is set to 60.

[0168] Segments that simultaneously enter the two higher-ranked results will have two scores added, resulting in a higher ranking.

[0169] After merging, four filters are executed sequentially: First, delete , or Missing, and and Simultaneously missing segments; then based on Delete segments of legal rules that cannot be found in the document source index or that are marked as invalid.

[0170] Second, after normalizing the main text, calculate the text hash, and for segments with the same hash, only retain the one with the highest RRF score.

[0171] Third, according to the RRF score from high to low, when the vector cosine similarity between the current segment and the retained segments is not less than 0.95, it is identified as a near duplicate and deleted.

[0172] Fourth, the same At most two highest-ranked segments will be retained. After filtering, the segments will be sorted in descending order of RRF score, and if the scores are the same, segments with higher vector similarity will be retained first.

[0173] (4) Generate professional context: Select the first five segments from the sorted results; if there are fewer than five, select all remaining segments to form a professional context, as shown below: Formula (28); in, This indicates the number of filtered segments. If there are fewer than 5 filtered segments, all are selected; if there are more than 5, only the first 5 are selected. The system selects segments based on their... The document title, publication date, and original file location are retrieved from the document source index to form a source record. And define the professional context item as: Formula (29); in, For fragment metadata, This is the main text of the excerpt. To pass Document-level source records obtained, The RRF fusion score is calculated. The system integrates the PC, user dialogue X, and 3D structured link. They are jointly passed to the subsequent report generation module; the report generation module uses... As professional content, and using , , and This creates source attribution, allowing the report content to be traced back to the original document.

[0174] If no qualified segments are found after filtering, output the professional context. The system will mark any missing publicly available information and will not allow the fabrication of professional evidence in subsequent steps.

[0175] This indicates "I need to search for publicly available information related to...", meaning what you want to find (your needs); Professional context This indicates "what was actually retrieved," consisting of the fragment text + chunk_id + source information, thus completing the transformation from a description of the data requirement to a traceable professional context. What was actually found (the result)?

[0176] Step S3 is to perform parallel keyword and semantic hybrid retrieval based on the three-dimensional structured link generated by the user dialogue and the first model, through the constructed public knowledge base, and generate traceable and redundant professional context after fusion sorting and filtering, so as to provide accurate and verifiable external knowledge basis for subsequent report generation.

[0177] Step S4 includes steps S41-S42.

[0178] This step is used to address the user's original question. 3D structured link and professional context By integrating these three elements, a well-structured, well-founded, and traceable decision support report is generated. Furthermore, through quantitative evaluation and targeted feedback mechanisms, a closed-loop optimization of report quality is achieved. For example... Figure 4 As shown.

[0179] S4 has three core objectives: ① Controlled generation: External facts, legal rules, research conclusions, and case content in the report can only use fragments from the PC and cannot be supplemented using the model's own memory.

[0180] ②Automatic source tracing: Based on the actual segment identifier chunk_id used in the report, a source list can be automatically generated to eliminate "fake citations".

[0181] ③ Closed-loop optimization: Through three-dimensional quantitative evaluation indicators, problem links are accurately located, and targeted return is made for reprocessing and correction, forming a limited number of quality closed loops.

[0182] The input received in this step As shown below: Formula (30); Where X represents a single-turn or multi-turn conversation between users; A three-dimensional structured link corresponding to user dialogue; The corresponding professional context.

[0183] Step S41: Generate a decision support report.

[0184] Based on the current user dialogue, three-dimensional structured links, and professional context, multi-agent collaborative generation of decision support reports is performed, including: Write the target results, key sub-problems, problem priorities, and constraint sets in the three-dimensional structured link into a preset report template; The report generation agent uses the text fragments in the professional context as the sole source of external facts and professional evidence, and generates a decision support report according to the preset report template. When the report generation agent uses the professional fragments in the professional context, it simultaneously outputs the fragment identifier corresponding to the fragment. Based on the fragment identifiers actually used by the report-generating agent, the corresponding document title, author or publishing organization, page number and source address are obtained from the fragment retrieval index and the document source index to form a data source list; the data source list is generated based on the index records and is not freely filled in by the report-generating agent. When the professional context is empty or insufficient to support any key sub-problem, mark the corresponding position of the decision support report as insufficient or pending verification of professional information. The report generating agent shall not use its own parameters to memorize and supplement the professional facts that were not provided. The evaluation feedback agent is invoked to perform a quantitative evaluation of the decision support report, and the evaluation results of the decision support report on each evaluation indicator are obtained.

[0185] First, based on the user's original dialogue The corresponding three-dimensional structured links and the professional context obtained from the search ; Next, The target results, key sub-problems, problem priorities, and constraints are written into a fixed report template and then... The professional segments in the text serve as the limiting source of external facts and professional evidence. The report is generated by calling the large language model of the report generation; then the report is automatically evaluated according to preset indicators, and targeted feedback is triggered according to the source of the score loss.

[0186] Finally, it provides a decision support report that includes a problem description, scenario analysis, professional basis, analytical conclusions, handling suggestions, risk warnings, and data sources, along with corresponding report evaluation results. This ensures the report generation process aligns with user intent and publicly available information, guarantees that key issues are addressed, professional content has a proven track record, and that report quality can be checked and continuously optimized using standardized rules.

[0187] Decision support report generation process: Step 1 (Information Assembly Generation): Prioritize key sub-problems according to their importance, and assemble the target results, relevant stakeholders, analysis objects, and constraints into the report generation requirements; at the same time, read fragments of text and source information from the professional context.

[0188] Step Two (Decision Support Report Generation): Input the user's original dialogue question, report generation requirements, and professional context into the first model. Following a fixed report template, the first model sequentially generates the problem and objectives, scenario and constraints, key sub-question analysis, final conclusions, processing recommendations, and risk or limitation statements. External facts, legal rules, research conclusions, and case content in the report can only be used for... Document snippets; when using professional snippets, output the corresponding snippet identifier simultaneously. .

[0189] Step 3 (Source Generation): Based on the actual fragments used in the report. The system retrieves document titles, authors or publishing organizations, page numbers, and source addresses from the fragment retrieval index and document source index, automatically generating a list of source sources. The source sources are generated by the system based on index records, rather than being freely filled in by the large language model, to avoid sources that cannot be traced back.

[0190] Step 4 (Handling Missing Data): When When the existing professional information is insufficient to support a key sub-issue, the report should mark "insufficient professional information" or "to be verified" in the corresponding position. The model should not use its own memory to supplement the professional facts that are not provided.

[0191] This leads to the report to be evaluated. As shown below: Formula (31); in, This indicates that the report generates a large language model, which is a base large language model. The model version, report template version, and generation parameters used are recorded to ensure that the generation process can be repeated under the same system version.

[0192] Step S42: Quantitatively evaluate the decision support report.

[0193] If the evaluation result does not reach the evaluation threshold, the process returns to steps S2, S3, or S4 for reprocessing until the preset conditions are met to obtain the final decision support report, including: The quantitative assessment includes the following three assessment indicators: Content completeness rate is used to check whether the decision support report answers the key sub-questions in the three-dimensional structured link item by item and retains the user's explicit constraints; Professional basis coverage is used to check whether the analytical items in the decision support report that require the use of external facts, legal rules, research conclusions or case content are related to the document fragments in the professional context. Citation traceability rate is used to check whether the citations in the decision support report can be traced back to the original document corresponding to the user dialogue through fragment identifiers and document identifiers. The evaluation result is obtained by weighting and summing the content completeness rate, professional basis coverage rate, and citation traceability rate according to preset weights. The decision support report passes the evaluation when the evaluation result is not lower than the preset threshold and the citation traceability rate is 100%. When the content completeness rate does not reach the preset threshold, it is determined that a key sub-problem or constraint is missing, and the processing flow is returned to step S4 to regenerate the decision support report. When the coverage of the professional basis does not reach the preset threshold, it is determined that the required professional basis has not entered the professional context, and the corresponding key sub-question and basis requirement are returned to step S3 for re-retrieval; When the fields in the problem dimension information or scene dimension information are significantly inconsistent with the user's original dialogue, it is determined that the problem target or scene information is incorrectly identified, and the process returns to step S2 to regenerate the three-dimensional structured link. If the evaluation still fails after the preset limit of re-executions is reached, the automatic loop stops and manual review begins.

[0194] The report was evaluated from three aspects: content completeness, coverage of sources, and traceability of citations. Perform automatic evaluation.

[0195] First, content completeness rate Used to check whether the report answers each item. The key sub-problems are identified, and the user-defined constraints are retained, as shown below: Formula (32); in, This indicates the total number of key sub-questions that should be answered and the total number of explicit constraints that should be retained. This indicates the number of responses that have been received in the report.

[0196] Second, professional basis coverage This is used to check whether the analytical content in the report that requires professional evidence is related to the document fragments and whether it is related to valid [documents / documents]. As shown below: Formula (33); in, This indicates the total number of analytical items in the report that use external facts, legal rules, research findings, or case content. Indicates that it has validity The number of analysis items.

[0197] Third, citation traceability rate This is used to check whether every citation in the report can pass the test. and The link is used to locate the original document, the text snippet, the page number, and the source URL; as shown below: Formula (34); in, This indicates the total number of citations in the report. This indicates the number of references that can be obtained for the corresponding text segment, document source record, and original file location.

[0198] Overall score of the report As shown below: Formula (35); in, In one implementation, , , ;when Not less than 0.90 and When the overall score of the report is not lower than 0.90 and the citation traceability rate reaches 100%, the report passes the evaluation. The weights and thresholds can be pre-configured according to the business scenario, but remain unchanged in the same set of evaluation tasks.

[0199] Citation traceability is a baseline for quality; if even one citation cannot be traced back to the original document, the report is considered to have "unreliable basis." In serious scenarios such as compliance audits and legal analysis, untraceable citations are equivalent to no citations at all, hence this is set as a mandatory requirement.

[0200] When a report fails the evaluation, instead of performing an indiscriminate retry, targeted processing is carried out based on the source of the failure.

[0201] ① If key sub-problems or constraints are missing, If the threshold is not met, the report is returned to step S4 for regeneration; the missing sub-problems and constraints are re-emphasized to the report generation model, requiring regeneration. ②If the required professional basis for the report is not included , If the threshold is not reached, then the corresponding key sub-issues and requirements will be addressed. Return to step S3 and perform the search again; ③ If the target or scene information is incorrectly identified, or If the fields in the query clearly do not match the user's original dialogue, then the user's original question is returned to step S2 to regenerate the three-dimensional structured link, and the user's original question is... Return to the first model Regenerate the three-dimensional structured link.

[0202] If the system still fails to pass the preset number of feedback attempts, a manual review flag is displayed to prevent infinite loops. For example, the preset number of feedback attempts is set to 5; however, this can be flexibly adjusted according to specific needs in actual applications.

[0203] Step S4 aims to generate a controlled and traceable decision support report through multi-agent collaboration based on user dialogue, three-dimensional structured links, and professional context. It then achieves closed-loop quality optimization through quantitative evaluation and targeted feedback mechanisms, ultimately outputting a validated final decision support report.

[0204] Specific implementation examples: The intelligent decision support report on the adjustment of ride-hailing driver access rules is generated as follows: A ride-hailing platform uses this invention to construct a decision support intelligence agent to assist compliance managers in analyzing the risks of adjustments to driver access rules in City A and generating decision support reports with publicly available data sources, such as... Figure 5 As shown, the steps are as follows: Step 1 (Deployment and Initialization): Based on ride-hailing and food delivery scenarios, cold-start dialogue samples are constructed. User dialogues are paired with corresponding three-dimensional structured links of "problem dimension - scenario dimension - basis dimension" to form supervised training data. The Qwen3.5-9B base large language model is then fine-tuned under supervision to obtain the first model. Simultaneously, the system collects publicly available laws, regulations, rule documents, papers, textbooks, and professional cases. After text extraction, fragment segmentation, metadata generation, and vectorization, document source indexes, BM25 inverted indexes, and HNSW vector indexes are established in Elasticsearch.

[0205] Step Two (User-Agent Dialogue): The compliance manager inputs to the agent: "We plan to adjust the admission rules for ride-hailing drivers in City A, considering using online training results as a supplementary judgment condition. Please analyze the risks and provide suggestions." The agent recognizes that the data boundaries and judgment conditions are not yet clear, and further inquires about the reporting deadline, data scope, and whether online training alone determines admission. The user adds: "Complete within three days, using only publicly available information, without accessing driver personal information or platform internal operating data; online training is only a supplementary condition." The agent reiterates and confirms the objectives and constraints, forming a complete multi-turn dialogue context.

[0206] Step 3 (Generation of 3D Structured Link): The system inputs the dialogue into the first model to obtain a 3D structured link. The problem dimension identifies the problem type as risk identification, with the goal of assessing the feasibility of rule adjustments and generating implementation recommendations. Key sub-problems include existing rule risks, applicable conditions for online training, and safeguards to be implemented before launch. The scenario dimension identifies the business scenario as ride-hailing, the questioning role as compliance management personnel, and the analysis object as driver access rules in City A, recording the region, time limit, data scope, and judgment boundaries. Based on this dimension, publicly available information related to driver access requirements, applicable conditions for online training, appeal and manual review mechanisms, and similar governance cases is generated. This link, as a parsable and verifiable thought chain analysis result, provides a unified input for subsequent retrieval and report generation.

[0207] Step 4 (Hybrid Retrieval): Based on user dialogue and the 3D structured link, a weighted keyword query and a 768-dimensional semantic vector query are generated. BM25 keyword retrieval and HNSW vector retrieval are performed respectively, and then the RRF algorithm is used to merge and rank the two results. The system deletes fragments with incomplete sources, untraceable original text, duplicate content, or expired content, and selects the top-ranked professional fragments to form a professional context including the fragment text, document identifier, page number, publishing organization, and source address.

[0208] Step 5 (Report Generation): Input the user's original dialogue, three-dimensional structured link, and professional context into the report to generate a large language model. Following a fixed template, generate the problem and objectives, scenario and constraints, key sub-problem analysis, final conclusions, processing suggestions, risk warnings, and data sources. External facts, rules, research conclusions, and case content in the report can only use professionally retrieved fragments; the system automatically binds the corresponding source when using fragments. If the existing data is insufficient to support a sub-problem, mark it as "Insufficient Professional Data" or "Pending Verification" in the report; do not use the model's own memory to supplement professional facts.

[0209] Step Six (Evaluation and Closed-Loop Feedback): The system evaluates the report from three aspects: content completeness, coverage of professional evidence, and traceability of citations. If key sub-issues or user constraints are missing, the system returns to the report generation module to regenerate the report; if professional evidence is insufficient, the corresponding sub-issues and evidence requirements are returned to the hybrid search module for supplementary retrieval; if the problem target or scenario is incorrectly identified, the system returns to the first model to regenerate the three-dimensional structured link. If the system still fails to pass after reaching the preset number of feedback attempts, it outputs a manual review mark; after passing the evaluation, it outputs a decision support report, a list of data sources, and the evaluation results.

[0210] Implementation Results: Through the above processing, this embodiment sequentially completes the continuous process of "user dialogue with intelligent agent - three-dimensional structured thinking chain analysis - mixed retrieval of public information - professional report generation - quality assessment feedback", which enables user goals and constraints to be preserved, professional content in the report to be traced back to public information, and reduces the risk of generating unfounded conclusions when data is insufficient.

[0211] In summary, the decision-making method based on structured inference chain fine-tuning and multi-agent collaboration of this invention has the following beneficial effects: 1. This invention improves the completeness of complex user intent parsing: by using the question dimension, scenario dimension, and basis dimension to provide a unified structured expression of single-turn or multi-turn user dialogues, it can clearly define the problem objectives, key sub-problems, relevant participants, analysis objects, constraints, and public information requirements, reducing the deviation of the main problem and the omission of constraints caused by relying solely on keywords or free text generation; 2. This invention improves the stability of model output in cold start scenarios: It uses scenario-based cold start dialogue samples and three-dimensional structured links to form supervised data, enabling the base model to learn the mapping relationship from natural language dialogue to fixed field links, and performs quantitative evaluation through structural integrity rate, field accuracy rate and requirement description compliance rate, which facilitates objective verification of fine-tuning effects and control of online output format. 3. This invention balances the accuracy and coverage of professional data retrieval: It employs a dual-path retrieval method combining BM25 (a bag-of-words retrieval model widely used in information retrieval search engines to evaluate the relevance of a document to a given query) keyword retrieval with HNSW (Hierarchical Navigable Small World, an efficient approximate nearest neighbor search algorithm) vector semantic retrieval. Furthermore, it utilizes RRF (Reciprocal Rank Fusion, a ranking fusion algorithm that does not require score normalization, specifically designed to merge multi-path retrieval results, such as the keyword retrieval list from BM25 and the vector retrieval list from HNSW, into a unified, superior ranking list) fusion, data type filtering, and multi-level deduplication. This approach not only preserves the precise matching ability of legal names, subject names, and professional terms, but also retrieves data with synonymous expressions and overall semantic relevance. 4. This invention enhances the traceability of professional evidence: It links the fragment retrieval index, document source index, and original file layer by using the fragment ID (chunk_id) and document ID (doc_id) in the document. The data source in the report is generated based on the actual fragment used. When professional data is insufficient, it clearly outputs missing or unverified marks, reducing the risk of the model fabricating professional facts or generating data whose source cannot be traced back. 5. This invention achieves quantifiable closed-loop optimization of report quality: It conducts a unified evaluation of the report from three aspects: content completeness, coverage of professional basis, and traceability of citations. Based on the reasons for the loss of points, the task is redirected back to the link generation, data retrieval, or report generation stage for reprocessing, so that the error can be located to the specific processing stage and avoids aimless regeneration. 6. This invention improves the efficiency of decision support report generation and system scalability: Through standardized data transfer between multiple agents, a continuous processing flow is formed from user problem analysis and professional information retrieval to report generation, evaluation, and manual review, reducing the workload of manually understanding problems item by item, searching for information across sources, and organizing citations; by expanding scenario samples, updating the public information knowledge base, and configuring evaluation thresholds, it can be adapted to different platform business scenarios.

[0212] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0213] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A decision-making method based on structured reasoning chain fine-tuning and multi-agent collaboration, characterized in that, Includes the following steps: Step S1: Generate a corresponding 3D structured link based on each user dialogue sample in the cold start dialogue dataset; pair each user dialogue sample with the corresponding 3D structured link to construct a supervised training sample set; Step S2: The base large language model is fine-tuned under supervision using the supervised training sample set to obtain the first model; wherein, the first model is used to generate the corresponding three-dimensional structured link in real time based on the user dialogue of the input natural language. Step S3: Using a pre-built public knowledge base, based on the current user dialogue and the three-dimensional structured link generated in real time by the first model, weighted keyword query and semantic vector query are generated respectively. BM25 keyword recall and HNSW vector semantic recall are executed in parallel. Based on the two recall results, the corresponding professional context is generated. Step S4: Based on the current user dialogue, three-dimensional structured link, and professional context, a multi-agent collaborative decision support report is generated; the decision support report is quantitatively evaluated; if the evaluation result does not reach the evaluation threshold, the process returns to steps S2, S3, or S4 for reprocessing until the preset conditions are met to obtain the final decision support report.

2. The decision-making method based on structured inference chain fine-tuning and multi-agent collaboration according to claim 1, characterized in that, Based on each user dialogue sample in the cold start dialogue dataset, a corresponding three-dimensional structured link is generated, including: Based on single-turn or multi-turn dialogues, and utilizing pre-defined link generation template constraints, each user dialogue sample is subjected to question identification, scenario identification, and requirement description identification to obtain the corresponding three-dimensional structured link. Among them, the user dialogue sample includes user questions, supplementary conditions, historical follow-up questions, business constraint information, and expected output targets. The three-dimensional structured link includes problem-dimensional information, scene-dimensional information, and basis-dimensional information; The issue dimension information includes the user's issue type, target result, set of key sub-issues, and issue priority: The scenario dimension information includes scenario type, questioning user role, set of relevant participants, analysis object, and set of constraints. The information provided is a description of the requirements and is used to indicate the direction of information to be focused on in subsequent public information retrieval. It does not indicate specific evidence content, does not limit specific information sources, and does not include internal business data of the platform.

3. The decision-making method based on structured inference chain fine-tuning and multi-agent collaboration according to claim 2, characterized in that, The preset link generation template constraints include top-level structure constraints, field-level format constraints, content boundary constraints, and serialization format constraints; The top-level structural constraints are used to force the three-dimensional structured link to include three dimensions: problem dimension information, scene dimension information, and basis dimension information, and the order of these dimensions must be fixed. No dimension may be missing, and no undefined additional fields may be added. The field-level format constraints are used to specify the data type and value specifications for each field in the problem dimension information, scenario dimension information, and basis dimension information. The content boundary constraints are used to prohibit the basis dimension information from containing specific evidence content, and to prohibit the self-creation of constraints in the constraint condition set of the scenario dimension information when the user has not explicitly given constraints. The serialization format constraint is used to require the three-dimensional structured link to be serialized into a parsable data format in a fixed field order.

4. The decision-making method based on structured reasoning chain fine-tuning and multi-agent collaboration according to claim 1, characterized in that, Each user dialogue sample is paired with its corresponding 3D structured link to construct a supervised training sample set, including: Each supervised training sample includes input data and output data; The input data is a single-turn or multi-turn dialogue between users; the output data is a three-dimensional structured link corresponding to the input data. All supervised training samples are divided into a training set and a test set according to a preset ratio; the training set is used to update model parameters during the training of the first model, and the test set is used to evaluate the performance after the first model is trained.

5. The decision-making method based on structured inference chain fine-tuning and multi-agent collaboration according to claim 4, characterized in that, The base large language model is fine-tuned under supervision using the aforementioned supervised training sample set to obtain a first model, including: Using the samples in the training set as training data, the input data as input to the base big language model, and the output data as the supervised target output of the base big language model, the base big language model learns the mapping relationship from user dialogue in natural language to standardized link data. The output of the supervision target is converted into a tag sequence, which consists of multiple tags formed by sequentially splitting the output of the supervision target; Using the cross-entropy loss function, for each labeled position in the supervised target output, the probability that the model predicts that position is correctly labeled is calculated. The predicted probabilities of all labeled positions are summed and the negative logarithm is taken to obtain the loss value of the training sample. The loss value is used to quantify the difference between the model prediction result and the standard output. Based on the loss value, the parameters of the base large language model are updated through backpropagation. The training is iterated with the goal of minimizing the loss value. After completing a preset number of training rounds, the model parameters are saved to obtain the first model.

6. The decision-making method based on structured inference chain fine-tuning and multi-agent collaboration according to claim 1, characterized in that, Based on the current user dialogue and the three-dimensional structured link generated in real time by the first model, weighted keyword queries and semantic vector queries are generated respectively, including: The problem dimension information, scenario dimension information and basis dimension information are read from the three-dimensional structured link. Chinese word segmentation and stop word filtering are performed on the read field content. According to the preset weight rules, the segmented words are assigned weights to form a weighted keyword query. The non-empty fields in the three-dimensional structured link are concatenated into semantic query text according to a preset order. A fixed retrieval instruction is appended before the semantic query text and then input into a pre-trained text embedding model to generate a semantic vector query. The weighted keyword query is used to perform BM25 keyword retrieval in the public information knowledge base, and the semantic vector query is used to perform HNSW vector semantic retrieval in the public information knowledge base.

7. The decision-making method based on structured reasoning chain fine-tuning and multi-agent collaboration according to claim 6, characterized in that, The BM25 keyword recall and HNSW vector semantic recall are performed in parallel. Based on the results of the two recalls, corresponding professional contexts are generated, including: Based on the requirement description in the three-dimensional structured link and the set of constraints in the scene dimension, the data type is determined and data type filtering conditions are generated. Based on the data type filtering conditions, the BM25 algorithm is used to perform keyword recall on the weighted keyword query to obtain the BM25 keyword recall list; in parallel, based on the data type filtering conditions, the HNSW algorithm is used to perform vector semantic recall on the semantic vector query to obtain the semantic recall list. A reciprocal ranking fusion algorithm is used to merge and sort the BM25 keyword recall list and the semantic recall list. The reciprocal ranking fusion algorithm calculates the fusion score only using the ranking of each segment in the BM25 keyword recall list and the semantic recall list. For segments that appear only in one recall list, only that ranking is used to calculate the score. For segments that appear in both the BM25 keyword recall list and the semantic recall list, the scores corresponding to the rankings in both lists are added together. All segments are then merged and sorted from high to low according to the fusion score. The results after fusion and sorting are sequentially subjected to source integrity verification, text hash deduplication, vector cosine similarity near-duplicate filtering, and control of the number of homologous segments; based on the fusion score, no more than five segments are selected to generate corresponding professional contexts.

8. The decision-making method based on structured reasoning chain fine-tuning and multi-agent collaboration according to claim 6, characterized in that, Pre-built public knowledge bases are obtained through the following methods: Publicly available information is obtained through automatic downloading, API call extraction, or manual import; the publicly available information includes at least one or more of the following: academic papers, professional textbooks, publicly available professional cases, laws and regulations, and publicly available rule documents. For each piece of publicly available information obtained, a unique document identifier is generated, and the document title, author, publishing organization, information type, publication date, source address, acquisition time, and original file hash value are recorded; when multiple original files have the same hash value, the one with the earliest acquisition time is retained; The publicly available information is processed sequentially to perform document format recognition and text extraction to obtain the document body. The document body is segmented into segments, with each document divided into several document segments, and the text of each segment controlled within a preset word count range; Generate a unique segment identifier for each document segment, and record its document identifier, author, publishing organization, page number, document type and source address to form a standardized segment text; The standardized fragment text is encoded using a pre-trained Chinese text embedding model, and fragment vectors are obtained after L2 normalization. A keyword inverted index of the standardized fragment text and a vector index of the fragment vector are established in the fragment retrieval index. The keyword index and the vector index point to the same fragment through the same fragment identifier.

9. The decision-making method based on structured reasoning chain fine-tuning and multi-agent collaboration according to claim 1, characterized in that, Based on the current user dialogue, three-dimensional structured links, and professional context, multi-agent collaborative generation of decision support reports is performed, including: Write the target results, key sub-problems, problem priorities, and constraint sets in the three-dimensional structured link into a preset report template; The report generation agent uses the text fragments in the professional context as the sole source of external facts and professional evidence, and generates a decision support report according to the preset report template. When the report generation agent uses the professional fragments in the professional context, it simultaneously outputs the fragment identifier corresponding to the fragment. Based on the fragment identifiers actually used by the report-generating agent, the corresponding document title, author or publishing organization, page number and source address are obtained from the fragment retrieval index and the document source index to form a data source list; the data source list is generated based on the index records and is not freely filled in by the report-generating agent. When the professional context is empty or insufficient to support any key sub-problem, mark the corresponding position of the decision support report as insufficient or pending verification of professional information. The report generating agent shall not use its own parameters to memorize and supplement the professional facts that were not provided. The evaluation feedback agent is invoked to perform a quantitative evaluation of the decision support report, and the evaluation results of the decision support report on each evaluation indicator are obtained.

10. The decision-making method based on structured reasoning chain fine-tuning and multi-agent collaboration according to claim 9, characterized in that, If the evaluation result does not reach the evaluation threshold, the process returns to steps S2, S3, or S4 for reprocessing until the preset conditions are met to obtain the final decision support report, including: The quantitative assessment includes the following three assessment indicators: Content completeness rate is used to check whether the decision support report answers the key sub-questions in the three-dimensional structured link item by item and retains the user's explicit constraints; Professional basis coverage is used to check whether the analytical items in the decision support report that require the use of external facts, legal rules, research conclusions or case content are related to the document fragments in the professional context. Citation traceability rate is used to check whether the citations in the decision support report can be traced back to the original document corresponding to the user dialogue through fragment identifiers and document identifiers. The evaluation result is obtained by weighting and summing the content completeness rate, professional basis coverage rate, and citation traceability rate according to preset weights. The decision support report passes the evaluation when the evaluation result is not lower than the preset threshold and the citation traceability rate is 100%. When the content completeness rate does not reach the preset threshold, it is determined that a key sub-problem or constraint is missing, and the processing flow is returned to step S4 to regenerate the decision support report. When the coverage of the professional basis does not reach the preset threshold, it is determined that the required professional basis has not entered the professional context, and the corresponding key sub-question and basis requirement are returned to step S3 for re-retrieval; When the fields in the problem dimension information or scene dimension information are significantly inconsistent with the user's original dialogue, it is determined that the problem target or scene information is incorrectly identified, and the process returns to step S2 to regenerate the three-dimensional structured link. If the evaluation still fails after the preset limit of re-executions is reached, the automatic loop stops and manual review begins.

Images