Task processing method, automatic question answering method, and computer-aided diagnosis method

Abstract

The embodiment of the application provides a task processing method, an automatic question answering method and a computer-aided diagnosis method, which comprises the following steps: obtaining an initial task processing result generated for a task processing request, wherein the initial task processing result is generated based on tool execution results of at least one tool; identifying a plurality of candidate reasoning information from the initial task processing result, wherein the plurality of candidate reasoning information comprises at least one to-be-traced reasoning information; determining a target reasoning evidence associated with the to-be-traced reasoning information from an evidence index dataset, wherein the evidence index dataset comprises a plurality of initial evidences, and the initial evidence comprises the tool execution result; and generating a target task processing result for responding to the task processing request, wherein the target task processing result comprises the plurality of candidate reasoning information and a traceability identifier set for the to-be-traced reasoning information, and the traceability identifier points to the target reasoning evidence. The application can realize point-level evidence tracing and support medical evidence-based auxiliary diagnosis.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a task processing method, an automatic question answering method, and a computer-aided diagnosis method, which can be applied to the field of medical question answering intelligent agents. Background Technology

[0002] While intelligent agents based on large-scale language models and tool calls possess powerful task execution capabilities, they still face technical challenges related to the credibility and interpretability of generated content. Specifically, existing intelligent agent systems lack effective mechanisms to verify the objective basis of generated content. On the one hand, this can lead to factual deviations or inconsistencies between the generated content and supporting evidence. On the other hand, users cannot easily verify or confirm the authenticity of the generated content. Summary of the Invention

[0003] In a first aspect, embodiments of this application provide a task processing method, comprising: obtaining an initial task processing result generated in response to a task processing request, the initial task processing result being generated based on the tool execution result of at least one tool; identifying multiple candidate inference information from the initial task processing result, the multiple candidate inference information including at least one inference information to be traced back to its source; determining target inference evidence that is associated with the inference information to be traced back to its source from an evidence index dataset, the evidence index dataset including multiple initial pieces of evidence, the initial pieces of evidence including the tool execution result; and generating a target task processing result for responding to the task processing request, the target task processing result including the multiple candidate inference information and a tracing identifier set for the inference information to be traced back to its source, the tracing identifier pointing to the target inference evidence.

[0004] Secondly, embodiments of this application provide an automatic question-answering method, comprising: obtaining an initial answer result generated for a question to be answered, the initial answer result being generated based on the tool execution result of at least one tool; identifying multiple candidate inference information from the initial answer result, the multiple candidate inference information including at least one inference information to be traced back to its source; determining target inference evidence that is associated with the inference information to be traced back to its source from an evidence index dataset, the evidence index dataset including multiple initial pieces of evidence, the initial pieces of evidence including the tool execution result; and generating a target answer result for responding to the question to be answered, the target answer result including the multiple candidate inference information and a tracing identifier set for the inference information to be traced back to its source, the tracing identifier pointing to the target inference evidence.

[0005] Thirdly, embodiments of this application provide a computer-aided diagnosis method, comprising: acquiring an initial medical diagnosis result generated in response to a medical diagnosis request, the initial medical diagnosis result being generated based on the tool execution result of at least one tool; identifying multiple candidate medical viewpoints from the initial medical diagnosis result, the multiple candidate medical viewpoints including at least one medical viewpoint to be traced; determining target inference evidence related to the medical viewpoint to be traced from an evidence index dataset, the evidence index dataset including multiple initial pieces of evidence, the initial pieces of evidence including the tool execution result; generating a target medical diagnosis result for responding to the medical diagnosis request, the target medical diagnosis result including the multiple candidate medical viewpoints and a traceability identifier set for the medical viewpoint to be traced, the traceability identifier pointing to the target inference evidence.

[0006] Fourthly, embodiments of this application provide a task processing system, including: a client for sending a task processing request; a server connected to the client for generating a target task processing result in response to the task processing request based on the task processing method described in the first aspect; the client is also used to output the target task processing result.

[0007] Fifthly, embodiments of this application provide a computer-aided diagnostic system, comprising: a client for sending a medical diagnostic request; a server connected to the client for generating a target medical diagnostic result in response to the medical diagnostic request based on the diagnostic method described in the third aspect; the client is also configured to output the target diagnostic result.

[0008] Sixthly, embodiments of this application provide an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor implements any of the methods of embodiments of this application when executing the computer program.

[0009] In a seventh aspect, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method of any one of the embodiments of this application.

[0010] Eighthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements any of the methods described in the embodiments of this application.

[0011] According to the technical solution of this application embodiment, by decomposing the initial task processing result into fine-grained reasoning information to be traced, each piece of reasoning information can be treated as an independent viewpoint for evidence tracing. This solves the technical problem that the generated content of traditional intelligent agents cannot be traced for evidence, enabling users to verify or confirm the information source of the reasoning information to be traced (such as specific factual judgments, logical inferences, or decision suggestions). Furthermore, it enables fine-grained evidence tracing at the viewpoint level. Additionally, since the target reasoning evidence originates from the execution results of the tool relied upon when generating the initial task processing result, each viewpoint (reasoning information to be traced) in the target task processing result can be matched to the original output content of the tool, thereby establishing an evidence tracing mechanism from viewpoint to the original tool execution result. This alleviates, to some extent, the illusion phenomenon caused by the lack of factual basis in large language models. Further, the generated target task processing result includes a tracing identifier pointing to the target reasoning evidence, allowing users to intuitively understand the target reasoning evidence corresponding to each viewpoint, thereby improving the credibility and interpretability of the target task processing result and meeting the stringent requirements of evidence-based decision-making in professional fields.

[0012] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application, it can be implemented according to the contents of the specification. In order to make the above and other objects, features and advantages of this application more easy to understand, specific embodiments of this application are given below. Attached Figure Description

[0013] In the accompanying drawings, unless otherwise specified, the same reference numerals throughout the various drawings denote the same or similar parts or elements. These drawings are not necessarily drawn to scale. It should be understood that these drawings depict only some embodiments according to this application and should not be construed as limiting the scope of this application.

[0014] Figure 1 This diagram illustrates the architecture of a task processing system 100 provided in an embodiment of this application.

[0015] Figure 2 This diagram illustrates an application example of the task processing system 100 provided in this application embodiment in an automatic question-and-answer scenario.

[0016] Figure 3A This diagram illustrates the architecture of a computer-aided diagnostic system 300 provided in an embodiment of this application.

[0017] Figure 3B The diagram illustrates an application example of the computer-aided diagnostic system 300 provided in the embodiments of this application.

[0018] Figure 4A flowchart illustrating the task processing method 400 provided in an embodiment of this application is shown.

[0019] Figure 5 A flowchart illustrating the task processing method 500 provided in an embodiment of this application is shown.

[0020] Figure 6 A flowchart of an automatic question-answering method 600 provided in an embodiment of this application is shown.

[0021] Figure 7 A flowchart of a computer-aided diagnostic method 700 provided in an embodiment of this application is shown.

[0022] Figure 8 and Figure 9 The diagram shows an application example of the computer-aided diagnostic method 700.

[0023] Figure 10 A block diagram of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0024] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the concept or scope of this application. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0025] To facilitate understanding of the technical solutions of the embodiments of this application, the relevant technologies of the embodiments of this application are described below. The following relevant technologies are optional solutions and can be combined with the technical solutions of the embodiments of this application in any way, and all of them fall within the protection scope of the embodiments of this application.

[0026] First, some of the terms or terms used in the embodiments of this application will be explained as follows.

[0027] Retrieval-Augmented Generation (RAG) is a technical architecture that improves the generation quality of large language models by combining the retrieval capabilities of external knowledge bases. In this technical process, the system first retrieves highly relevant document fragments or knowledge entries from large-scale external data sources based on the user's input question or instruction. Then, the retrieved contextual information and the original prompt words are input into the generative model.

[0028] Intelligent Agent: An intelligent agent is an artificial intelligence system capable of autonomous environmental perception, independent decision-making, and tool invocation. Unlike traditional models that passively respond to instructions, intelligent agents can understand complex high-level goals and autonomously break them down into a series of executable sub-task sequences. During execution, the intelligent agent can dynamically select and invoke various external tools (such as search engines, code interpreters, database interfaces, etc.) to obtain information or perform operations, and adjust its strategy in real time based on execution feedback.

[0029] Open-source Interactive Interface Framework: An open-source interactive interface framework is a large language model interaction system architecture that supports local deployment and multi-model access. This type of framework provides standardized front-end display and back-end service interfaces, allowing users to flexibly configure and connect to various underlying language model engines, while also supporting plugin extensions, session management, and multimodal data input.

[0030] Multimodal data refers to data sets that simultaneously contain two or more different forms of representation or perceptual dimensions. For example, multimodal data not only covers traditional text data, but also integrates various data types such as image data (e.g., medical images, pathological slide scans, and screenshots of examination reports), structured data (e.g., medical record forms, laboratory test numerical matrices, and time series monitoring data), and audio data (e.g., doctor-patient communication recordings and phonocardiograms).

[0031] Computed Tomography (CT) is a technique that uses precisely collimated X-ray beams, gamma rays, ultrasound, etc., along with highly sensitive detectors, to scan a specific part of the human body section by section, and then reconstructs the image using a computer.

[0032] Magnetic Resonance Imaging (MRI) is an imaging technique that uses signals generated by the resonance of atomic nuclei in a strong magnetic field to reconstruct images. It does not use ionizing radiation and has extremely high resolution for soft tissues (such as the brain, spinal cord, muscles, and ligaments). It is often used for detailed examination of the nervous system, joints, and abdominal organs.

[0033] Randomized controlled trials (RCTs) are the gold standard research experiments used in fields such as medicine, psychology, and social sciences to evaluate the effectiveness and safety of interventions (such as new drugs, new therapies, and new policies).

[0034] While intelligent agents based on large-scale language models and tool calls possess powerful task execution capabilities, they still face technical challenges related to insufficient credibility and interpretability of generated content. This is particularly true in the medical field, where, despite their vast medical knowledge and strong logical reasoning abilities, intelligent agents exhibit significant technical shortcomings in practical clinical applications. The primary issue is the risk of hallucinations. Research shows that such agents have a high rate of factual errors in medical scenarios, especially when dealing with critical information such as medication dosage and contraindications, where erroneous content could directly threaten patient safety. Secondly, these intelligent agent systems generally suffer from a lack of evidence; their output process often appears as a black box, preventing doctors from verifying the specific source of the answers. This lack of interpretability violates the core principles of Evidence-Based Medicine (EBM) and fails to meet regulatory requirements for the traceability and transparency of medical device algorithms. Therefore, constructing a medical intelligent processing solution that eliminates the risk of hallucinations and complies with the principles of evidence-based medicine has become a critical technical problem that urgently needs to be solved.

[0035] To address the aforementioned technical issues, the first type of solution employs a general search-enhanced generation (RAG) approach. This approach typically transforms medical guidelines, textbooks, and other documents into vector databases, retrieving relevant paragraphs during the question-and-answer process as contextual input models. However, its retrieval granularity is relatively coarse, usually remaining at the document or paragraph level, failing to precisely map search results to specific viewpoints within the answers. Furthermore, this approach relies heavily on pre-built static knowledge bases, failing to effectively utilize dynamic new information generated during the conversation. In addition, this type of solution lacks a mechanism for differentiating levels of evidence; for example, it cannot identify the differences in credibility between evidence from different sources such as randomized controlled trials (RCTs), meta-analyses, or case reports, resulting in output results lacking rigorous evidence-based support.

[0036] The second type of solution is the ex-post citation approach. This approach primarily employs a generation-then-retrieve strategy, where a large language model first generates the complete answer, then an independent search is performed, and relevant links are appended to the end of the answer. The fundamental flaw of this approach lies in the logical separation between the generation and retrieval processes, making it highly susceptible to inconsistencies between the answer description and the cited literature. Furthermore, the citation of relevant links is in the form of entire paragraphs or articles appended, lacking fine-grained correspondences. Doctors must manually determine which statement corresponds to which article, resulting in poor clinical practicality. Especially when facing complex medical problems involving multiple sub-propositions, this approach cannot achieve separate tracing of the source of each independent viewpoint.

[0037] The third type of solution relies on traditional Clinical Decision Support Systems (CDSS). These systems primarily use rule engines or knowledge graphs to derive conclusions according to pre-defined logical chains. Their limitations include limited coverage, only handling pre-defined scenarios and failing to address open-ended consultation needs. Furthermore, due to reliance on manual maintenance, their knowledge base updates are often outdated, making it difficult to keep up with the latest literature (such as clinical guidelines). In terms of user experience, these systems lack natural language dialogue capabilities, have rigid interaction methods, and are ill-suited to the flexibility and intelligence requirements of smart healthcare scenarios.

[0038] In summary, none of the relevant technical solutions have achieved precise claim-level tracing. Therefore, this application aims to provide an agent-based task processing scheme, an automatic question-answering scheme, and a computer-aided diagnostic scheme. By performing fine-grained claim-level evidence tracing on generated content, it addresses the technical deficiency of logical separation between key information and evidence. A detailed description follows.

[0039] To facilitate understanding, let's first combine Figure 1 This paper introduces the application scenarios of the technical solution presented in this application. Figure 1 This diagram illustrates the architecture of a task processing system 100 provided in an embodiment of this application. Figure 1 As shown, the task processing system 100 includes a client 101 and a server 102.

[0040] The client 101 is used to send task processing requests. For example, a user can input natural language commands or upload files containing multimodal data through the user interface provided by the client 101, thereby triggering a task processing request. In this embodiment, the task processing request includes, but is not limited to, requests for answers to questions, medical diagnoses, legal advice, and complex data analysis.

[0041] An intelligent agent is deployed on the server 102. This intelligent agent receives task processing requests from the client 101 and is used to implement the method 400 or method 500 provided in the embodiments of this application. For example, in response to the task processing request, the intelligent agent invokes at least one tool and generates an initial task processing result based on the tool's execution result. Multiple local or fine-grained candidate inference information are identified from the macroscopic or global initial task processing result. One or more of these candidate inference information can be used as inference information to be traced (such as key conclusions or opinions). That is, some or all of the inference information to be traced can be used as candidate inference information. The inference information to be traced is the object of subsequent evidence tracing.

[0042] Furthermore, the agent will identify target reasoning evidence that is related to the reasoning information to be traced from the evidence index dataset, which includes multiple initial pieces of evidence. The target reasoning evidence serves as the factual basis supporting the reasoning information to be traced. The initial evidence in the evidence index dataset includes the tool execution results. Then, the agent will generate a target task processing result to respond to the task processing request. The target task processing result not only retains the reasoning result constructed from multiple candidate reasoning information, but also embeds a traceability identifier pointing to the target reasoning evidence.

[0043] Furthermore, the client 101 outputs the result of the target task processing, enabling users to directly trace the factual basis of each piece of reasoning information to be traced, thereby improving the credibility and interpretability of the result of the target task processing.

[0044] In an optional application example, the evidence index dataset in this embodiment is a dynamic session-based indexing. Specifically, the initial evidence in the evidence index dataset includes not only the tool execution results but also the session context data corresponding to the task processing request. The session context data includes, but is not limited to, user-uploaded data and historical session text. User-uploaded data can be multimodal data; historical session text refers to session interaction records generated before the current session round. In other words, this evidence index dataset is a dynamically constructed index dataset for the current task processing request. It can capture user-uploaded data (such as data uploaded by the user in real-time) and historical session text during the inference process in real time during multiple rounds of interaction in the task processing request, effectively overcoming the problem that traditional methods relying on pre-built static knowledge bases cannot effectively utilize the dynamic new information generated during the session.

[0045] Figure 2 This diagram illustrates an application example of the task processing system 100 provided in this embodiment of the application in an automatic question-answering scenario. For example... Figure 2 As shown, users can input questions to be answered based on the user interface provided by client 101. The questions to be answered can come from different scenarios, such as subject knowledge Q&A, professional field consultation, itinerary planning, code debugging, etc. This application embodiment does not specifically limit them. For example, the question to be answered is "Based on the sales data of the past five years, analyze the market trend of a certain electronic product in the third quarter".

[0046] An agent is deployed on server 102. Upon receiving the question to be answered, the agent generates an initial solution by invoking at least one tool (such as a data statistical analysis engine, a historical database query interface, or a trend prediction model). Then, the agent performs semantic parsing on the initial solution to identify multiple candidate inference information, such as candidate inference information A, candidate inference information B, candidate inference information C, and candidate inference information D. Among them, candidate inference information C and candidate inference information D are the source objects that need to be traced; that is, candidate inference information C and candidate inference information D are the inference information to be traced.

[0047] Furthermore, the agent performs retrieval and matching from the evidence index dataset to determine the target inference evidence C1 that has a semantic relationship with the inference information C to be traced, such as the original sales statistics report or calculation log output by the historical database query interface; and the target inference evidence D2 that has a semantic relationship with the inference information D to be traced, such as the prediction curve data generated by the trend prediction model.

[0048] The target solution generated by the agent fully retains multiple candidate reasoning information, including candidate reasoning information A, candidate reasoning information B, reasoning information to be traced C, and reasoning information to be traced D, and also includes a tracing identifier set for reasoning information to be traced C. Figure 2 The "tracing identifier 1" in the text, and the tracing identifier set for the information D to be traced ( Figure 2 (Referring to "Tracing Identifier 2" in the context of the target inference evidence). Tracing Identifier 1 points to target inference evidence C1. By clicking Tracing Identifier 1 on the user interface provided by client 101, users can obtain evidence details for target inference evidence C1, such as evidence summary, evidence level, and evidence type. Tracing Identifier 2 points to target inference evidence D2. By clicking Tracing Identifier 2 on the user interface provided by client 101, users can obtain evidence details for target inference evidence D2, allowing them to intuitively view the specific data support or factual basis behind each piece of inference information to be traced in the target solution result.

[0049] The following is combined Figure 3A and Figure 3B This paper introduces another application scenario of the technical solution of this application. Figure 3A This diagram illustrates the architecture of a computer-aided diagnostic system 300 provided in an embodiment of this application. Figure 3A As shown, the computer-aided diagnostic system 300 includes a client 301 and a server 302.

[0050] The client 301 is used to send medical diagnosis requests. For example, a user can input a description of the patient's chief complaint, upload medical imaging data (such as CT images, MRI scans), and test reports through the user interface provided by the client 301, thereby triggering a medical diagnosis request. An intelligent agent is deployed on the server 302, which receives the medical diagnosis request from the client 301 and is used to implement the method 700 provided in this embodiment.

[0051] For example, in response to a medical diagnosis request, the agent invokes at least one tool and generates an initial medical diagnosis result based on the tool's execution result. From the macroscopic or global initial medical diagnosis result, multiple local or fine-grained candidate medical viewpoints are identified, one or more of which can serve as the source medical viewpoint to be traced. Further, the agent identifies target inference evidence related to the source medical viewpoint from an evidence index dataset containing multiple initial pieces of evidence, serving as factual basis supporting that viewpoint. The initial evidence in the evidence index dataset includes the tool execution result. The agent then generates a target medical diagnosis result to respond to the medical diagnosis request. This target medical diagnosis result not only retains the inference result constructed from multiple candidate medical viewpoints but also embeds a source identifier pointing to the target inference evidence. Finally, the client outputs this target medical diagnosis result, allowing the user to directly trace the factual basis of each source medical viewpoint, improving the credibility and interpretability of the target medical diagnosis result.

[0052] Figure 3B This diagram illustrates an application example of the computer-aided diagnostic system 300 provided in an embodiment of this application. For example... Figure 3B As shown, users can perform operations on the user interface provided by client 301, such as providing multimodal data: uploading CT images and test reports, and entering specific clinical consultation content in the text query area. This can trigger a medical diagnosis request.

[0053] The medical diagnosis request is sent to server 302, where it is received and processed by an agent deployed on the server. The agent first invokes at least one tool from the toolset based on the content of the medical diagnosis request, including but not limited to medical literature retrieval tools, medication dosage calculation tools, image feature extraction tools, and clinical guideline query tools. The agent then obtains the execution results of each tool and generates an initial medical diagnosis based on these results.

[0054] Furthermore, the agent performs semantic parsing on the initial medical diagnosis results, identifying multiple semantically complete and independent candidate medical viewpoints, such as diagnostic conclusions, treatment recommendations, and prognostic analysis results. The agent identifies candidate medical viewpoints containing key factual judgments or data inferences as the medical viewpoints to be traced back to. For example, in this application example, each candidate medical viewpoint (including diagnostic conclusion 10, treatment recommendation 11, treatment recommendation 12, treatment recommendation 13, and prognostic analysis result 14) is treated as an independent medical viewpoint to be traced back to.

[0055] Furthermore, the agent retrieves target reasoning evidence from the evidence index dataset that is related to the medical viewpoint to be traced. The initial evidence in the evidence index dataset encompasses the tool execution results generated during tool execution, such as retrieved medical literature or excerpts from medical literature, formulas or drug interaction records on which drug dosage calculations are based, image features extracted from medical images, and the original text of queried guideline clauses. In an optional example, the evidence index dataset can be a dynamic index dataset, meaning the initial evidence in the evidence index dataset can also include session context data (including user-uploaded data and historical session text), such as multimodal data like newly uploaded CT images and test reports during the agent's generation of the final target medical diagnosis.

[0056] For example, such as Figure 3B As shown, the agent performs retrieval and matching from the evidence index dataset: the target inference evidence associated with diagnosis conclusion 10 is the relevant clause on white blood cell count threshold and bacterial infection probability in the "Guidelines for the Diagnosis and Treatment of Community-Acquired Pneumonia in Adults" output by the medical literature retrieval tool; the target inference evidence associated with treatment opinion 11 is the authoritative antibiotic recommendation entry returned by the clinical guideline query tool; the target inference evidence associated with treatment opinion 12 is the liver and kidney function metabolic risk assessment report generated by the drug dosage calculation tool; the target inference evidence associated with treatment opinion 13 is the record of incompatible combinations of cephalosporins and other drugs extracted from the drug interaction database; and the target inference evidence associated with prognostic analysis result 14 is the 72-hour symptom relief rate curve obtained from the statistical analysis of similar case follow-up data in the historical session context.

[0057] The intelligent agent binds the target reasoning evidence with the corresponding medical viewpoint to be traced, generating a target medical diagnosis result containing a traceability identifier. For example, the target medical diagnosis result is returned to client 301 for display. For instance, as... Figure 3BAs shown, the target medical diagnosis result generated by the agent fully retains five candidate medical viewpoints, namely diagnosis conclusion 10, treatment opinion 11, treatment opinion 12, treatment opinion 13 and prognosis analysis result 14, and sets corresponding traceability identifiers for them:

[10] ,

[11] ,

[12] ,

[13] ,

[14] .

[0058] Among them, the source identification

[10] points to the target reasoning evidence supporting the diagnostic conclusion 10. When the user clicks the source identification

[10] based on the user interface provided by the client 301, the user can obtain the evidence details supporting the diagnostic conclusion 10, namely the summary of the relevant clauses on the white blood cell count threshold and the probability of bacterial infection in the "Guidelines for the Diagnosis and Treatment of Community-Acquired Pneumonia in Adults", the level of evidence (such as high confidence level) and the type of evidence (such as clinical practice guidelines); the source identification

[11] points to the target reasoning evidence supporting the treatment opinion 11. When the user clicks the source identification

[11] , the user can obtain the evidence details supporting the treatment opinion 11; and so on.

[0059] This allows users to intuitively view the specific data support or factual basis behind each traceable medical viewpoint in the target medical diagnosis results, realizing the traceability, verifiability, and auditability of the entire diagnostic process, effectively improving the interpretability and credibility of computer-aided diagnostic results, and enhancing the transparency and clinical credibility of the medical intelligent agent system.

[0060] It is understood that the other components of the aforementioned task processing system, computer-aided diagnostic system, client, and server can employ various technical solutions now and in the future known to those skilled in the art, and will not be described in detail here. Furthermore, the deployment form or other components of the aforementioned task processing system, computer-aided diagnostic system, client, and server may differ in different application scenarios, and the embodiments of this application do not specifically limit them in this regard.

[0061] For example, a connection is established between the client and the server via a network. The network provides the medium for the communication link between the client and the server. Networks can include various connection types, such as wired, wireless communication links, or fiber optic cables. Data transmitted by the client can be processed such as encoding, transcoding, and compression before being sent to the server. There can be multiple clients; users can interact with the server through their clients to receive data sent by other clients or send data to other clients.

[0062] For example, the client can be hardware, such as mobile phones, personal computers, tablets, wearable devices, and other electronic devices; it can also be a browser, application (APP), web application, mini-program, cloud application, or microservice deployed on the electronic device. The server can include servers that provide various services, such as servers that provide communication services to clients, servers that support models used on clients, and servers that process data sent by clients. The server can be deployed as a distributed server cluster consisting of multiple servers, or it can be implemented as a single server, which can be a physical server or a cloud server.

[0063] It should be noted that the application scenarios or examples provided in this application are for ease of understanding, and this application does not specifically limit the application of the technical solutions. Furthermore, the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. The collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0064] The technical solution of this application and how it solves the aforementioned technical problems are described in detail below with specific embodiments. The listed specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0065] Figure 4 A flowchart illustrating a task processing method 400 according to an embodiment of this application is provided. This method 400 can be applied to a server, for example, by... Figure 1 The server 102 shown is executed. Figure 4 As shown, the method 400 may include steps S401 to S404.

[0066] Step S401: Obtain the initial task processing result generated in response to the task processing request. The initial task processing result is generated based on the tool execution result of at least one tool.

[0067] A task processing request is an instruction initiated by a user that requires the assistance of an intelligent agent to complete a complex task in a specific domain. For example, a user might input natural language instructions and / or upload multimodal data through a user interface provided by the client, thus triggering a task processing request. The server then receives the task processing request from the client. Task processing requests include, but are not limited to, requests for answers to unanswered questions, medical diagnoses, legal advice, and complex data analysis.

[0068] Upon receiving a task processing request, the server-side agent dynamically invokes at least one tool from the toolset (such as retrieval tools, calculation tools, feature extraction tools, analysis tools, etc.) to perform corresponding operations based on the specific content of the request, and collects the execution results returned by each tool (such as document excerpts, calculated values, feature extraction results, data charts, etc.). Subsequently, the agent performs semantic parsing, logical reasoning, and content generation on the execution results of one or more tools, thereby forming a preliminary response containing conclusive statements or suggestions. This response constitutes the initial task processing result.

[0069] Step S402: Identify multiple candidate inference information from the initial task processing results, including at least one inference information to be traced.

[0070] Candidate inference information refers to multiple local or fine-grained candidate inference information identified from the initial task processing results, such as specific factual judgments, logical inferences, or decision suggestions. One or more of these candidate inference information can serve as source-tracing inference information (such as key conclusions or viewpoints). That is, some or all of the source-tracing inference information can serve as candidate inference information. Source-tracing inference information is the source object for subsequent evidence tracing, such as information involving key factual basis, data support, or professional standard citations, which needs to show its source to the user to verify or confirm its credibility. For example, the intelligent agent uses natural language processing technology to structurally divide the initial task processing results, breaking them down into several independent semantic paragraphs or sentences as candidate inference information.

[0071] This application does not specifically limit the ratio of candidate reasoning information to reasoning information to be traced. In one example, some candidate reasoning information can be used as reasoning information to be traced. That is, based on preset traceability requirement rules (such as filtering objective arguments containing numerical values, diagnostic conclusions, regulations, etc., and excluding subjective assertions), candidate reasoning information that needs to be traced can be selected as reasoning information to be traced. For example, in a medical diagnosis scenario, only statements involving specific disease diagnosis conclusions, medication dosage recommendations, and prognostic assessments are identified as reasoning information to be traced, while general greetings, background introductions, or subjective assertions are excluded from the traceability scope. In another example, each candidate reasoning information can be used as an independent reasoning information to be traced. For example, in a legal opinion consultation scenario, each interpretation of a legal clause, each case analogy, and the final legal advice in the initial task processing results are considered as reasoning information to be traced, so as to achieve full-chain evidence traceability.

[0072] Step S403: Identify target reasoning evidence that is related to the reasoning information to be traced from the evidence index dataset. The evidence index dataset includes multiple initial pieces of evidence, including the tool execution results.

[0073] The evidence index dataset contains multiple initial pieces of evidence. Target inference evidence refers to initial evidence that has a direct semantic support, derivation, or citation relationship with the inference information to be traced back to its source. This initial evidence provides factual basis, data sources, or normative references for the inference information to be traced back to, thereby proving its accuracy and reliability. In this embodiment, the initial evidence includes tool execution results. That is, for the inference information to be traced back to its source, one can directly trace back to the specific output content (such as specific literature paragraphs, calculation formulas, database records, etc.) generated by the tool called when the initial task processing result was produced, ensuring the authenticity and verifiability of the evidence source.

[0074] Among these, the relationships include, but are not limited to, semantic implication relationships, logical deduction relationships, data support relationships, and normative reference relationships. For example, determining the target reasoning evidence that is related to the reasoning information to be traced from the evidence index dataset can be achieved through an evidence retrieval algorithm, that is, based on the extracted reasoning information to be traced, a search is performed in the evidence index dataset composed of multiple initial pieces of evidence.

[0075] Step S404: Generate a target task processing result for responding to the task processing request. The target task processing result includes multiple candidate inference information and a tracing identifier set for the inference information to be traced. The tracing identifier points to the target inference evidence.

[0076] The target task processing result includes multiple candidate inference information, meaning that the target task processing result can completely retain the initial task processing result to ensure the semantic integrity of the generated content. Furthermore, the target task processing result also includes a source tracing identifier; that is, for a specific inference information to be traced, a source tracing identifier is set pointing to the target inference evidence. This allows users to intuitively view the target inference evidence supporting the information to be traced when viewing the target task processing result by triggering the source tracing identifier.

[0077] According to the technical solution of this application embodiment, by decomposing the initial task processing result into fine-grained reasoning information to be traced, each piece of reasoning information can be treated as an independent viewpoint for evidence tracing. This solves the technical problem that the generated content of traditional intelligent agents cannot be traced for evidence, enabling users to verify or confirm the information source of the reasoning information to be traced (such as specific factual judgments, logical inferences, or decision suggestions). Furthermore, it enables fine-grained evidence tracing at the viewpoint level. Additionally, since the target reasoning evidence originates from the execution results of the tool relied upon when generating the initial task processing result, each viewpoint (reasoning information to be traced) in the target task processing result can be matched to the original output content of the tool, thereby establishing an evidence tracing mechanism from viewpoint to the original tool execution result. This alleviates, to some extent, the illusion phenomenon caused by the lack of factual basis in large language models. Further, the generated target task processing result includes a tracing identifier pointing to the target reasoning evidence, allowing users to intuitively understand the target reasoning evidence corresponding to each viewpoint, thereby improving the credibility and interpretability of the target task processing result and meeting the stringent requirements of evidence-based decision-making in professional fields.

[0078] In one implementation, the target task processing result includes a target response text. In step S404, generating a target task processing result for responding to a task processing request may include: constructing an initial response text using multiple candidate inference information; determining the insertion position of the tracing identifier in the initial response text; and inserting the tracing identifier into the insertion position to obtain the target response text.

[0079] The initial response text is a coherent text formed by piecing together and refining multiple candidate inference information in a natural language logical order. This initial response text constitutes a preliminary framework for answering the task processing request, but it does not yet include source identifiers pointing to specific evidence. For example, the identified multiple candidate inference information can be used as semantic nodes. A natural language generation model can then logically connect these semantic nodes using conjunctions and transitional sentences to form a coherent paragraph or passage, thus obtaining the initial response text. The target response text refers to the text generated after embedding source identifiers in a non-intrusive visual form (such as badges, highlighted links, etc.) into the initial response text. This target response text is the final response content presented to the user.

[0080] For example, the syntactic structure of the initial response text can be analyzed, and the insertion position can be automatically locked based on semantic similarity calculation. For instance, when the agent generates the source inference information "the patient's white blood cell count is elevated, indicating bacterial infection", the insertion position is calculated to be between "the patient's white blood cell count is elevated, indicating bacterial infection" and the period through syntactic structure and semantic matching. The source identifier of the target inference evidence, such as the superscript "

[411] ", is automatically inserted at this insertion position, thereby generating the target response text containing "the patient's white blood cell count is elevated, indicating bacterial infection

[411] .", where the superscript "

[411] " points to Article 3.2 of the "Guidelines for the Diagnosis and Treatment of Pneumonia", which is the target inference evidence.

[0081] In one application example, in response to a triggered action on the source identification icon, such as a user clicking or hovering over the icon on the user interface, evidence details of the target reasoning can be displayed. The evidence details include at least one of the following: evidence summary, evidence level, and evidence type.

[0082] In this context, "evidence summary" refers to a concise description of the target reasoning evidence. For example, it's a simplified description of key information extracted from initial evidence (such as long documents or complex reports) and summarized using natural language, allowing users to grasp the core basis without reading the full text of the target reasoning evidence. "Evidence level" is an indicator used to quantitatively evaluate the credibility and authority of the target reasoning evidence. For example, it can be determined based on preset evidence grading rules, such as Grade A / Strong Recommendation, Grade B / Moderate Recommendation, etc. "Evidence type" refers to the data type or document attribute classification of the target reasoning evidence. Evidence types can be differentiated according to different application scenarios. For example, in a medical intelligent question-answering scenario, evidence types include, but are not limited to: research types (such as RCTs, prospective cohort studies, case-control studies, meta-analyses, etc.), clinical guidelines, drug instructions, real-time computing logs, or database records, etc.

[0083] Based on this implementation method, by utilizing multiple candidate inference information to construct the initial response text, the semantic coherence and logical integrity of the target task processing results are ensured; by determining the insertion position of the tracing identifier in the initial response text and inserting the tracing identifier at the insertion position, users can intuitively and accurately identify which information has undergone evidence tracing when reading the target response text, thereby improving the user experience.

[0084] In one implementation, in step S402, identifying multiple candidate inference information from the initial task processing result may include: semantically decomposing the initial task processing result, extracting multiple independent atomic semantic units, and obtaining multiple candidate inference information, with each candidate inference information corresponding to an atomic semantic unit.

[0085] Among them, an atomic semantic unit refers to a declarative proposition with independent semantic integrity that is automatically identified and extracted from unstructured text data. Each atomic semantic unit can represent a complete semantic, such as a clear statement of fact, conclusion or hypothesis, without containing redundant background descriptions, irrelevant modifiers or other information.

[0086] For example, semantic decomposition of the initial task processing results can be performed using an opinion extraction algorithm. This algorithm automatically identifies and extracts declarative propositions with independent semantic integrity from unstructured text data, thereby breaking down long documents or complex paragraphs into several atomic semantic units. For instance, a description in the initial task processing results, "Given that several recent randomized controlled trials have shown that compound X can significantly inhibit the activity of enzyme Y, and no serious side effects have been observed in clinical trials, it is recommended that compound X be used as a first-line treatment," can be transformed into three independent atomic semantic units through fine-grained semantic decomposition (opinion extraction): the first atomic semantic unit is "compound X has the effect of inhibiting the activity of enzyme Y," the second atomic semantic unit is "compound X has not been observed to have serious side effects in clinical trials," and the third atomic semantic unit is "it is recommended that compound X be used as a first-line treatment."

[0087] Based on this implementation method, by semantically decomposing the initial task processing results and extracting multiple independent atomic semantic units, redundant text descriptions can be transformed into structured knowledge nodes, providing a standardized input basis for subsequent evidence tracing, thereby improving the precision and recall rate of evidence matching.

[0088] In one implementation, in step S403, determining the target reasoning evidence that is related to the reasoning information to be traced from the evidence index dataset may include: determining at least one candidate reasoning evidence that is related to the reasoning information to be traced from multiple initial pieces of evidence; in response to the candidate reasoning evidence including positive evidence and negative evidence, obtaining the positive evidence and negative evidence to obtain the target reasoning evidence, and generating a dispute status marker, which is used to identify the target reasoning evidence as evidence with conflict and dispute.

[0089] Candidate reasoning evidence refers to the initial set of evidence retrieved from the evidence index dataset that is related to the reasoning information to be traced. If the candidate reasoning evidence includes both positive and negative evidence—that is, if there is positive evidence supporting the validity of the reasoning information to be traced, and negative evidence refuting the validity of the reasoning information or proposing a contrary conclusion—then it can be determined that there are multiple target reasoning pieces of evidence, including both positive and negative evidence.

[0090] Furthermore, a dispute status marker is generated. The dispute status marker is used to identify the target reasoning evidence as evidence that is in conflict or dispute. That is, in the result of the target task processing, a specific metadata label or visual indicator is attached to the reasoning information to be traced as a dispute status marker to prompt the user that the target reasoning evidence is evidence that is in conflict or dispute.

[0091] For example, when the information to be inferred is "Drug X can be used to treat disease Y", the evidence retrieval module retrieves multiple articles from the evidence index dataset as candidate inference evidence. Among them, a randomized controlled trial report published in 2025 (positive evidence) states that "Drug X can significantly improve the clinical symptoms of disease Y", while another systematic review published at the same time (negative evidence) states that "existing data are insufficient to support the efficacy of drug X for disease Y, and there are potential side effects". At this time, the system identifies both articles as target inference evidence and generates a dispute status marker. In the final target task processing result, a dispute status marker will be displayed next to "Drug X can be used to treat disease Y". Users can click on the dispute status marker to view the positive evidence supporting the efficacy and the negative evidence questioning the efficacy, respectively.

[0092] This implementation provides an evidence conflict detection mechanism that can automatically identify and mark the disputed state when multiple initial pieces of evidence lead to opposite conclusions or have substantial differences regarding the same source information to be traced, enabling users to make independent judgments or evidence-based decisions based on complete evidence from both sides.

[0093] Figure 5 A flowchart illustrating a task processing method 500 according to an embodiment of this application is provided. This method 500 can be applied to a server, for example, by... Figure 1 The server 102 shown is executed. Figure 5 As shown, the method 500 may include steps S501 to S506.

[0094] Step S501: Generate multiple subtasks based on the task processing request and initialize the inference results; Step S502: Perform at least one round of iterative processing on multiple subtasks until all subtasks have been completed. Each round of iterative processing includes: generating the current subtask to be processed based on the previous inference result and the unfinished subtasks; calling at least one tool to execute the current subtask to be processed; and generating the current inference result based on the tool execution result returned by at least one tool. Step S503: In response to the completion of multiple subtasks, generate the initial task processing result based on the current inference result.

[0095] Step S504: Identify multiple candidate inference information from the initial task processing results, among which at least one inference information to be traced is included; Step S505: Identify target reasoning evidence that is related to the reasoning information to be traced from the evidence index dataset. The evidence index dataset includes multiple initial pieces of evidence, including the tool execution results. Step S506: Generate a target task processing result for responding to the task processing request. The target task processing result includes multiple candidate inference information and a tracing identifier set for the inference information to be traced. The tracing identifier points to the target inference evidence.

[0096] Step S504 can be implemented in the same or similar way as step S402, step S505 can be implemented in the same or similar way as step S403, and step S506 can be implemented in the same or similar way as step S404, and achieve the corresponding technical effects, which will not be elaborated here.

[0097] In step S501, the agent can perform the first round of reasoning for the task processing request, such as using a large language model to perform in-depth analysis of the semantic intent, constraints and required knowledge domain of the task processing request, thereby decomposing the complex task processing request into multiple sub-tasks that are logically interdependent or parallel; at the same time, the reasoning result is initialized, for example, the reasoning result is initialized as an empty string, an empty list or an initial context state containing background information of the task processing request.

[0098] In step S502, at least one round of iterative processing is performed on multiple subtasks until all subtasks are completed. Each round of iterative processing includes: generating a currently pending subtask based on the previous reasoning result and the unfinished subtasks; invoking at least one tool to execute the currently pending subtask; and generating a current reasoning result based on the tool execution result returned by at least one tool. For example, this process adopts the Reasoning Acting (ReAct) paradigm, which is a large language model example that tightly intertwines logical reasoning processes with external action execution. Specifically, in each iteration, the agent first executes the "Thought" phase, analyzing the current state and planning the next strategy based on the current inference result generated in the previous iteration and the list of remaining unfinished subtasks, generating the current subtask to be processed. Then, it enters the "Action" phase, where the agent selects and calls the appropriate tools to execute specific operations based on the current subtask. Finally, it enters the "Observation" phase, where the agent receives the tool execution results returned by at least one tool, integrates them as new factual evidence into the previous inference result, corrects or supplements the original inference path, and constructs the current inference result. By constructing this iterative mechanism, the large language model can perform logical deduction before generating the final answer, then call external tools to obtain real-time feedback, and dynamically adjust the subsequent subtask generation and inference path based on the feedback results until all subtasks are completed.

[0099] The process of invoking at least one tool to execute the currently pending subtask can include: dynamically selecting and invoking a matching tool based on the task type of the currently pending subtask. For example, if the currently pending subtask is an information retrieval task, a retrieval tool (such as a search engine interface, medical literature retrieval tool, clinical guideline query tool, etc.) is invoked to search for the specified keywords or query conditions in the currently pending subtask, obtaining the tool's execution results, such as unstructured text data including relevant literature abstracts, web page fragments, or database records. If the currently pending subtask contains image data, a visual analysis tool (such as an optical character recognition engine, image feature extraction model, etc.) is invoked to convert the image data into semantic text description data. For example, a user-uploaded CT image is converted into a text description such as "The image shows a patchy high-density shadow in the lower lobe of the left lung with blurred edges." If the current subtask to be processed contains structured data, then a data parsing tool (such as a table conversion plugin, data extraction script, medication dosage calculation tool, etc.) is invoked to convert the structured data into text data. For example, an Excel table containing the patient's various biochemical indicators is converted into a text description such as "The patient's alanine aminotransferase value is 85 U / L, which exceeds the upper limit of the normal reference range" so that the large language model can perform subsequent semantic understanding and reasoning.

[0100] In step S503, in response to the completion of multiple subtasks, an initial task processing result is generated based on the current inference result. Specifically, when the iterative processing loop terminates, that is, when all the decomposed subtasks have been executed and the corresponding tool execution results have been integrated into the current inference result, the agent performs final logical sorting, deduplication, polishing, and formatting encapsulation on the current inference result to generate a coherent and complete natural language text or multimodal data report, thereby obtaining the initial task processing result, which is the complete response content before the addition of the source identification.

[0101] According to the technical solution of this application, by introducing a multi-round iterative mechanism based on the ReAct paradigm during task processing, dynamic interaction between the generation of multiple sub-tasks, tool invocation, and the generation of current inference results is achieved. This avoids the illusion accumulation problem that may occur in traditional chain-of-thought due to the lack of real-time feedback. Because tool invocation is embedded in the iterative processing, the inference process and the acquisition of tool execution results are synchronized. The generation of each current inference result directly depends on the actual tool execution result returned by the tool in the preceding step, thereby fundamentally avoiding the mismatch between the target task processing result and the target inference evidence, and eliminating the uncertainty and bias of retrospectively searching for evidence.

[0102] In one implementation, method 500 may further include: storing the tool execution results generated in each round of iterative processing as initial evidence in the evidence index dataset, and storing the evidence details corresponding to the tool execution results in the evidence index dataset after associating them with the corresponding initial evidence.

[0103] The evidence details include at least one of the following: evidence summary, evidence level, and evidence type, as detailed above.

[0104] For example, after the "observation" phase of each iteration, at least one tool execution result returned by the tool is captured, assigned a unique evidence identifier, and stored in the evidence index dataset as a key-value pair or document object. This allows the tool execution results generated in each iteration to be stored as initial evidence in the evidence index dataset. Further, the evidence details corresponding to the tool execution results are associated with the corresponding initial evidence and stored in the evidence index dataset. For instance, for tool execution results retrieved from a medical guideline database, their core conclusions are automatically extracted as an evidence summary. Based on the authority of the guideline's issuing institution, their evidence level is marked as "Level A / Strong Recommendation," and their evidence type is marked as "Clinical Guideline." These metadata elements are then bound and stored with the original content of the tool execution results to form structured evidence details.

[0105] Based on this implementation method, a dynamic evidence index dataset synchronized with the reasoning process is constructed by capturing and storing the tool execution results and their associated evidence details in real time during the iterative processing. At the same time, the pre-stored evidence details (such as evidence level and evidence type) can enable users to quickly verify or validate the credibility of their viewpoints (inference information to be traced back to the source).

[0106] In one implementation, there are multiple tool execution results. The current inference result is generated based on the tool execution result returned by at least one tool, including: determining the evidence level of each tool execution result based on a preset evidence grading rule; and generating the current inference result based on each tool execution result and its corresponding evidence level.

[0107] The evidence grading rules can be preset according to the needs of the application scenario. For example, in a legal consultation scenario, the evidence grading rules can be configured as follows: legal provisions promulgated by national legislative bodies and judicial interpretations issued by the Supreme People's Court are set as the highest level of evidence (e.g., Level 1 evidence); local regulations or institutional rules are set as the second-lower level of evidence (e.g., Level 2 evidence); and legal practice articles or online legal Q&As are set as the lowest level of evidence (e.g., Level 3 evidence), to ensure that the authority of the reasoning basis matches the level of legal effect. As another example, in a computer-aided diagnosis scenario, the evidence grading rules can be configured as follows: internationally recognized clinical diagnostic guidelines, systematic reviews, and meta-analyses are set as high level of evidence (e.g., Level A evidence); the results of randomized controlled trials (RCTs) are set as the second-lower level of evidence (e.g., Level B evidence); expert consensus, case reports, or basic medical textbook content are set as the third-lower level of evidence (e.g., Level C evidence); and information from non-professional medical websites or unverified online rumors are set as the lowest level of evidence or not accepted, to meet the strict standards of evidence-based medicine.

[0108] Furthermore, the current inference result is generated based on the execution results of each tool and their corresponding evidence levels. For example, when there are content conflicts between the execution results of multiple tools, the agent prioritizes the execution results of tools with higher evidence levels as the basis for inference, discarding or reducing the weight of execution results of tools with lower evidence levels.

[0109] Based on this implementation method, on the one hand, by pre-setting evidence grading rules to differentiate the execution results of multi-source tools, the generation process of the current reasoning result can be logically weighted according to the evidence level. This helps to prioritize the use of the more credible tool execution result when there are content conflicts or quality differences among the execution results of multiple tools, thereby improving the accuracy and logical consistency of the current reasoning result. On the other hand, the evidence level of each tool execution result is dynamically determined and stored during the iterative processing, realizing the automated labeling and structured retention of evidence levels. This provides a retrieval basis including the quality dimension for fine-grained evidence tracing of subsequent reasoning information to be traced.

[0110] In one application example, the initial evidence in the evidence index dataset can employ vector indexing, text indexing, and multimodal indexing. In step S505, target reasoning evidence that is related to the information to be traced back to the source is determined from the evidence index dataset, which can be done using a hybrid retrieval method. For example, firstly, vector indexing technology is used to transform the information to be traced back to the source into a high-dimensional semantic vector. Cosine similarity is calculated to retrieve semantically highly matching initial evidence in the vector space to capture deep semantic connections. A text-based matching algorithm is used to extract and match keywords from the information to be traced back to the source, ensuring the recall accuracy of technical terms, specific entities, or precise numerical values. Furthermore, the retrieval process can introduce a time decay factor to weight the generation or publication time of the initial evidence in the evidence index dataset, prioritizing recently generated tool execution results or newly published document records, thereby further improving the timeliness of the evidence while ensuring semantic relevance and keyword matching. For example, semantic similarity score, keyword matching score, and time decay weight can be comprehensively considered to select the initial evidence with the highest comprehensive score from the evidence index dataset as the target reasoning evidence.

[0111] In another application example, in step S505, determining the target reasoning evidence that is related to the reasoning information to be traced from the evidence index dataset may include: determining at least one candidate reasoning evidence that is related to the reasoning information to be traced from multiple initial pieces of evidence; and selecting the target reasoning evidence from multiple candidate reasoning evidence based on the evidence details of the candidate reasoning evidence.

[0112] Evidence details include information such as the evidence summary, evidence level, and evidence type mentioned above. After conducting an initial search of the initial evidence and identifying multiple candidate inference evidences with related relationships, the evidence details corresponding to each candidate inference evidence can be further analyzed. For example, candidate inference evidence with higher evidence levels can be prioritized as target inference evidence; or, in the case of the same evidence level, a secondary screening can be conducted based on the evidence type (such as prioritizing "clinical guidelines" over "case reports") or the logical fit between the evidence summary and the inference information to be traced.

[0113] If the evidence details of multiple candidate inference evidences all meet the preset selection criteria, these candidate inference evidences can be jointly determined as the target inference evidence to provide multi-dimensional supporting evidence; or, the candidate inference evidences can be sorted according to the semantic similarity between the candidate inference evidences and the inference information to be traced, and the candidate inference evidence with the higher ranking can be selected as the target inference evidence.

[0114] Based on this implementation method, by introducing evidence details as a screening dimension for evidence tracing, the evidence selection tracing logic can be realized from semantic relevance to quality relevance, thereby enhancing the rigor and user trust of the entire task processing method when applied in vertical professional fields.

[0115] In one implementation, during the iterative processing, a current subtask to be processed is generated based on the previous inference result and the unfinished subtasks. This may include generating the current subtask to be processed based on the previous inference result, the unfinished subtasks, and the session context data corresponding to the task processing request. The session context data includes user-uploaded data and / or historical session text.

[0116] User-uploaded data can be multimodal data, including at least one of text data, image data, and structured data. Historical conversation text refers to conversation interaction records generated before the current conversation round.

[0117] For example, when generating the current subtask to be processed, the agent first constructs the current prompt context by combining the facts confirmed in the previous inference result, the list of incomplete subtasks, and the session context data corresponding to the task processing request. Then, it analyzes the prompt context. If it finds that the previous inference result mentions key features in the user-uploaded image data but has not yet been deeply interpreted, or if the user has added specific constraints in the historical session text (such as "focus only on children's medication"), the agent will combine this dynamic information to generate a targeted current subtask to be processed. For example, it may generate a subtask of "calling the drug database to query the list of antibiotics that meet the safety standards for children's medication". This ensures that the generated current subtask to be processed not only inherits the inference logic of the previous round, but also fully integrates the latest multimodal data and historical interaction intentions provided by the user during the session.

[0118] Furthermore, method 500 may also include: storing the session context data from each round of iterative processing as initial evidence in the evidence index dataset.

[0119] In other words, the initial evidence in the evidence index dataset not only includes the tool execution results, but also covers the session context data corresponding to the task processing request. Therefore, the evidence index dataset in this application embodiment is a dynamic session-level index, that is, the evidence index dataset is an index dataset dynamically constructed for the current task processing request, which can capture user-uploaded data (such as data uploaded by the user in real time) and historical session text in the reasoning process in real time during the multi-round interaction of the task processing request.

[0120] Based on this implementation method, by incorporating the session context data corresponding to the task processing request into the evidence index dataset, the problem of not being able to effectively utilize the dynamic new information generated during the session caused by relying on the pre-built static knowledge base is effectively overcome. This dynamic session-level indexing mechanism ensures that multimodal data dynamically input by the user can be covered when tracing evidence, thereby improving the response accuracy of the agent.

[0121] Figure 6 A flowchart of an automatic question-answering method 600 provided in an embodiment of this application is shown. This method 600 can be applied to a server, for example, by... Figure 1 The server 102 shown is executed. Figure 6As shown, the method 600 may include: step S601, obtaining an initial solution result generated for the question to be answered, the initial solution result being generated based on the tool execution result of at least one tool; step S602, identifying multiple candidate inference information from the initial solution result, the multiple candidate inference information including at least one inference information to be traced back to its source; step S603, determining target inference evidence that is related to the inference information to be traced back to its source from the evidence index dataset, the evidence index dataset including multiple initial pieces of evidence, the initial pieces of evidence including the tool execution result; step S604, generating a target solution result for responding to the question to be answered, the target solution result including multiple candidate inference information and a traceability identifier set for the inference information to be traced back to its source, the traceability identifier pointing to the target inference evidence.

[0122] Method 600 can serve as an example of method 400 or method 500 in an automatic question-answering scenario. That is, the question to be answered belongs to the task processing request, the initial answer result belongs to the initial task processing result, and the target answer result belongs to the target task processing result. Therefore, the implementation method and technical effect of method 600 can be found in the relevant descriptions of method 400 or method 500.

[0123] Figure 7 A flowchart illustrating a computer-aided diagnostic method 700 provided in an embodiment of this application is shown. This method 700 can be applied to a server, for example, by... Figure 3A or Figure 3B The server-side 302 execution is shown. (As shown...) Figure 7 As shown, the method 700 may include: step S701, obtaining an initial medical diagnosis result generated in response to a medical diagnosis request, the initial medical diagnosis result being generated based on the tool execution result of at least one tool; step S702, identifying multiple candidate medical viewpoints from the initial medical diagnosis result, the multiple candidate medical viewpoints including at least one medical viewpoint to be traced; step S703, determining target reasoning evidence that is related to the medical viewpoint to be traced from an evidence index dataset, the evidence index dataset including multiple initial pieces of evidence, the initial pieces of evidence including the tool execution result; step S704, generating a target medical diagnosis result for responding to a medical diagnosis request, the target medical diagnosis result including multiple candidate medical viewpoints and a traceability identifier set for the medical viewpoint to be traced, the traceability identifier pointing to the target reasoning evidence.

[0124] Method 700 can serve as an example of method 400 or method 500 in an automatic question-answering scenario. Specifically, the medical diagnosis request belongs to the task processing request, the initial medical diagnosis result belongs to the initial task processing result, the target medical diagnosis result belongs to the target task processing result, the candidate medical viewpoint belongs to the candidate reasoning information, and the medical viewpoint to be traced belongs to the reasoning information to be traced. Therefore, the implementation method and technical effect of method 700 can be found in the relevant descriptions of method 400 or method 500.

[0125] In this application embodiment, medical opinions (such as candidate medical opinions or medical opinions to be traced) include, but are not limited to: diagnostic conclusions, such as specific disease names, disease stages or differential diagnosis lists inferred based on patient symptoms and examination results; treatment opinions, such as drug prescription plans (including drug names, dosages, routes of administration and courses of treatment), surgical intervention suggestions, physical therapy plans or lifestyle intervention guidance, etc.; prognostic analysis results, such as the probability of recurrence of old diseases predicted based on clinical data, assessment of expected survival, probability of complications or prediction of treatment response.

[0126] Users can trigger a medical diagnosis request by inputting multimodal data, such as structured medical record text (e.g., chief complaint, present illness, past medical history), medical imaging data (e.g., X-rays, CT images, MRI scans), pathological slide images, laboratory test reports (e.g., blood routine tests, biochemical indicator values), and patient-reported symptom audio or video recordings. Method 700 generates a target medical diagnosis result with a traceability identifier, allowing users to intuitively view the target reasoning evidence supporting each traceable medical viewpoint. For example, by manipulating the traceability identifier, users can further obtain details of the target reasoning evidence to independently verify and evaluate the reliability of each traceable medical viewpoint. This transparent presentation helps medical professionals quickly verify the authority and applicability of information, reducing the risk of misdiagnosis due to unclear or low-quality evidence, while enhancing patients' trust in the computer-aided diagnostic system and promoting the effective implementation of computer-aided diagnostic models in clinical settings.

[0127] Figure 8 and Figure 9 The diagram illustrates an application example of the computer-aided diagnostic method 700. Among other things, Figure 8 This illustrates the process of generating preliminary medical diagnostic results, such as... Figure 8 As shown, the process includes steps S801 to S807.

[0128] Step S801: The user inputs: symptom description / medication consultation / disease inquiry / file upload, thereby triggering a medical diagnosis request. For example, the patient describes the symptoms of "sudden chest pain accompanied by shortness of breath" in detail through the text input box, or uploads a recent electrocardiogram, or uploads a test report.

[0129] Step S802: Safety Pre-check: Critical Symptom Recognition. For example, the system uses natural language processing technology to quickly scan user input and compare it with preset critical symptom recognition guidelines to determine whether there are critical symptoms that require immediate intervention, such as chest pain, difficulty breathing, stroke, etc.

[0130] Step S803: Is a critical symptom detected? If a critical symptom is detected (yes), proceed to step S804; if no critical symptom is detected (no), proceed to step S805.

[0131] Step S804: Generate an immediate response: Please seek medical attention immediately or call emergency services. For example, when the system determines that the user has critical symptoms that require immediate intervention, it will no longer perform complex reasoning, but will directly pop up a red warning window, clearly informing the user to "go to the nearest hospital emergency department or call 120 for emergency services" in order to buy time for rescue.

[0132] Step S805: Iterative processing. This process first generates multiple subtasks based on the medical diagnosis request and initializes the inference results. Iterative processing is performed on the multiple subtasks at least once until all subtasks are completed.

[0133] Each iteration of the process includes a thinking phase, specifically encompassing intent analysis, emergency assessment, medical entity identification, and tool selection strategy. In the thinking phase, based on the previous reasoning result and unfinished sub-tasks, a current sub-task to be processed is generated. For example, if the user intent is analyzed as "diagnosing the cause of illness," medical entities such as "hypertension" and "dizziness" are identified, and a tool selection strategy is formulated based on the currently missing information. This strategy includes whether a tool needs to be invoked, the tools to be invoked, and the invocation parameters. Invoked tools include, but are not limited to, medical literature retrieval tools, medication dosage calculation tools, image feature extraction tools, clinical guideline lookup tools, symptom assessment tools, and drug databases. Further, it is determined whether a tool needs to be invoked. If so (yes), the process proceeds to the action phase, where at least one tool is invoked to execute the current sub-task. If no tool is needed (no), the process proceeds to direct reasoning (knowledge base answer). For example, for simple common-sense questions, answers are generated directly based on the built-in medical knowledge base without invoking external tools.

[0134] After the action phase ends, the observation phase begins. On one hand, it generates the current inference result based on the tool execution results returned by at least one tool (or directly generates the result based on the knowledge base). On the other hand, it determines whether the obtained current inference result is sufficiently informative and has clear safety boundaries, i.e., whether each subtask has been completed. If the information is insufficient or the safety boundaries are unclear (no), it indicates that there are incomplete subtasks, and the next iteration begins, repeating the "thinking phase - action phase - observation phase" iterative process. If the information is sufficient and the safety boundaries are clear (yes), it indicates that multiple subtasks have been completed, and the iterative process ends, proceeding to step S806.

[0135] Step S806: Generate preliminary medical diagnosis results, that is, generate preliminary medical diagnosis results based on the current reasoning results, including: mandatory safety statement (such as "AI suggestions are for reference only and cannot replace doctor consultation"), reasoning process display (showing the logical chain from user input to the current reasoning result), candidate medical viewpoints (such as diagnosis conclusions, treatment opinions, prognostic analysis results), and initial evidence chain (listing the initial evidence supporting the candidate medical viewpoints).

[0136] Step S807: Safety review, which is a safety review conducted before the output of the preliminary medical diagnosis results. Examples include: efficacy commitment check (e.g., ensuring no illegal commitments are made), folk remedy detection (e.g., filtering out unverified folk remedies), and diagnostic inference compliance (e.g., ensuring that the diagnostic logic complies with medical ethics and laws and regulations).

[0137] Figure 9 This illustrates the process of generating the target medical diagnosis result, such as... Figure 9 As shown, the process includes steps S901 to S904.

[0138] Step S901: Obtain the preliminary medical diagnosis results after security review, which include candidate medical viewpoints, a summary of the reasoning process and the initial chain of evidence, but have not yet added structured traceability identifiers.

[0139] Step S902: Obtain the evidence index dataset. The evidence index dataset is dynamically constructed and includes... Figure 8 The tool execution results and session context data (such as new user-uploaded data and historical session text) are shown in the iterative process.

[0140] Step S903: Evidence tracing, for example, each candidate medical viewpoint in step S901 is taken as an independent medical viewpoint to be traced. Based on step S703, target reasoning evidence that is related to the medical viewpoint to be traced is determined from the evidence index dataset, thereby establishing the mapping relationship from viewpoint to evidence. This step can be standardized by using citation format examples, such as: "Aspirin can inhibit platelets[1]" indicates that the pharmacological effect originates from target reasoning evidence 1, and "reduce the risk of myocardial infarction[2][3]" indicates that the clinical benefit is supported by target reasoning evidence 2 and target reasoning evidence 3.

[0141] Step S904: Generate the target medical diagnosis result, which includes not only the mandatory safety statement, reasoning process demonstration, and candidate medical viewpoints from the initial medical conclusion, but also source identification markers pointing to the target reasoning evidence. For example, the final output target medical diagnosis result not only includes the original diagnostic suggestion, but also forcibly embeds three enhanced contents: first, a prominent mandatory safety statement (such as "This suggestion is generated by AI and must be reviewed by a licensed physician"); second, a clearly presented reasoning process demonstration (such as "Based on the patient's blood pressure value >140 / 90mmHg + fundus artery stenosis → inferring the existence of target organ damage → initiating secondary prevention"); and third, binding each medical viewpoint to be traced to its corresponding source identification marker in the output, forming a structured and interactive final text. For example, the output format is: "It is recommended to take 100mg of aspirin[1] daily[2], which may reduce the risk of cardiovascular events[3]." Here, [1] refers to Chapter 5 of the drug instructions, [2] refers to the recommended dosage table of clinical guidelines, and [3] refers to the analysis research literature. Users can view the complete evidence details by clicking the traceability mark, thereby realizing the transparent, verifiable and credible delivery of diagnostic results.

[0142] Corresponding to the application scenarios and methods 400 and 500 provided in the embodiments of this application, the embodiments of this application also provide a task processing apparatus, including: an initial task processing result acquisition module, used to acquire an initial task processing result generated in response to a task processing request, wherein the initial task processing result is generated based on the tool execution result of at least one tool; a candidate inference information identification module, used to identify multiple candidate inference information from the initial task processing result, wherein the multiple candidate inference information includes at least one inference information to be traced; a target inference evidence determination module, used to determine target inference evidence that is associated with the inference information to be traced from an evidence index dataset, wherein the evidence index dataset includes multiple initial pieces of evidence, wherein the initial pieces of evidence include the tool execution result; and a target task processing result generation module, used to generate a target task processing result for responding to the task processing request, wherein the target task processing result includes the multiple candidate inference information and a traceability identifier set for the inference information to be traced, wherein the traceability identifier points to the target inference evidence.

[0143] In one implementation, the target task processing result includes a target response text, and the target task processing result generation module is specifically used to: construct an initial response text using the plurality of candidate inference information; determine the insertion position of the tracing identifier in the initial response text; insert the tracing identifier into the insertion position to obtain the target response text.

[0144] In one embodiment, the candidate reasoning information identification module is specifically used to: semantically decompose the initial task processing result, extract multiple independent atomic semantic units, and obtain the multiple candidate reasoning information, each of the candidate reasoning information corresponding to an atomic semantic unit with complete semantics.

[0145] In one implementation, the target reasoning evidence determination module is specifically used to: determine at least one candidate reasoning evidence that is related to the reasoning information to be traced from a plurality of initial evidences; in response to the candidate reasoning evidence including positive evidence and negative evidence, obtain the positive evidence and the negative evidence to obtain the target reasoning evidence, and generate a dispute status marker, wherein the dispute status marker is used to identify the target reasoning evidence as evidence with conflict and dispute.

[0146] In one implementation, the initial task processing result acquisition module is specifically used to: generate multiple sub-tasks based on the task processing request and initialize the inference result; perform at least one round of iterative processing on the multiple sub-tasks until all the multiple sub-tasks have been completed, wherein each round of iterative processing includes: generating a currently pending sub-task based on the previous inference result and the unfinished sub-tasks; calling at least one tool to execute the currently pending sub-task; generating a current inference result based on the tool execution result returned by at least one tool; and generating the initial task processing result according to the current inference result in response to the completion of all the multiple sub-tasks.

[0147] In one embodiment, the task processing device further includes: an evidence index dataset storage module, used to store the tool execution results generated in each round of iterative processing as the initial evidence in the evidence index dataset, and to store the evidence details corresponding to the tool execution results in the evidence index dataset after associating them with the corresponding initial evidence, wherein the evidence details include at least one of the following: evidence summary, evidence level, and evidence type.

[0148] In one implementation, the tool execution results are multiple, and the initial task processing result acquisition module is further configured to: determine the evidence level of each tool execution result based on a preset evidence grading rule; and generate a current reasoning result based on each tool execution result and its corresponding evidence level.

[0149] In one implementation, the target reasoning evidence determination module is specifically used to: determine at least one candidate reasoning evidence that is related to the reasoning information to be traced from a plurality of initial evidences; and select the target reasoning evidence from a plurality of candidate reasoning evidences based on the evidence details of the candidate reasoning evidence.

[0150] In one implementation, the initial task processing result acquisition module is further configured to: generate the currently pending subtask based on the previous inference result, the unfinished subtask, and the session context data corresponding to the task processing request, wherein the session context data includes user-uploaded data and / or historical session text, and the user-uploaded data includes at least one of text data, image data, and structured data; and the evidence index dataset storage module is further configured to: store the session context data in each round of iteration processing as the initial evidence in the evidence index dataset.

[0151] Corresponding to the application scenario and method 600 provided in the embodiments of this application, the embodiments of this application also provide an automatic question answering device, including: an initial answer result acquisition module, used to acquire an initial answer result generated for a question to be answered, the initial answer result being generated based on the tool execution result of at least one tool; a candidate reasoning information identification module, used to identify multiple candidate reasoning information from the initial answer result, the multiple candidate reasoning information including at least one reasoning information to be traced back to its source; a target reasoning evidence determination module, used to determine target reasoning evidence that is associated with the reasoning information to be traced back to its source from an evidence index dataset, the evidence index dataset including multiple initial pieces of evidence, the initial pieces of evidence including the tool execution result; and a target answer result generation module, used to generate a target answer result for responding to the question to be answered, the target answer result including the multiple candidate reasoning information and a traceability identifier set for the reasoning information to be traced back to its source, the traceability identifier pointing to the target reasoning evidence.

[0152] Corresponding to the application scenario and method 700 provided in the embodiments of this application, the embodiments of this application also provide a computer-aided diagnostic device, including: an initial medical diagnosis result acquisition module, used to acquire an initial medical diagnosis result generated in response to a medical diagnosis request, the initial medical diagnosis result being generated based on the tool execution result of at least one tool; a candidate medical opinion identification module, used to identify multiple candidate medical opinions from the initial medical diagnosis result, the multiple candidate medical opinions including at least one medical opinion to be traced; a target inference evidence determination module, used to determine target inference evidence that is associated with the medical opinion to be traced from an evidence index dataset, the evidence index dataset including multiple initial pieces of evidence, the initial pieces of evidence including the tool execution result; and a target medical diagnosis result generation module, used to generate a target medical diagnosis result in response to the medical diagnosis request, the target medical diagnosis result including the multiple candidate medical opinions and a traceability identifier set for the medical opinion to be traced, the traceability identifier pointing to the target inference evidence.

[0153] The functions of each module in each device in the embodiments of this application can be found in the corresponding description in the above method, and they have corresponding beneficial effects, which will not be repeated here.

[0154] Figure 10 This is a block diagram of an electronic device used to implement embodiments of this application. For example... Figure 10As shown, the electronic device includes a memory 1001 and a processor 1002. The memory 1001 stores a computer program that can run on the processor 1002. When the processor 1002 executes the computer program, it implements the method described in the above embodiments. The number of memories 1001 and processors 1002 can be one or more. In a specific implementation, the electronic device may also include a communication interface 1003 for communicating with external devices and performing data exchange and transmission.

[0155] In practical implementation, if the memory 1001, processor 1002, and communication interface 1003 are implemented independently, they can be interconnected via a bus to communicate with each other. This bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 10 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0156] Optionally, in a specific implementation, if the memory 1001, processor 1002 and communication interface 1003 are integrated on a single chip, the memory 1001, processor 1002 and communication interface 1003 can communicate with each other through an internal interface.

[0157] This application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method provided in this application.

[0158] This application provides a computer program product, including a computer program that, when executed by a processor, implements the method provided in this application.

[0159] This application also provides a chip, which includes a processor for calling and executing instructions stored in a memory, causing a communication device with the chip installed to perform the method provided in this application.

[0160] This application also provides a chip, including: an input interface, an output interface, a processor, and a memory. The input interface, output interface, processor, and memory are connected through an internal connection path. The processor is used to execute code in the memory. When the code is executed, the processor is used to execute the method provided in the application embodiment.

[0161] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. General-purpose processors can be microprocessors or any conventional processor. It is worth noting that the processor can be a processor supporting Advanced Reduced Instruction Set Machines (ARM) architecture.

[0162] Further, optionally, the aforementioned memory may include read-only memory and random access memory. The memory may be volatile memory or non-volatile memory, or may include both. Non-volatile memory may include read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory may include random access memory (RAM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available. Examples include Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced Synchronous DRAM (ESDRAM), Sync Link DRAM (SLDRAM), and Direct Rambus RAM (DR RAM).

[0163] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another.

[0164] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0165] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified.

[0166] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process. Furthermore, the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functionality involved.

[0167] The logic and / or steps described in the flowchart or otherwise herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-included system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0168] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. All or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware, the program being stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiments.

[0169] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. This storage medium can be a read-only memory, a disk, or an optical disk, etc.

[0170] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope described in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A task processing method, comprising: Obtain the initial task processing result generated in response to the task processing request, the initial task processing result being generated based on the tool execution result of at least one tool; Multiple candidate inference information are identified from the initial task processing results, and the multiple candidate inference information includes at least one inference information to be traced back to its source; The evidence index dataset is used to identify target reasoning evidence that is related to the reasoning information to be traced, wherein the evidence index dataset includes multiple initial pieces of evidence, and the initial pieces of evidence include the execution results of the tool. Generate a target task processing result in response to the task processing request. The target task processing result includes the plurality of candidate inference information and a tracing identifier set for the inference information to be traced, the tracing identifier pointing to the target inference evidence.

2. The method according to claim 1, wherein, The target task processing result includes the target response text, and the generation of the target task processing result for responding to the task processing request includes: The initial response text is constructed using the multiple candidate inference information; Determine the insertion position of the tracing identifier in the initial response text; The source tracing identifier is inserted into the insertion position to obtain the target response text.

3. The method according to claim 1, wherein, The identification of multiple candidate inference information from the initial task processing results includes: The initial task processing result is semantically decomposed to extract multiple independent atomic semantic units, resulting in multiple candidate inference information, each of which corresponds to an atomic semantic unit.

4. The method according to claim 1, wherein, The step of determining the target reasoning evidence that is related to the reasoning information to be traced from the evidence index dataset includes: From the multiple initial pieces of evidence, at least one candidate piece of inference evidence is identified that is related to the inference information to be traced. In response to the candidate reasoning evidence including positive evidence and negative evidence, the positive evidence and the negative evidence are obtained to obtain the target reasoning evidence, and a dispute status marker is generated. The dispute status marker is used to identify the target reasoning evidence as evidence with conflict and dispute.

5. The method according to any one of claims 1 to 4, wherein, The step of obtaining the initial task processing result generated in response to the task processing request includes: Based on the task processing request, multiple sub-tasks are generated, and the inference results are initialized; The plurality of subtasks are processed in at least one round of iteration until all the subtasks have been completed. Each round of iteration includes: generating a current subtask to be processed based on the previous inference result and the unfinished subtasks; calling at least one tool to execute the current subtask to be processed; and generating a current inference result based on the tool execution result returned by at least one tool. In response to the completion of all the subtasks, the initial task processing result is generated based on the current inference result.

6. The method according to claim 5, wherein, The task processing method further includes: The tool execution results generated in each round of iteration are stored as the initial evidence in the evidence index dataset. The evidence details corresponding to the tool execution results are associated with the corresponding initial evidence and then stored in the evidence index dataset. The evidence details include at least one of the following: evidence summary, evidence level, and evidence type.

7. The method according to claim 6, wherein, The tool execution results are multiple, and the generation of the current inference result based on the tool execution result returned by at least one tool includes: Based on preset evidence grading rules, the evidence level of the execution results of each tool is determined; The current reasoning result is generated based on the execution results of each tool and its corresponding level of evidence.

8. The method according to claim 6, wherein, The step of determining the target reasoning evidence that is related to the reasoning information to be traced from the evidence index dataset includes: From the multiple initial pieces of evidence, at least one candidate piece of inference evidence is identified that is related to the inference information to be traced. Based on the evidence details of the candidate inference evidence, the target inference evidence is selected from multiple candidate inference evidence.

9. The method according to claim 5, wherein, The process of generating a currently pending subtask based on the previous reasoning result and the unfinished subtasks includes: Based on the previous reasoning result, the unfinished subtasks, and the session context data corresponding to the task processing request, the current subtask to be processed is generated. The session context data includes user-uploaded data and / or historical session text. The user-uploaded data includes at least one of text data, image data, and structured data. The task processing method also includes: The session context data from each round of iteration is stored as the initial evidence in the evidence index dataset.

10. An automatic question-answering method, comprising: Obtain an initial solution result generated for the question to be answered, the initial solution result being generated based on the tool execution result of at least one tool; Multiple candidate inference information are identified from the initial solution results, and the multiple candidate inference information includes at least one inference information to be traced back to the source. The evidence index dataset is used to identify target reasoning evidence that is related to the reasoning information to be traced, wherein the evidence index dataset includes multiple initial pieces of evidence, and the initial pieces of evidence include the execution results of the tool. A target solution result is generated to respond to the question to be answered. The target solution result includes the plurality of candidate reasoning information and a source identification set for the reasoning information to be traced. The source identification points to the target reasoning evidence.

11. A computer-aided diagnostic method, comprising: Obtain an initial medical diagnosis result generated in response to a medical diagnosis request, the initial medical diagnosis result being generated based on the tool execution result of at least one tool; Multiple candidate medical viewpoints are identified from the initial medical diagnosis results, including at least one medical viewpoint to be traced back to its source. The evidence index dataset is used to identify target inference evidence that is related to the medical viewpoint to be traced, the evidence index dataset includes multiple initial pieces of evidence, the initial pieces of evidence include the execution results of the tool; Generate a target medical diagnosis result in response to the medical diagnosis request. The target medical diagnosis result includes the plurality of candidate medical viewpoints and a source identification identifier set for the medical viewpoint to be traced, and the source identification identifier points to the target reasoning evidence.

12. A task processing system, comprising: The client is used to send task processing requests; The server, connected to the client, is used to generate a target task processing result in response to the task processing request based on the task processing method according to any one of claims 1 to 9. The client is also used to output the processing results of the target task.

13. A computer-aided diagnostic system, comprising: The client is used to send medical diagnosis requests; The server, connected to the client, is used to generate a target medical diagnostic result in response to the medical diagnostic request based on the diagnostic method of claim 11. The client is also used to output the target diagnostic results.

14. An electronic device comprising a memory, a processor, and a computer program stored in the memory, wherein the processor, when executing the computer program, implements the method of any one of claims 1 to 11.

15. A computer-readable storage medium storing a computer program that, when executed by a processor, implements the method of any one of claims 1 to 11.

16. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 11.

Images