Question and answer method based on large model and data agent

Abstract

The invention belongs to the technical field of large models and data agents, and particularly relates to a question answering method based on a large model and a data agent, and the method comprises the steps: receiving and preprocessing a user question, and generating a question data set; judging whether external data support is needed or not by utilizing a large model, and if so, dynamically generating an agent calling strategy; real-time data acquisition, domain knowledge analysis, multi-source data fusion and logical reasoning verification intelligent agent cooperative execution tasks are scheduled according to a strategy; fusing the traceability information returned by the intelligent agent and the knowledge of the large model to generate a preliminary answer, and outputting the preliminary answer after interpretable optimization; and finally, optimizing the case library according to user feedback and iterating an agent calling strategy. All data agent processing results are attached with full-link traceability information, data sources and reasoning processes are presented in an explanatory optimization mode, and a user can verify answer basis.

Description

Technical Field

[0001] This invention belongs to the field of large model and data intelligent agent technology, specifically relating to a question-answering method based on large model and data intelligent agent. Background Technology

[0002] With the widespread application of large language model (GPT series, LLaMA series, etc.) technologies, question answering systems have evolved from the traditional "rule matching + simple retrieval" model to a "large model generation + external knowledge enhancement" model. However, existing technologies still have several key limitations, making it difficult to meet the demands for high accuracy and high reliability in complex application scenarios: Large language models rely heavily on pre-trained corpora, which leads to a lag in knowledge updates and an inability to reflect the latest clinical guidelines, drug changes, or epidemic dynamics in a timely manner. For example, a model trained on 2023 data cannot obtain the latest revision in the 2024 "Guidelines for the Prevention and Treatment of Type 2 Diabetes in China" which states that "the eGFR threshold for SGLT2 inhibitors is lowered to 30 ml / min / 1.73 m²". Existing systems often use fixed combinations of intelligent agents and preset execution sequences, which cannot be dynamically adjusted according to the complexity and type of the problem; the systems often lack a comprehensive and transparent traceability mechanism, making it difficult for doctors to easily verify the literature, guidelines or data sources on which the answers are based, thus reducing clinical credibility; Medical knowledge is frequently updated, but existing systems only store binary data of "question-answer" and do not link the complete chain of "agent invocation strategy-processing-error cause".

[0003] There is an urgent need for an innovative approach that integrates dynamic agent scheduling, full-link tracing, interpretable reasoning, and data-driven iteration to break through existing technological bottlenecks and improve the practicality and reliability of question-answering systems. Summary of the Invention

[0004] In view of the above-mentioned shortcomings in the existing technology, the present invention provides a question-answering method based on large models and data agents to solve the problems in the background technology.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The question-answering method based on large models and data agents includes the following steps: S1: Receive a natural language question input by the user, wherein the natural language question includes text form or text form converted by speech recognition; S2: Preprocess the natural language questions to generate a question dataset, which includes key entities, question types, and difficulty coefficients, where the difficulty coefficients are divided into five levels according to the complexity of the questions; S3: Input the question dataset from step S2 into the large model, and determine whether it can generate an accurate and complete answer based on the pre-trained knowledge. If it cannot answer, proceed to step S4. S4: Based on the problem dataset in step S2, dynamically generate an agent invocation strategy, including selecting the type of agent to be invoked and setting the execution order. Low-difficulty problems are executed in parallel, while high-difficulty problems are executed serially, and the weight of the logic reasoning verification agent is increased. S5: Based on the strategy, schedule and execute at least one of the following intelligent agents: real-time data acquisition, domain knowledge parsing, multi-source data fusion, or logical reasoning verification. Each intelligent agent returns a secondary answer with traceability information. S6: The large model integrates all secondary answers and its own knowledge to generate a primary answer, and outputs the final answer after interpretability optimization. The interpretability optimization includes embedding data source annotations or presenting a visualization of the reasoning process. S7: Update the case library based on user feedback, and use machine learning algorithms to optimize the agent calling strategy and adjust the weight of each agent based on the case library data. Optimize evaluation indicators including question answering accuracy, user satisfaction and calling efficiency, and set a target threshold for each indicator.

[0006] Furthermore, the preprocessing in step S2 involves using natural language processing techniques to segment the question, tag parts of speech, and identify named entities to extract key entities; using an intent classification model finely tuned with a domain question-answering dataset to determine the question type; and generating a difficulty coefficient based on a weighted scoring rule that considers question length, the number of key entities, and logical complexity, where logical complexity is the number of derivation steps required to derive the target answer from the known information of the question.

[0007] Furthermore, the agent type selection rule in step S4 is as follows: when the problem type is a fact query, the real-time data acquisition agent is called; when the problem type is a domain consultation, the domain knowledge analysis agent is called; when the problem type is a numerical calculation, the real-time data acquisition agent and the multi-source data fusion agent are called; when the problem type is a logical reasoning, the domain knowledge analysis agent and the logical reasoning verification agent are called, and the agent combination is adjusted according to the difficulty coefficient.

[0008] Furthermore, the execution flow of the real-time data acquisition agent in step S5 is as follows: generating a structured search query based on the key entities in step S2; calling pre-configured data source interfaces according to preset priorities, wherein the preset priorities are official data source interfaces over authoritative database interfaces, and authoritative database interfaces over trusted web page interfaces; cleaning the original data to remove duplicate and abnormal data, converting the cleaned data into a structured format, and outputting it after attaching traceability information.

[0009] Furthermore, the execution flow of the domain knowledge parsing agent in step S5 is as follows: load the vertical domain knowledge base according to the question type in step S2. The vertical domain knowledge base supports automatic synchronization of official updated content according to a preset period. Based on the key entities in step S2, the target chapters and entries in the knowledge base are located by combining keyword precise matching and semantic similarity comparison. The located content is deeply analyzed to extract concept definitions, association rules and case reference information, which are then organized into a structured domain knowledge package and output with source information attached.

[0010] Furthermore, the execution flow of the multi-source data fusion agent in step S5 is as follows: receiving the processing results of at least two different data agents, wherein the processing results contain structured data and corresponding traceability information; associating data pointing to the same entity in different processing results through an entity alignment algorithm, wherein the entity alignment algorithm achieves matching based on entity attribute features and context information; detecting data conflicts and resolving conflicts based on the data source credibility description, selecting high-credibility data as valid data; assigning weights to valid data according to the data source credibility, fusing them to form a unified structured dataset, and outputting it after attaching a traceability information summary of each data source.

[0011] Furthermore, the execution flow of the logical reasoning verification agent in step S5 is as follows: a structured reasoning chain is constructed based on the input data. The input data is the domain knowledge package of the domain knowledge parsing agent or the structured dataset of the multi-source data fusion agent. The structured reasoning chain includes premises, reasoning steps, and conclusions. The validity of the reasoning steps is verified by applying preset verification rules, which are logical rules or domain-specific causal verification models. If there are nodes in the reasoning chain that are questionable in terms of validity, the associated data source information or domain rules are supplemented as verification basis, and the location of the questionable node, the reason for the question, and the traceability information of the supplementary basis are marked.

[0012] Furthermore, the interpretability optimization in step S6 specifically involves: when embedding data source annotations, presenting them in the final answer in a specified format, the specified format including the data source identifier, data acquisition time, and data source credibility description; when presenting visualization information of the reasoning process, displaying the structured reasoning chain in the form of dotted arrows, presenting the premise, reasoning steps, and conclusion in sequence, and annotating the verification results of each reasoning step.

[0013] Furthermore, the update of the case library in step S7 specifically involves: when the feedback is positive, the question ID, question text, preprocessing result of step S2, agent invocation strategy of step S4, processing results of each data agent, final answer, and feedback label are associated and stored in the case library; when the feedback is negative, the negative reasons are analyzed, such as omission of agent invocation, data tracing error, reasoning logic defect, or ambiguous answer expression. Then, the original question text, negative reason analysis report, corrected agent invocation strategy, corrected processing result, corrected answer, and feedback label are stored as new cases in the case library, and specific labels are marked on the new cases.

[0014] Furthermore, in step S7, the fine-tuning of the large model module is initiated according to a preset cycle or when the amount of new data in the case library reaches a preset threshold. The machine learning algorithm used is a reinforcement learning algorithm or a supervised learning algorithm, and the fine-tuned question-answering accuracy is not less than 90%. User feedback satisfaction is not less than 85%. The efficiency requirements for agent invocation are: difficulty level 1-2 execution ≤ 1 second, difficulty level 3-4 execution ≤ 3 seconds, difficulty level 5 execution ≤ 5 seconds, with a compliance rate of not less than 90%. The fine-tuned agent invocation strategy is verified offline. If the strategy matching accuracy for the target question type improves by a preset percentage compared to before fine-tuning, the fine-tuning result is solidified. If the core evaluation indicators do not improve or decrease, the parameters of the large model module before fine-tuning are rolled back.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. By generating a calling strategy through three-dimensional features of question type, key entity, and difficulty coefficient, and combining it with agent weight scheduling for high-difficulty questions, redundant and insufficient calls are avoided, balancing processing efficiency and answer rigor, thereby improving the completeness rate of answers to complex questions. 2. All data processing results are accompanied by full-link traceability information, with optimized interpretability to present the data source and reasoning process, allowing users to verify the basis of their answers; 3. Multi-data intelligent agents collaboratively integrate real-time data and vertical domain knowledge to solve the problems of delayed knowledge and insufficient domain depth in large models. This improves the accuracy of vertical domain question answering. The closed-loop mechanism of feedback, case library, and machine learning fine-tuning enables precise optimization of intelligent agent calling strategies. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall process of the question-answering method based on large models and data agents of the present invention. Figure 2 This is a diagram illustrating the problem preprocessing workflow; Figure 3 This is a schematic diagram of the data structure of the case library; Figure 4 Here is a flowchart of the agent scheduling algorithm; Detailed Implementation To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0017] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual images. They should not be construed as limiting the scope of this application. To better illustrate the embodiments of the present invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0018] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "inner," and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present application. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0019] In the description of this invention, unless otherwise explicitly specified and limited, the term "connection" or similar designation indicating a connection between components should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0020] Example 1: like Figure 1-4 As shown, the question-answering method of this invention based on large models and data agents includes the following steps: S1: Receives natural language questions input by the user, including text or text converted from speech recognition. S2: Preprocess the natural language questions to generate a question dataset. The question dataset includes key entities, question types, and difficulty coefficients, with the difficulty coefficients divided into five levels according to the complexity of the questions. S3: Input the question dataset from step S2 into the large model, and determine whether it can generate an accurate and complete answer based on the pre-trained knowledge. If it cannot answer, proceed to step S4. S4: Based on the problem dataset in step S2, dynamically generate an agent invocation strategy, including selecting the type of agent to be invoked and setting the execution order. Low-difficulty problems are executed in parallel, while high-difficulty problems are executed serially, and the weight of the logic reasoning verification agent is increased. S5: Based on the strategy, schedule and execute at least one of the following intelligent agents: real-time data acquisition, domain knowledge parsing, multi-source data fusion, or logical reasoning verification. Each intelligent agent returns a secondary answer with traceability information. S6: The large model integrates all secondary answers and its own knowledge to generate a primary answer. After interpretability optimization, the final answer is output. Interpretability optimization includes embedding data source annotations or presenting a visualization of the reasoning process. S7: Update the case library based on user feedback, and use machine learning algorithms to optimize the agent calling strategy and adjust the weight of each agent based on the case library data. Optimize evaluation indicators including question answering accuracy, user satisfaction and calling efficiency, and set a target threshold for each indicator.

[0021] The preprocessing in step S2 involves using natural language processing techniques to segment the question, tag parts of speech, and identify named entities to extract key entities; using an intent classification model finely tuned with a domain question-answering dataset to determine the question type; and generating a difficulty coefficient based on a weighted scoring rule that considers question length, number of key entities, and logical complexity. The logical complexity is the number of derivation steps required to derive the target answer from the known information of the question.

[0022] Specifically, key entities are accurately extracted through word segmentation, part-of-speech tagging, and named entity recognition. The question type is determined by fine-tuning the intent model, and a difficulty coefficient is generated according to weighted rules. This primarily provides accurate question features for subsequent dynamic policy generation, avoiding agent invocation bias caused by feature ambiguity and improving policy matching accuracy.

[0023] The agent type selection rules in step S4 are as follows: when the problem type is a fact query, the real-time data acquisition agent is called; when the problem type is a domain consultation, the domain knowledge analysis agent is called; when the problem type is a numerical calculation, the real-time data acquisition agent and the multi-source data fusion agent are called; when the problem type is a logical reasoning, the domain knowledge analysis agent and the logical reasoning verification agent are called, and the agent combination is adjusted according to the difficulty coefficient.

[0024] Specifically, basic intelligent agent combinations are matched according to problem type, and the combinations are adjusted based on the difficulty level. This achieves a basic solution for each type of problem, plus difficulty-adaptive optimization, avoiding redundant calls for simple problems and insufficient calls for complex problems, thus balancing efficiency and rigor.

[0025] The execution flow of the real-time data acquisition agent in step S5 is as follows: generate a structured search query based on the key entities in step S2; call the pre-configured data source interface according to the preset priority, with the preset priority being that the official data source interface takes precedence over the authoritative database interface, and the authoritative database interface takes precedence over the trusted web page interface; clean the original data to remove duplicate and abnormal data, convert the cleaned data into a structured format, attach traceability information, and then output it.

[0026] Specifically, search queries are generated based on key entities, data sources are invoked according to priority, structured data is cleaned and traceable. This ensures that the acquired data is authoritative, standardized, and traceable, addresses the issue of missing real-time data for large models, and improves data credibility.

[0027] The execution flow of the domain knowledge parsing agent in step S5 is as follows: 1. Load the vertical domain knowledge base according to the question type in step S2. The vertical domain knowledge base supports automatic synchronization of official updated content according to a preset period. 2. Based on the key entities in step S2, locate the target chapters and entries in the knowledge base using a combination of precise keyword matching and semantic similarity comparison. 3. Perform in-depth analysis on the located content, extract concept definitions, association rules, and case reference information, organize them into a structured domain knowledge package, attach source information, and then output it.

[0028] Specifically, the system loads domain knowledge bases according to question type, accurately locates and extracts content, and generates structured knowledge packages with source tracing. This achieves deep adaptation of vertical domain knowledge, solves the problem of shallow domain knowledge in general large models, and provides reliable knowledge support for professional question answering.

[0029] The execution flow of the multi-source data fusion agent in step S5 is as follows: receiving the processing results of at least two different data agents, wherein the processing results contain structured data and corresponding traceability information; associating data pointing to the same entity in different processing results through an entity alignment algorithm, the entity alignment algorithm achieves matching based on entity attribute features and context information; detecting data conflicts and resolving conflicts based on the data source credibility description, selecting high-credibility data as valid data; assigning weights to valid data according to the data source credibility, fusing them to form a unified structured dataset, and outputting it after attaching a traceability information summary of each data source.

[0030] Specifically, it involves linking multi-source data, resolving conflicts, and weighted fusion to output a unified dataset with a source summary. By integrating multi-dimensional information, it avoids data contradictions, improves data integrity and consistency, and provides comprehensive data support for complex question answering.

[0031] The execution flow of the logical reasoning verification agent in step S5 is as follows: a structured reasoning chain is constructed based on the input data. The input data is the domain knowledge package of the domain knowledge parsing agent or the structured dataset of the multi-source data fusion agent. The structured reasoning chain includes premises, reasoning steps, and conclusions. The validity of the reasoning steps is verified by applying preset verification rules, which are symbolic logic rules or domain-specific causal verification models. If there are nodes in the reasoning chain that are questionable in terms of validity, the relevant data source information or domain rules are supplemented as the basis for verification, and the location of the questionable node, the reason for the question, and the traceability information of the supplementary basis are marked.

[0032] Specifically, this involves constructing a chain of reasoning and verifying its validity, supplementing evidence and marking any doubts. It ensures rigorous reasoning logic, exposes and corrects reasoning flaws, and improves the accuracy and traceability of answers to complex questions.

[0033] The interpretability optimization in step S6 specifically involves: when embedding data source annotations, presenting them in the final answer in a specified format, which includes the data source identifier, data acquisition time, and a description of the data source's credibility; when presenting visualization information of the reasoning process, displaying the structured reasoning chain in the form of dotted arrows, presenting the premises, reasoning steps, and conclusions in sequence, and annotating the verification results of each reasoning step.

[0034] Specifically, it embeds data source annotations, uses dotted arrows to display the reasoning chain, and annotates the verification results. This breaks down the "black box," allowing users to verify the basis and logic of the answers, significantly improving user trust in high-credibility scenarios (such as healthcare).

[0035] The specific steps for updating the case library in step S7 are as follows: When the feedback is positive, the question ID, question text, preprocessing result of step S2, agent invocation strategy of step S4, processing results of each data agent, final answer, and feedback label are associated and stored in the case library; when the feedback is negative, the negative reasons are analyzed. Negative reasons include omission of agent invocation, data tracing error, reasoning logic defect, or ambiguous answer expression. The original question text, negative reason analysis report, corrected agent invocation strategy, corrected processing result, corrected answer, and feedback label are then stored as new cases in the case library, and specific labels are marked on the new cases.

[0036] Specifically, positive feedback is associated with and stored across the entire data chain, while negative feedback is analyzed for its causes and revised examples are stored. This provides comprehensive data support for system optimization, avoiding the problem of being unable to locate defects solely through a "problem-answer" approach, and laying the foundation for iteration.

[0037] Step S7 involves fine-tuning the large model module. The fine-tuning process is initiated according to a preset cycle or when the amount of new data in the case library reaches a preset threshold. The machine learning algorithm used is reinforcement learning or supervised learning. The fine-tuned question-answering accuracy is no less than 90%. User feedback satisfaction is no less than 85%. The agent invocation efficiency requirements are: difficulty level 1-2 execution ≤ 1 second, difficulty level 3-4 execution ≤ 3 seconds, difficulty level 5 execution ≤ 5 seconds, with a compliance rate of no less than 90%. The fine-tuned agent invocation strategy is verified offline. If the strategy matching accuracy for the target question type improves by a preset percentage compared to before fine-tuning, the fine-tuning result is solidified. If the core evaluation indicators do not improve or decrease, the parameters of the large model module before fine-tuning are rolled back.

[0038] Specifically, fine-tuning is initiated periodically or based on data volume, optimizing the strategy generation module with a specified algorithm, and then either solidifying or rolling back the results after verification. This enables the system to self-evolve, improves strategy matching accuracy, dynamically adapts to new scenarios, and reduces manual intervention.

[0039] Operating principle: Taking the consultation in the medical field on "the first-line hypoglycemic drugs and applicable range of renal function indicators for patients with type 2 diabetes mellitus complicated with chronic kidney disease" as an example, the implementation process of this invention will be explained in detail.

[0040] S1: Receives natural language input from the user. The user enters the following question through the medical consultation platform: What is the first-line hypoglycemic drug for patients with type 2 diabetes and chronic kidney disease? What are the required kidney function indicators? S2: Preprocess the problem by extracting key entities using a medical-domain fine-tuned named entity recognition model. Key entities extracted include type 2 diabetes (disease entity), chronic kidney disease (disease entity), preferred hypoglycemic drug (medical intervention entity), and kidney function indicators (detection indicator entity). The problem type is determined to be a domain consultation problem.

[0041] The difficulty level is generated based on a problem length of 35 characters, 4 key entities, and a logical complexity requiring 3 steps of derivation: disease combination → guideline matching → medication and indicator recommendation. The difficulty level is generated as Level 4 (high difficulty) through a weighted scoring rule.

[0042] S3: Preliminary judgment of the large-scale model. The preprocessed question model is then used. If the model determines that the question involves the latest clinical guidelines and indicator limitations, exceeds the scope of its own pre-trained knowledge, and requires external domain knowledge and logical verification, it cannot be answered directly. Proceed to S4.

[0043] S4: Agent Generation Invocation Strategy. Based on the domain consultation class + medical entity + level 4 difficulty generation strategy, it invokes the domain knowledge parsing agent and the logical reasoning verification agent, adopting a serial execution order. The invocation weight of the logical reasoning verification agent is increased by 20%, giving it priority in allocating computing resources.

[0044] S5: Scheduling and Execution Data Agent. This domain knowledge parsing agent loads the "2024 Guidelines for the Prevention and Treatment of Diabetes" knowledge base, which has been updated periodically with official sources. Based on key entities, it uses keyword matching and semantic similarity comparison to locate medication chapters; it parses and extracts rules to extract the eGFR-estimated glomerular filtration rate (eGFR) of ≥45 ml / min / 1.73 m², prioritizing SGLT2 inhibitors (such as dapagliflozin); and generates a structured knowledge package with traceability information (source: Section 3.2.4 of the guidelines, updated in March 2024, credibility: authoritative release).

[0045] The logical reasoning verification agent constructs a reasoning chain based on a knowledge package: patient type (type 2 diabetes mellitus with chronic kidney disease) → applicable guidelines → medication recommendations and eGFR index requirements; the validity of the reasoning is verified using a medical causal verification model; if no questionable nodes are found, the reasoning chain with verification results is output.

[0046] S6: Generate and optimize the final answer. The large model integrates knowledge packages and reasoning chains to generate an initial answer; it then performs interpretability optimization to output the final answer, embedding source traceability annotations to visualize the reasoning process in the order of "premise → steps → conclusion".

[0047] S7: Process feedback and optimize the system. User feedback is assessed for accuracy, and guidelines cite authoritative positive feedback. The case library is updated, associating and storing question IDs, question texts, preprocessing results, invocation strategies, agent processing results, final answers, and positive labels. At the end of the month, once the case library reaches its data addition target, the fine-tuning process is initiated. The strategy generation module is fine-tuned using the PPO reinforcement learning algorithm, with the strategy matching accuracy for medical consultation questions as the core indicator; this improves the strategy matching accuracy for medical questions.

[0048] The above are merely embodiments of the present invention. The circuits, electronic components, and modules involved are all prior art, fully achievable by those skilled in the art, and require no further explanation. The scope of protection in this application does not involve improvements to the software and methods. Commonly known structures and characteristics in the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all prior art in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A question-answering method based on large models and data agents, characterized by: Includes the following steps: S1: Receive a natural language question input by the user, wherein the natural language question includes text form or text form converted by speech recognition; S2: Preprocess the natural language questions to generate a question dataset, which includes key entities, question types, and difficulty coefficients, where the difficulty coefficients are divided into five levels according to the complexity of the questions; S3: Input the question dataset from step S2 into the large model, and determine whether it can generate an accurate and complete answer based on the pre-trained knowledge. If it cannot answer, proceed to step S4. S4: Based on the problem dataset in step S2, dynamically generate an agent invocation strategy, including selecting the type of agent to be invoked and setting the execution order. Low-difficulty problems are executed in parallel, while high-difficulty problems are executed serially, and the weight of the logic reasoning verification agent is increased. S5: Based on the strategy, schedule and execute at least one of the following intelligent agents: real-time data acquisition, domain knowledge parsing, multi-source data fusion, or logical reasoning verification. Each intelligent agent returns a secondary answer with traceability information. S6: The large model integrates all secondary answers and its own knowledge to generate a primary answer, and outputs the final answer after interpretability optimization. The interpretability optimization includes embedding data source annotations or presenting a visualization of the reasoning process. S7: Update the case library based on user feedback, and use machine learning algorithms to optimize the agent calling strategy and adjust the weight of each agent based on the case library data. Optimize evaluation indicators including question answering accuracy, user satisfaction and calling efficiency, and set a target threshold for each indicator.

2. The question-answering method based on large models and data agents as described in claim 1, characterized in that: The preprocessing in step S2 involves using natural language processing techniques to segment the question, tag parts of speech, and identify named entities to extract key entities; using an intent classification model finely tuned with a domain question-answering dataset to determine the question type; and generating a difficulty coefficient based on a weighted scoring rule that considers question length, number of key entities, and logical complexity, where logical complexity is the number of derivation steps required to arrive at the target answer from known information.

3. The question-answering method based on large models and data agents as described in claim 1, characterized in that: In step S4, the agent is selected based on the problem type: for fact query, a real-time data acquisition agent is called; for domain consultation, a domain knowledge analysis agent is called; for numerical calculation, a real-time data acquisition agent and a multi-source data fusion agent are called; and for logical reasoning, a domain knowledge analysis agent and a logical reasoning verification agent are called. The combination of agents is adjusted according to the difficulty level.

4. The question-answering method based on large models and data agents as described in claim 1, characterized in that: The execution flow of the real-time data acquisition agent in step S5 is as follows: generating a structured search query based on the key entities in step S2; generating a structured search query based on the key entities; calling interfaces according to the priority order of official data source interface - authoritative database interface - trusted web page interface; cleaning the original data to remove duplicate and abnormal data, converting it into a structured format, and outputting it after attaching traceability information.

5. The question-answering method based on large models and data agents as described in claim 1, characterized in that: The execution flow of the domain knowledge parsing agent in step S5 is as follows: load the vertical domain knowledge base that can be automatically synchronized with the official updates according to the question type; locate the content based on key entities using keyword precise matching and semantic similarity comparison; parse and extract concept definitions, association rules, and case reference information, organize them into a structured knowledge package and output it with traceability information.

6. The question-answering method based on large models and data agents as described in claim 1, characterized in that: The execution process of the multi-source data fusion intelligent agent in step S5 involves receiving structured data and traceability information from at least two intelligent agents. Data is linked using entity alignment algorithms; Conflicts are resolved and weights are assigned based on the credibility of the data source; The data is integrated to form a unified structured dataset and output with a summary of source information.

7. The question-answering method based on large models and data agents as described in claim 1, characterized in that: The execution process of the logical reasoning verification agent in step S5 involves constructing a structured reasoning chain containing premises, reasoning steps, and conclusions based on the input data; applying symbolic logic rules or domain causal models to verify the steps; and supplementing evidence and marking the location, cause, and source information of the doubt when there is any doubt.

8. The question-answering method based on large models and data agents as described in claim 1, characterized in that: The interpretability optimization of step S6 is to include data source labeling, data acquisition time, and credibility. The reasoning process is visualized in the form of a structured reasoning chain.

9. The question-answering method based on large models and data agents as described in claim 1, characterized in that: In step S7, the case library is updated. Positive feedback stores the question ID, text, preprocessing result, calling strategy, secondary answer, final answer, and feedback tag together. Negative feedback analyzes the reasons as omission of intelligent agent calling, data tracing error, reasoning logic defect, or ambiguous answer expression, and stores the original question, analysis report, corrected strategy, answer, and tag as a new case and adds specific tags.

10. The question-answering method based on large models and data agents as described in claim 1, characterized in that: In step S7, fine-tuning is initiated according to a preset period or when the amount of new data in the case library reaches a threshold, using reinforcement learning or supervised learning algorithms. Fine-tuning requires meeting the following criteria: question-answering accuracy rate of no less than 90%, user satisfaction rate of no less than 85%, and intelligent agent invocation efficiency of ≤1 second for difficulty levels 1-2, ≤3 seconds for difficulty levels 3-4, and ≤5 seconds for difficulty level 5, with a compliance rate of no less than 90%. If the offline verification of the target question type strategy matching accuracy rate improves to a preset proportion, the result is solidified; if the core indicators do not improve or decrease, the parameters are rolled back.

Images