Data processing method and electronic device

Abstract

This application discloses a data processing method and an electronic device. The data processing method includes: receiving a task request input by a user; processing the task request to generate a target task request, wherein the target task request represents a request that meets processing requirements after intent analysis and complexity classification; generating a task execution plan based on the complexity classification information of the target task request, wherein the task execution plan represents the execution plan of at least one subtask contained in the target task request and the dynamic dependencies between the subtasks; and executing the task execution plan to obtain a target processing result.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and more specifically to a data processing method and electronic device. Background Technology

[0002] With the development of artificial intelligence technology, deep research systems, as an important application of large models, can assist users in conducting complex research tasks such as industry analysis and market research. Currently, the processing method using a single general-purpose large language model employs prompts to guide the model to complete tasks autonomously. However, this approach suffers from high inference costs, uncontrollable research processes, difficulty in guaranteeing the stability and accuracy of research results, and high overall cost. Summary of the Invention

[0003] In view of the above, this application provides the following technical solution:

[0004] A data processing method, comprising:

[0005] Receive task requests from user input;

[0006] The task request is processed to generate a target task request, which represents a request that meets the processing requirements after intent analysis and complexity classification.

[0007] Based on the complexity classification information of the target task request, a task execution plan is generated. The task execution plan represents the execution plan of at least one subtask included in the target task request and the dynamic dependencies between the subtasks.

[0008] Execute the task execution plan to obtain the target processing result.

[0009] Optionally, processing the task request to generate a target task request includes:

[0010] The task request is subjected to intent recognition and complexity classification to obtain the classification result;

[0011] In response to the classification result being a complex classification, the task request is processed to complete the information, thereby obtaining the target task request.

[0012] Optionally, it also includes:

[0013] In response to the classification result being invalid, a rejection message corresponding to the task request is output.

[0014] In response to the classification result being a simple classification, the processing result corresponding to the task request is output.

[0015] Optionally, the intent recognition and complexity classification of the task request are implemented through a first model; wherein, the first model is a language model based on training data labeled with intent categories and complexity levels, and optimized through supervised fine-tuning; the reward function used by the first model is constructed according to the classification accuracy and output normalization of the first model.

[0016] Optionally, generating a task execution plan based on the complexity classification information of the target task request includes:

[0017] Based on the complexity classification information of the target task request, the target task request is split into at least one sub-task;

[0018] Analyze the semantic relationships and execution logic between each subtask to generate a dependency graph representing the execution order of the subtasks, wherein the dependency graph contains at least one non-fixed preset custom dependency relationship;

[0019] Based on the subtasks and the dependency graph, the task execution plan is constructed.

[0020] Optionally, the generation of the task execution plan is implemented through a second model, wherein the second model receives the target task request as input and outputs the task execution plan containing the dependency graph; the second model represents a planning model trained by reinforcement learning, and the reward function of the reinforcement learning is constructed based on at least the completeness of the task execution plan's coverage of the original request, the execution efficiency reflected by the dependency graph, and the quality of the results obtained by executing the plan.

[0021] Optionally, executing the task execution plan to obtain the target processing result includes:

[0022] For each subtask in the task execution plan, an information search and extraction operation is performed, wherein the information search and extraction operation includes: generating a search query based on the current subtask; calling a retrieval service to obtain retrieval results matching the search query; in response to the sufficiency condition not being met by the retrieval results, generating an optimized search query based on the retrieval results, and performing information search and extraction operations based on the optimized search query until the target retrieval result meets the sufficiency condition;

[0023] Based on the target retrieval results corresponding to each subtask, the target processing results are generated.

[0024] Optionally, the execution information search and extraction operation is implemented through a third model; the third model is a search model trained by reinforcement learning, which evaluates the sufficiency of information and generates or optimizes search queries based on the current subtask and the retrieved search results; wherein, the reward function of the reinforcement learning is constructed based on the deviation of the retrieved search results obtained under the guidance of the third model from the benchmark search results in terms of the quality of the final generated target processing result.

[0025] Optionally, the step of calling the retrieval service to obtain retrieval results matching the search query includes calling a multi-source retrieval system; the multi-source retrieval system integrates at least two of the following retrieval services: web page content crawling service based on condition parsing, network search engine application programming interface service, and local document retrieval service based on semantic understanding of a large language model.

[0026] An electronic device, comprising:

[0027] A memory for storing computer programs and the data generated by the execution of said computer programs;

[0028] A processor for executing the computer program to achieve:

[0029] Receive task requests from user input;

[0030] The task request is processed to generate a target task request, which represents a request that meets the processing requirements after intent analysis and complexity classification.

[0031] Based on the complexity classification information of the target task request, a task execution plan is generated. The task execution plan represents the execution plan of at least one subtask included in the target task request and the dynamic dependencies between the subtasks.

[0032] Execute the task execution plan to obtain the target processing result. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0034] Figure 1 A flowchart illustrating a data processing method provided in an embodiment of this application;

[0035] Figure 2 A schematic diagram of the core architecture of a multi-source retrieval system provided in this application embodiment;

[0036] Figure 3 A schematic diagram illustrating a document index creation process provided in an embodiment of this application;

[0037] Figure 4 A schematic diagram of a system architecture for in-depth research applications is provided as an embodiment of this application;

[0038] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0039] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0040] The terms "first" and "second," etc., used in this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units may include steps or units not listed, but may include steps or units not listed.

[0041] This application provides a data processing method applicable to electronic devices (such as servers, terminal computers, and smart terminals). For any task request submitted by a user, the method first processes the request to obtain a target task request that meets the execution requirements. Then, based on the request complexity, it generates an execution plan containing dynamic dependencies between subtasks. Finally, it executes the plan and outputs the target processing result, achieving optimized resource allocation and precise task processing. See also... Figure 1 The diagram illustrates a data processing method provided in an embodiment of this application, which may include the following steps:

[0042] S101, Receive task requests input by the user.

[0043] User-input task requests can be various information processing needs submitted by users to electronic devices through input devices (keyboards, touchscreens, voice input devices). These requests can take the form of text, voice, or a combination of text and images, and can include various types such as life consultations, business analysis, and academic research, serving as the initial input for data processing. For example, the electronic device receives user-input task requests through preset interactive interfaces (such as APP clients, web pages, and voice assistant interfaces). If the request is in voice format, it is first converted into text using a speech recognition model; if it is in image-text format, the text information is extracted using image recognition and OCR (Optical Character Recognition) technologies. Finally, all types of task requests are uniformly converted into a standardized text format and stored in the electronic device's temporary storage for subsequent processing.

[0044] S102. Process the task request and generate the target task request.

[0045] The target task request is represented as a request that meets processing requirements after intent analysis and complexity classification. During the processing of user-input task requests, task intent analysis and complexity classification are performed. Intent analysis identifies the core intent of the task request, such as determining the information query intent, content creation intent, planning intent, or invalid information processing intent based on the application scenario. Complexity classification represents the task request's intent type, the number of processing steps required, and the difficulty of information acquisition, categorizing it into simple, complex, and invalid categories. Furthermore, for complex task requests, information completion processing can be performed, such as requesting the user to supplement missing key information, ultimately forming a target task request with a clear intent and complete information. Requests in the simple and invalid categories can be processed directly, as will be explained in subsequent embodiments of this application.

[0046] For example, if a user inputs a task request of "Write a market analysis report on a certain type of home appliance in 2024," the electronic device identifies the user's intent as content creation intent through intent analysis and classifies it as complex (requiring multi-step processing and multi-source information retrieval). It then sends a completion query to the user, such as "Please confirm whether the report needs to include price range analysis, core brand comparison, and the time range for data statistics?" The user replies, "It needs to include price ranges and brand comparisons, with a time range of 2023-2024." The electronic device integrates the original request and the completion information to generate the target task request: "Write a market analysis report on a certain type of home appliance in 2024, including market price range analysis for 2023-2024, core brand comparison, and comprehensive coverage of other content."

[0047] In this step, intent analysis and complexity classification enable refined identification of task requests, distinguishing requests with different processing needs from the source. Information completion is used to generate target task requests, ensuring the accuracy of subsequent task planning and execution. This solves the problem of poor task request identification accuracy or resource waste caused by using a uniform processing model.

[0048] S103. Generate a task execution plan based on the complexity classification information of the target task request.

[0049] A task execution plan represents the execution plan of at least one subtask contained in a target task request and the dynamic dependencies between these subtasks. During the generation of the task execution plan based on the complexity classification information of the target task request, subtask decomposition and dynamic dependency construction steps can be performed. Subtask decomposition refers to breaking down the target task request into at least one independently executable subtask. The granularity of the decomposition can be determined based on the complexity; for example, the higher the complexity, the more detailed the subtask decomposition. Dynamic dependency construction involves analyzing the semantic relationships and execution logic between subtasks, determining the execution order of the subtasks, and constructing dynamic dependencies, such as serial, parallel, or custom dependencies. This allows the electronic device to integrate subtasks and dynamic dependencies to generate a task execution plan, facilitating subsequent processing of the target task request. This subtask decomposition based on complexity classification information achieves refined task processing. By constructing dynamic dependencies, it replaces the traditional fixed pipeline, solves the problem of weak generalization ability of the planning module, and enables the task execution plan to adapt to target task requests of different complexities, improving the efficiency and rationality of task execution.

[0050] For example, consider the task request "Write a market analysis report on a certain type of home appliance in 2024, including market price range analysis for 2023-2024, comparison of core brands, and comprehensive coverage of other content." This task is classified as high complexity. The electronic device breaks it down into six sub-tasks: Sub-task 1 (Industry environment analysis), Sub-task 2 (Market size statistics for 2023-2024), Sub-task 3 (Comparison of product parameters of core brands), Sub-task 4 (Market price range analysis for 2023-2024), Sub-task 5 (Technology development trend assessment), and Sub-task 6 (Report integration and writing). Dynamic dependencies are then constructed: Sub-task 1 and Sub-task 2 are executed in parallel; Sub-task 3 is executed independently; Sub-task 4 depends on the execution results of Sub-task 2 and Sub-task 3; Sub-task 5 depends on the execution results of Sub-task 1 and Sub-task 3; Sub-task 6 depends on all execution results of Sub-task 1, Sub-task 2, Sub-task 3, Sub-task 4, and Sub-task 5. The electronic device then draws a dependency graph of these sub-tasks and dependencies, generating a complete task execution plan.

[0051] S104. Execute the task execution plan and obtain the target processing results.

[0052] The electronic device executes each subtask according to the dependency graph in the task execution plan. For each subtask, it calls a retrieval service to obtain and extract relevant information (the iterative retrieval logic will be explained in detail in subsequent embodiments). After all subtasks are completed, the result generation module in the electronic device is invoked to integrate, analyze, and format the processing results of each subtask, generating a standardized target processing result. Finally, the target processing result is fed back to the user through an interactive interface. In this way, executing subtasks according to dynamic dependencies ensures the logical rationality of task execution. Through multi-source information retrieval and result integration, the generated target processing result is comprehensive and accurate, solving the problems of information gaps and low result quality in traditional solutions, while also adapting to the personalized needs of users.

[0053] This application provides a data processing method. After receiving a user-input task request, the method first performs intent analysis and complexity classification on the task request to generate a standardized target task request that meets processing requirements. Then, based on the complexity classification information of the target task request, it generates a task execution plan containing at least one subtask and dynamic dependencies between subtasks. Finally, it executes the task execution plan to obtain the target processing result. This method achieves accurate request routing and optimized resource allocation through pre-processing task evaluation and classification, adaptive generation of execution paths through dynamic dependency task planning, and end-to-end result output through automated execution and integration. Therefore, it comprehensively solves the problems of rigid processing links, high resource consumption, passive information acquisition, and unstable result quality in existing technologies, achieving the goals of reducing inference costs, improving processing efficiency, and increasing result accuracy.

[0054] The data processing method provided in this application can be applied to image-based multimodal task processing scenarios without reconstructing the overall data processing architecture. When a user inputs a multimodal task request containing an image, only the original lightweight text model in the system needs to be replaced with a visual language (VL) model (such as InternVL, Qwen-VL, etc.). This allows for professional analysis and in-depth interpretation of image information through the VL model, while retaining the entire process logic of intent analysis, task planning, and information retrieval. Finally, the image analysis results are deeply integrated with the text retrieval information to output a multimodal in-depth research report containing image and text annotations, data support, and conclusions and recommendations. This multimodal extension capability can adapt to the needs of multiple fields such as medical image analysis, industrial quality inspection and evaluation, and product design judgment. For example, if a user uploads a product defect image and inputs "analyze the defect type and rectification suggestions," the system can identify the defect features through the VL model and simultaneously retrieve relevant process standard text information, ultimately generating a multimodal report containing defect image annotations, defect analysis, and rectification plans, thus expanding the applicability of this application.

[0055] Based on the above embodiments, in order to further optimize the line-of-sight approach in each stage of the data processing process, this application also provides a variety of implementation methods to enhance the flexibility, accuracy and intelligent processing of the above data processing process.

[0056] In some embodiments of this application, the process of processing a task request to generate a target task request may include: performing intent recognition and complexity classification on the task request to obtain a classification result; and in response to the classification result being a complex classification, performing information completion processing on the task request to obtain the target task request.

[0057] In this embodiment, intent recognition can identify the core intent in a task request, such as information retrieval, content creation, or planning, through semantic analysis technology. Complexity classification can be based on preset quantitative indicators (e.g., "simple with ≤2 processing steps, complex with ≥3 processing steps"; "simple requiring retrieval of 1 data source, complex with ≥3 data sources"; "simple without information completion, complex with information completion required"), combined with the intent type. To improve processing efficiency, a classification model can be pre-configured in the electronic device. This model is a pre-trained artificial intelligence model for classifying task requests. Inputting relevant information from the task request into this model will output the corresponding classification result. When the classification result is a complex classification, based on the information gap analysis and incomplete interaction information processing model for complex classification requests, the missing key information can be retrieved from the user to supplement the missing information in the original task request, ensuring the completeness of the target task request.

[0058] Specifically, information gap analysis can determine the core information dimensions required to complete the task based on the intent of the task request (e.g., core dimensions for travel planning include travel demographics, budget, food preferences, attraction preferences, and travel time; core dimensions for industry report writing include time frame, core analytical dimensions, and data source requirements). This is then compared to the original request to identify the missing information dimensions. Interactive completion sends targeted completion queries to the user via an interactive interface, inquiring about the missing core information dimensions and avoiding ineffective follow-up questions. The system then receives the user's completed responses, integrates the original task request with the completed information, and generates a target task request according to a preset format, ensuring that the request contains all the key information required for subsequent task planning and execution.

[0059] This targeted information completion for complex category requests resolves the issue of missing information in the original task request, ensuring that the subsequently generated task execution plan accurately matches user needs and improving the accuracy of the final processing result. Simultaneously, the interactive completion method avoids inconsistencies between system completion and user requirements, guaranteeing information accuracy.

[0060] Furthermore, in this embodiment, in response to an invalid classification result, a rejection message corresponding to the task request is output; in response to a simple classification result, a processing result corresponding to the task request is output.

[0061] Invalid classification requests can include illegal, rule-violating, or valueless requests. Rejection messages can be directly output to avoid invalid processing. For simple classification requests, such as single-information retrieval, task planning and iterative retrieval can be skipped, and the retrieval result can be output directly. Specifically, after receiving invalid classification results from the classification model, the electronic device first uses semantic verification to confirm whether the request is an illegal request (e.g., obtaining trade secrets, illegal transaction inquiries), a violation request (e.g., vulgar or violent requests), or a meaningless request (e.g., garbled text without clear content). After confirmation, a preset rejection message template is invoked to generate targeted rejection information based on the type of invalid request. This information is then fed back to the user through the interactive interface, and all subsequent processing steps are terminated.

[0062] After receiving the simple classification result from the classification model, the electronic device identifies the core intent of the request as a single information query or a common-sense question answer. It then calls a common-sense knowledge base or a single-request interface to quickly retrieve the corresponding answer (e.g., common-sense questions are retrieved directly from the knowledge base, while single-information queries are obtained through a single-request query). The retrieved answer is standardized (e.g., text simplification, format standardization), and the result is fed back to the user through the interactive interface, simultaneously terminating subsequent task planning and iterative retrieval steps.

[0063] In this embodiment, invalid classification requests are processed using a fast rejection mode, avoiding illegal and invalid processing operations while saving computational resources. Furthermore, the rapid processing of simple classification requests skips complex task planning and iterative retrieval steps, improving the processing efficiency of simple requests and resolving the resource waste caused by overprocessing. Thus, this embodiment achieves optimal allocation of system resources and improves overall processing efficiency through differentiated paths of refined processing of complex requests and rapid routing of non-complex requests.

[0064] In this embodiment of the application, to improve the processing efficiency of intent recognition and complexity classification of task requests, a first model can be used. The first model is a language model based on training data labeled with intent categories and complexity levels, and optimized through supervised fine-tuning. The reward function used by the first model is constructed based on the classification accuracy and output normalization of the first model.

[0065] The first model is a language model adapted to the semantic analysis requirements of text task requests. Model training is based on training data labeled with intent categories and complexity levels, and optimized through supervised fine-tuning (SFT). The reward function is constructed based on two dimensions: classification accuracy and output canonicity, achieving precise optimization of the model. For example, the first model can use a lightweight language model (such as Qwen3-4B) as the base model, and complete training and deployment through the following steps to ultimately achieve intent recognition and complexity classification functions:

[0066] First, a raw dataset containing multiple scenarios and types of task requests is constructed, including scenarios such as life consultation, business analysis, academic research, and invalid requests. The dataset size meets the needs of supervised fine-tuning (e.g., 150 supervised fine-tuning cold start data and 1000 reinforcement learning validation data). Then, each task request in the raw dataset is labeled in two dimensions. Intent category labeling can be preset categories such as information query, content creation, planning and arrangement, illegal invalid, violation invalid, and meaningless invalid; complexity level labeling can be three categories: simple, complex, and invalid. Further subdivisions for simple and complex labeling can be made based on criteria such as the number of processing steps and the number of data sources. Finally, a standardized training dataset labeled with intent category and complexity level is formed, divided into a training set, a validation set, and a test set, with a ratio of 8:1:1.

[0067] The labeled training dataset is input into a basic lightweight language model for supervised fine-tuning training. The training objective is to enable the model to learn the mapping relationship between task requests and "intent categories and complexity categories." Then, mini-batch gradient descent can be used, with appropriate learning rates and iteration counts. The error between the predicted and labeled results is calculated through forward propagation, and the model parameters are updated through backpropagation. Finally, the model's classification performance is validated on a validation set. If the classification accuracy does not reach a preset threshold (e.g., 95%), the training parameters (e.g., learning rate, batch size) are adjusted, and retraining is performed until the accuracy requirement is met.

[0068] To further improve the classification accuracy and output standardization of the first model, a two-dimensional reward function can be constructed, such as the reward function R = 0.4R_class + 0.3R_clarify + 0.2R_format + 0.1R_brevity.

[0069] Among them, R_class (classification accuracy, weight can be set to 0.4) measures the matching degree between the model's output intent category and complexity level and the labeling results. A matching degree of 1 results in full marks, and a matching degree of 0 results in 0 marks. R_clarify (clarification rationality, weight can be set to 0.3) measures whether the model's generated completion query accurately covers the missing information points for complex classification requests. Reasonable clarification results in high marks, and unreasonable clarification results in low marks. R_format (output standardization, weight can be set to 0.2) measures whether the model's output classification results and completion queries conform to the preset format (such as standardized labels and concise questions). Standard format results in high marks, and otherwise, low marks. R_brevity (output conciseness, weight can be set to 0.1) measures the conciseness of the model's output content, avoiding redundant expressions. Conciseness results in high marks, and redundancy results in low marks.

[0070] The reward function is integrated into the supervised fine-tuning process of the model, with the goal of maximizing the reward value, to further update the model parameters and complete the optimization of the first model. The trained and optimized first model is deployed to the model inference module of the electronic device. When the electronic device receives a standardized task request, it calls the first model, inputs the task request text, and the model outputs results containing "intent category, complexity level, and completion query (for complex classifications)," providing a basis for subsequent processing.

[0071] In this embodiment, a lightweight language model can be used as the base model, reducing the resource consumption of model inference and adapting to the deployment requirements of electronic devices. Supervised fine-tuning based on training data labeled with dual-dimensional information enables the model to accurately learn the mapping relationship between intent recognition and complexity classification, improving classification accuracy. The construction of a dual-dimensional reward function not only optimizes the model's classification accuracy but also standardizes the model's output format and clarification logic, ensuring the smoothness of subsequent request processing and solving the problem of non-standard output in traditional classification models.

[0072] In embodiments of this application, the process of generating a task execution plan based on the complexity classification information of the target task request may include: splitting the target task request into at least one subtask based on the complexity classification information of the target task request; analyzing the semantic relationships and execution logic between the subtasks to generate a dependency graph representing the execution order of the subtasks; and constructing a task execution plan based on the subtasks and the dependency graph.

[0073] The dependency graph must contain at least one custom dependency relationship that is not pre-defined. During the process of splitting the target task request based on its complexity classification information, the granularity and number of splits are related to the complexity level of the target task request. The split sub-tasks represent independently executable, smallest processing units with clear objectives. For example, different splitting dimensions can be preset according to the task intent type. For instance, tasks with content creation intent can be split into sub-tasks according to the processes of material collection, content writing, and review optimization; planning and scheduling tasks can be split into sub-tasks according to the processes of information query, solution customization, and optimization adjustment. Furthermore, for high-difficulty, complex category requests, the splitting can be done at the smallest granularity, such as into 6 to 8 sub-tasks; for low-difficulty, complex category requests, the splitting can be simplified, such as into 2 or 3 sub-tasks. Then, a unique identifier is set for each split sub-task, clearly defining its objective and execution requirements to ensure that the sub-task can be executed independently. This fine-grained breakdown based on complexity ensures the rationality and feasibility of subtasks, avoiding the problems of overly coarse breakdowns leading to execution difficulties or overly fine breakdowns leading to inefficiency, and laying the foundation for subsequent dependency building.

[0074] In this embodiment, the dependency graph representing the execution order of subtasks includes at least one non-fixed, pre-defined custom dependency. This custom dependency is a hybrid dependency, not simply sequential or parallel, but allows a subtask to depend on partial execution results of multiple other subtasks. The dependency graph visually presents the execution order of the subtasks. For example, an electronic device uses semantic analysis to analyze the goals and execution requirements of each subtask, constructs a semantic association matrix between subtasks, and then combines this with execution logic to determine the corresponding dependencies and generate a dependency graph. These dependencies can include sequential dependencies, parallel dependencies, and custom dependencies (which can also be called hybrid dependencies). A sequential dependency means that subtask B can only start after subtask A has been completed (e.g., "content writing" depends on "material collection"). A parallel dependency means that subtask C and subtask D can start simultaneously without affecting each other (e.g., "keyword A search" and "keyword B search"). Custom dependencies allow subtask C to start without waiting for subtasks A and B to complete fully; it only needs to retrieve "partial search results" from subtasks A and B (e.g., "core technology extraction" can start when "paper retrieval" is 50% complete). Furthermore, electronic devices can employ directed graph drawing algorithms, using subtasks as nodes and dependencies as directed edges, to generate a visual dependency graph, annotating the execution order and dependency type of each node. In this way, through semantic association and execution logic analysis, the constructed dependency relationships closely match the actual processing needs of the task. The introduction of custom dependencies overcomes the binary limitations of traditional serial or parallel execution, enabling subtasks to be executed concurrently, significantly improving the execution efficiency of complex tasks.

[0075] In the process of constructing a task execution plan based on subtasks and dependency graphs, priority can be assigned to each subtask according to the dependency graph (e.g., subtasks executed in parallel have the same priority, and subtasks that depend on other subtasks have a lower priority than the subtasks they depend on). Based on the complexity of each subtask, the execution time of each subtask is estimated to facilitate resource scheduling. Finally, a standardized task execution plan is generated according to a preset format, stored in the electronic device's memory, and a visual version is also generated for subsequent execution and monitoring, improving task processing efficiency.

[0076] In some embodiments of this application, the generation of the task execution plan is achieved through a second model, wherein the second model receives the target task request as input and outputs a task execution plan containing a dependency graph. The second model represents a planning model trained by reinforcement learning, and the reward function of the reinforcement learning is constructed based at least on the evaluation of the completeness of the task execution plan's coverage of the original request, the execution efficiency reflected by the dependency graph, and the quality of the results obtained by the execution plan.

[0077] In this embodiment, the second model uses a planning-specific language model as the base model, which is optimized through reinforcement learning (RL) training and finally deployed to the planning module of the electronic device. First, a language model adapted to the task planning scenario (such as Qwen3-8B) can be selected as the base model, and the model parameters are initialized to give it basic semantic understanding and task decomposition capabilities. Then, a reinforcement learning training framework is built, defining the core elements of agent, environment, action, and reward. The agent, the second model, is responsible for generating the task execution plan. The environment refers to the task processing simulation environment, which can simulate the sub-task execution process and output execution efficiency and result quality. Actions refer to the agent's behavior in generating sub-task decomposition schemes and dependency graphs. State refers to the semantic information of the target task request, the generated sub-tasks, and dependencies. The reward is a reward value calculated based on three-dimensional indicators, used to guide model optimization. In this embodiment, a three-dimensional function can be constructed, such as R_plan = w1R_completeness + w2R_efficiency + w3R_accuracy.

[0078] Among them, R_completeness (coverage completeness, weight w1) measures whether the generated subtasks completely cover all the requirements of the target task request. Coverage = number of actual covered requirement points / total number of requirement points. The higher the coverage, the higher the score. R_efficiency (execution efficiency, weight w2) measures the execution efficiency of the dependency graph, using the total task execution time as an indicator. The shorter the execution time, the higher the score. R_accuracy (result quality, weight w3) measures whether the processing results generated after executing the plan meet the user's needs. It is scored by human evaluation or automatic evaluation model. The higher the score, the higher the overall score.

[0079] The dataset labeled with target task requests and optimal task execution plans is input into the reinforcement learning framework for training. In the initial training phase, the agent generates an initial task execution plan, executes it in a simulation environment, and calculates the reward value. Then, an iterative optimization phase begins, updating the model parameters based on the reward value to enable the model to learn high-reward planning strategies. Subsequent validation and tuning involve verifying the effectiveness of the model-generated plans on a test set. If the average score of the three-dimensional metrics does not reach a preset threshold (e.g., 90 points), the reward function weights or training parameters are adjusted, and retraining is performed until the requirements are met. The trained and optimized second model is deployed to the planning module of an electronic device. When the electronic device generates a target task request, it calls the second model, inputs the target task request text, and the model outputs a task execution plan containing a "subtask list, dependency graph, and execution priority," which is directly used for subsequent task execution.

[0080] During reinforcement learning training, to address the issue of invalid sampling that easily occurs when the second model handles long-chain complex task planning, this embodiment incorporates the CISPO (Clipped Importance Sampling Policy Optimization) strategy: non-critical subtasks in task planning are split into redundant steps with reduced sampling weights, while retaining the sampling weights of core planning steps such as subtask dependency analysis and custom dependency construction. During the inference phase, the CISPO strategy is also used to filter the planning candidate results generated by the model, prioritizing the retention of task execution plans with complete core logic. This improves the training and inference efficiency of the model while avoiding the dilution of core planning logic in traditional sampling strategies, enabling the second model to maintain high coverage completeness and execution efficiency when handling multi-level, long-chain complex task planning.

[0081] In this embodiment, the second model employs reinforcement learning training, enabling it to continuously learn the optimal planning strategy and improve the quality of task execution plan generation. The construction of a three-dimensional reward function ensures that the generated plan not only fully covers user needs and possesses high execution efficiency but also guarantees the quality of the final processing result. The second model directly outputs a task execution plan containing a dependency graph, simplifying the planning process, enhancing the generalization ability of the planning module, and adapting to various complex target task requests.

[0082] Furthermore, the second model in this embodiment provides the architectural foundation for the multimodal expansion of the entire data processing system. Because the task execution plan generation logic is decoupled from the modality type of the processed object, model replacement is only required during the result generation stage to achieve multimodal task planning and execution. When processing image-based multimodal tasks, the lightweight text model in the report generation stage is replaced with a visual language VL model (such as InternVL or Qwen-VL) based on the task execution plan generated by the second model. The VL model can take over the task execution plan output by the second model, simultaneously completing the analysis and interpretation of image information and the integration and processing of text information, ultimately outputting a multimodal in-depth research report. The entire expansion process does not require adjustment to the training logic and task planning rules of the second model; it can be achieved solely through modular model replacement, realizing flexibility and scalability in architectural design and solving the technical problems of single modality and high expansion costs in traditional data processing systems.

[0083] In some embodiments of this application, the process of executing a task execution plan and obtaining a target processing result includes: performing information search and extraction operations for each subtask in the task execution plan, wherein the information search and extraction operations include: generating a search query based on the current subtask; calling a retrieval service to obtain retrieval results matching the search query; in response to the sufficiency condition of the retrieval results, generating an optimized search query based on the retrieval results, and performing information search and extraction operations based on the optimized search query until the target retrieval result meets the sufficiency condition; and generating a target processing result based on the target retrieval results corresponding to each subtask.

[0084] The electronic device executes iterative information search and retrieval operations for each subtask sequentially, according to the dependency graph of the task execution plan. First, based on the goal and execution requirements of the current subtask, a precise initial search query is generated. The query statement includes core keywords and limiting conditions (such as time, data source, and content type). The electronic device's retrieval service module is invoked to obtain search results matching the initial search query. These results include web pages, papers, data reports, and documents. Then, a sufficiency check is performed. Pre-set sufficiency check indicators (such as "relevance ≥ 90%, comprehensiveness ≥ 80%, timeliness ≤ 1 year") are used to evaluate the search results. If all indicators are met, the search result is taken as the "target search result." If not, the search query is optimized, analyzing the shortcomings of the search results (such as "low relevance," "incomplete information," and "poor timeliness"), and optimizing the search query based on these shortcomings (such as adding keywords, adding limiting conditions, and adjusting the query order). Based on the optimized search query, the retrieval service is invoked again to obtain search results, and the sufficiency check is repeated until the target search result that meets the conditions is obtained. By employing iterative information search and extraction operations, the problems of insufficient information and low accuracy in a single retrieval are solved. Through continuous optimization of query statements, it is ensured that the obtained target retrieval results can accurately, comprehensively, and timely meet the execution requirements of subtasks, laying the foundation for the final generation of high-quality target processing results.

[0085] After obtaining the target retrieval results for each subtask, an information preprocessing step can be performed. This involves deduplicating (removing duplicate information), reducing noise (filtering out invalid and low-authority information), and classifying (by content type, such as data, opinions, and cases) the target retrieval results for each subtask. Then, based on the requirements of the target task request, the preprocessed information is integrated, the relationships between information are analyzed, and key content such as core conclusions, data support, and action recommendations are extracted. Finally, based on the type of task request, a preset result template (such as a report template, itinerary template, or data list template) is selected, and the integrated and analyzed information is filled into the template to generate a structured target processing result, ensuring that the result format is standardized, logically clear, and comprehensive. In this way, through information preprocessing and integration analysis, the problems of duplicate and redundant result content are solved. The structured result generation method makes the target processing result conform to the user's reading and usage habits, improving the usability and satisfaction of the result.

[0086] Correspondingly, if the current task is an image-based multimodal task, this step will be specifically adapted during the structured result generation stage: First, the original lightweight text model used for information integration and result generation will be replaced with a visual language VL model (such as InternVL or Qwen-VL). This VL model will perform object detection, feature extraction, and semantic interpretation on the image information in the multimodal task, and convert the image analysis results into standardized structured text information. Subsequently, the structured text derived from the image will be deeply integrated with the text-based target retrieval results corresponding to each subtask. A preset image-text combination template will be selected according to the requirements of the multimodal task. The results will retain the key image region annotations, the correlation between image analysis conclusions and text information, and finally generate a multimodal in-depth research report that combines image visualization and text analysis. This processing method ensures both the integrity and accuracy of the multimodal results and the standardization and readability of the result format. At the same time, the image analysis and text integration process fully follows the original task execution plan without the need for additional adjustments to the subtask execution logic, thus improving the efficiency and stability of multimodal task processing.

[0087] In some embodiments of this application, the information search and extraction operations are performed through a third model. The third model is a search model trained by reinforcement learning, which evaluates the sufficiency of information and generates or optimizes search queries based on the current subtask and the retrieved results; wherein, the reward function of reinforcement learning is constructed based on the deviation of the retrieved results obtained under the guidance of the third model from the benchmark retrieved results in terms of the final generated target processing result.

[0088] The benchmark search result is characterized by using conventional search schemes (such as the RAG system based on keyword matching, or the default search results of a single search engine) as a benchmark, and is denoted as RAG. baseThe target retrieval result represents the retrieval results obtained after generating or optimizing the query using a third-party model, denoted as RAG. model The reward function can be expressed as follows:

[0089] R search =Acc final (RAG model )- Acc final (RAG base )

[0090] Among them, Acc final (RAG model ) indicates the use of RAG model After the information is generated and processed to achieve the target, the quality accuracy of the results is evaluated (covering information accuracy, completeness, and relevance); Acc final (RAG base ) indicates the use of RAG base After the information is generated and processed to achieve the target result, the quality and accuracy of the result are evaluated.

[0091] Furthermore, if R search A value greater than 0 indicates that the optimized retrieval results of the third model have brought information gain compared to the baseline solution, and the model receives a positive reward; if R0... search A value ≤0 indicates that the search results did not bring incremental value, and the model received zero or negative rewards. Furthermore, to accelerate model convergence, auxiliary reward terms can be introduced, such as a search round penalty (the fewer the iterations, the higher the reward), to prevent the model from getting stuck in infinite iteration.

[0092] The reinforcement learning training and iterative optimization process of the third model includes: For a batch of sub-tasks, the third model generates initial queries and executes iterative retrieval, recording the trajectory data of the query sequence, retrieval results, and final reward value. Then, using the PPO (Proximal Policy Optimization) algorithm, the policy gradient is calculated based on the reward value of the trajectory data, updating the model's query generation and sufficiency evaluation network parameters. During training, for long-chain retrieval tasks, the CISPO (Clipped Importance Sampling Policy Optimization) training strategy is adopted to prune the sampling weights of unimportant retrieval steps, rather than directly pruning key reflective tokens (such as optimization instructions like "supplementing retrieval of newly released guidance strategies in 2026"), ensuring the logical coherence of long-chain iterative retrieval. Furthermore, a test set can be set, such as one containing sub-tasks of different domains and complexities. If the average R-value of the third model on the test set... searchTraining is complete when a preset threshold (e.g., ≥15%) is reached and the average number of iterations is ≤3. Otherwise, the reward function weights or model hyperparameters are adjusted, and training continues. The trained third model is deployed to the enhanced search module of the electronic device and linked with the retrieval service module. The specific inference process may include: receiving the current subtask issued by the planning module; generating an initial search query and calling the retrieval service; receiving the retrieval results and determining whether the conditions are met through the model's sufficiency evaluation head; if not, generating a new query through the model's query optimization head and entering the next round of retrieval; if satisfied, outputting the target retrieval result and feeding back the "query optimization trajectory" to the model for offline fine-tuning. In this embodiment, through reinforcement learning and round penalty, the model can quickly learn the strategy of obtaining optimal information with the fewest iterations, avoiding resource waste. Combined with the CISPO (Clipped Importance Sampling Policy Optimization) strategy, it is ensured that key optimization logic is not lost when handling complex and long-step retrieval tasks, improving the robustness of the model.

[0093] Furthermore, in this embodiment of the application, calling the retrieval service to obtain retrieval results matching the search query includes calling a multi-source retrieval system, which integrates at least two retrieval services: web page content crawling service based on condition parsing, network search engine application programming interface service, and local document retrieval service based on semantic understanding of a large language model.

[0094] Based on the above embodiments, to further improve the comprehensiveness and accuracy of information acquisition, this application also provides a preferred embodiment of a multi-source retrieval system. Specifically, when performing information search and extraction operations, calling a retrieval service to obtain retrieval results matching the search query includes calling a multi-source retrieval system, which integrates at least two of the following retrieval services:

[0095] This web content crawling service, based on conditional parsing, uses automated tools to crawl and structure target web pages. In its implementation, open-source toolchains such as Jina AI and Crawl4AI can be used to target relevant web pages based on user query intent and preset crawling conditions. During the crawling process, the system first parses the web page's DOM structure, extracting key metadata such as the main content, publication time, and title. Then, it uses a large language model to filter the crawled content, removing irrelevant information such as advertisements and navigation bars, retaining only highly relevant and valid content. For example, when the subtask is "to obtain the latest revision status of a policy document," the system can target relevant pages on government websites and extract core information such as the policy document's publication time and revised clauses through conditional parsing.

[0096] The system provides an application programming interface (API) service for a web-based search engine. This service connects to the open APIs of mainstream search engines, enabling rapid retrieval of real-time information from the internet. In its implementation, the system concurrently calls multiple search engine APIs based on the search query generated by the current subtask, obtaining a diverse list of search results. Subsequently, the system deduplicates, sorts, and performs preliminary filtering on the returned results, extracting key information such as titles, summaries, and links for subsequent in-depth crawling or direct citation. By concurrently calling multiple search engines, the system effectively improves information recall and avoids the problems of biased or incomplete results from a single search engine.

[0097] This local document retrieval service, based on large language model semantic understanding, targets local knowledge bases or offline document sets. Leveraging the enhanced semantic understanding capabilities of a large language model, it achieves high-precision information retrieval. In its implementation, the system first preprocesses local documents: dividing the document content into paragraphs, using a long context model to perform semantic segmentation and summary generation on the segmented paragraphs, and constructing an index triple (Index, Title, summary) for each paragraph to form a semantically enhanced retrieval index. When a search query is received, the system uses the large language model to perform semantic understanding of the query, semantically matching it with the summary in the index, and recalling the document paragraphs that best match the query intent. Compared to traditional keyword-based or embedding vector-based retrieval methods, this method can more accurately understand the deeper intent of the query, making it particularly suitable for retrieval scenarios involving specialized fields and complex concepts. For example, when searching for "solid-state battery interface stability issues," the system can accurately recall document paragraphs involving in-depth technical discussions such as interface reactions and material matching, rather than just fragments that superficially mention keywords.

[0098] In actual operation, the multi-source retrieval system intelligently selects or combines the aforementioned retrieval services based on the needs and context of the current subtask. For example, for subtasks requiring the latest updates, it prioritizes calling online search engine services; for subtasks requiring in-depth professional knowledge or authoritative documents, it prioritizes calling local document retrieval services; and for subtasks requiring specific webpage content, it calls webpage content crawling services. The system can also merge and reorder the results returned by multiple retrieval sources, further improving the accuracy and reliability of information through cross-validation and redundancy elimination. By integrating multiple heterogeneous retrieval services, the multi-source retrieval system of this application achieves a unity of breadth and depth in information acquisition, ensuring both timely acquisition of real-time dynamic information and accurate retrieval of in-depth professional information, providing a rich and reliable information foundation for subsequent report generation.

[0099] Furthermore, the multi-source retrieval system in this embodiment can work collaboratively with the third model, adapting to the CISPO (Clipped Importance Sampling Policy Optimization) strategy. That is, during the fusion and ranking stage of multi-source retrieval results, the system trims the fusion weights of retrieval results with low authority and low relevance, focusing on retaining the fusion weights of high-value webpage parsing results and local document results with core semantic matching. This logic aligns with the CISPO strategy's objective of "trimming low-value results and retaining core results," enabling efficient linkage between the multi-source retrieval system's result fusion and the third model's sufficiency judgment, further improving the overall efficiency and accuracy of data processing.

[0100] like Figure 2 The diagram illustrates the core architecture of a multi-source retrieval system provided in this application embodiment. The multi-source retrieval system uses a search tool server as the top-level scheduling node, and its query requests are distributed to a parallel web search module and a web page parser module. The web search module further connects to at least three independent search engines (search engine 1, search engine 2, and search engine 3) to acquire a wide range of web resources, and its output is preliminary search results containing URLs and summary fragments. The web page parser module connects to a dedicated web crawling and parsing tool, responsible for deep cleaning and core information extraction of high-value web pages selected by the web search module, and inputting the parsed content into a large language model for semantic understanding and targeting, ultimately generating accurate target information.

[0101] This architecture achieves comprehensive coverage and efficient utilization of publicly available online information through the coordinated efforts of multi-engine parallel retrieval and deep webpage parsing. Data flow shows that query commands are distributed downwards from the search tool server, forming broad preliminary search results in the web search module, and then refined into precise information by the webpage parser module. Both types of results are ultimately integrated and fed back by the search tool server, providing solid and high-quality data support for the information sufficiency assessment of the third model.

[0102] See Figure 3This diagram illustrates a document index creation process provided in an embodiment of this application. This process corresponds to the implementation stage of the local document retrieval service based on semantic understanding using a large language model in this application. It aims to improve retrieval accuracy through refined semantic processing. The specific implementation steps are as follows: First, obtain the original document set to be processed. These documents can be various textual materials such as research reports, technical documents, and policy documents from a local knowledge base, serving as the basic data source for subsequent information retrieval. Second, segment the original documents. According to the natural paragraph structure of the document or a preset text length threshold, each document is divided into several continuous and semantically relatively complete text blocks, including text block 0, text block 1, text block 2, up to text block N. The choice of segmentation granularity balances semantic integrity and retrieval efficiency, ensuring that each text block contains sufficient contextual information to support independent understanding. Next, for each text block, call the large language model for segmented semantic understanding and summary generation. The large language model performs deep analysis on the current text block, extracting its core theme and key information, and generating the structured index information corresponding to that text block. Specifically, for each text block, the large language model outputs an index triple containing the following elements: Title: A high-level summary of the core theme of the text block, usually presented in the form of a phrase, which facilitates quick identification of the content scope of the text block; Content Summary: A concise summary of the key information of the text block, covering the main viewpoints, data conclusions, or event descriptions; Content: Sub-points or key entity information that can be further subdivided as needed, as a supplement to the content summary.

[0103] Repeat the above processing steps to generate corresponding index triples for all text blocks until the index of the entire document collection is completed. The final index is stored in a structured format of "title-content summary-supplementary information," ensuring that the semantic meaning of each text block is accurately extracted and solidified. In the subsequent retrieval stage, when the system receives a search query, it uses a large language model to semantically understand the query and performs semantic matching with the title and content summary in the index to quickly locate the text block most relevant to the query intent. Compared to traditional keyword-based or embedding vector-based retrieval methods, this method fully utilizes the deep semantic understanding capabilities of the large language model, enabling it to more accurately capture the implicit intent and contextual relationships of the query, especially suitable for retrieval scenarios with dense technical terms and complex conceptual relationships. For example, when the query involves the technical concept of "solid-state battery interface stability," the system can accurately recall text blocks in the index that involve in-depth technical discussions such as interface reaction mechanisms and material matching issues, rather than just fragments that superficially mention keywords. Through the document index creation process described above, this application achieves high-precision semantic modeling of local document resources, providing core support for local document retrieval services in multi-source retrieval systems, and significantly improving the accuracy of information retrieval and the semantic relevance of retrieval results.

[0104] See Figure 4 This illustration shows a schematic diagram of a system architecture for deep research applications provided in an embodiment of this application. The architecture presents a planning agent (such as...) Figure 4 The planner shown), and the retrieval agent (such as...) Figure 4 The intelligent search engine shown), and the tool service intelligent agent (such as...) Figure 4 The tool service section shown), generating intelligent agents (such as...) Figure 4 The following section details the implementation process of the four main core intelligent agents (shown in the generative large language model) working together to process user task requests, using application scenarios from the solid-state battery industry as an example.

[0105] Users submit task requests through an interactive interface, such as "Please analyze the progress and commercialization challenges of solid-state battery technology in 2024." After receiving the task request, the system passes it to the planning agent. Figure 4 The planner shown here uses a large language model to perform a preliminary analysis of the task and determine whether additional information from the user is needed. In this scenario, the planner identifies that the task request does not clearly specify the technical indicators, company scope, or data timeframes. Therefore, it initiates a user confirmation and rewrite process, sending a supplementary question to the user: "Please confirm whether specific technical parameter comparisons, core company dynamics, and cost analysis are required? Is it necessary to limit the specific region or time range?" The user confirms: "Energy density parameters, major companies' mass production plans, and cost comparisons are required, with a time range of 2024 and no regional restrictions." After receiving the supplementary information, the planner generates a standardized target task request based on the original request: "Analyze the progress and commercialization challenges of solid-state battery technology in 2024, focusing on energy density parameters, major companies' mass production plans, and cost comparisons."

[0106] The planning agent (i.e., the planner) generates an evaluation result based on the target task request, including a difficulty level classification of complex, a reason for classification indicating the need for multi-dimensional information retrieval and cross-analysis, a label indicating supplementary information, and an output of a rewritten task description. Since the task is classified as complex, the planning agent (i.e., the planner) breaks it down into multiple sub-tasks, specifically: Sub-task 1 "Collecting progress on solid-state battery energy density technical parameters in 2024", Sub-task 2 "Obtaining mass production plans and commercialization dynamics of solid-state batteries from major companies", Sub-task 3 "Analyzing the cost structure and cost reduction paths of solid-state batteries", Sub-task 4 "Investigating the policy support for solid-state batteries in various countries", and Sub-task 5 "Integrating the above information to generate a structured analysis report". During the decomposition process, the planning agent (i.e., the planner) analyzes the semantic relationships and execution logic between the sub-tasks, generating dependencies: Sub-task 1 and Sub-task 2 can be executed in parallel; Sub-task 3 depends on the cost-related data obtained from Sub-task 1 and Sub-task 2; Sub-task 4 is executed independently; Sub-task 5 depends on all execution results of Sub-task 1, Sub-task 2, Sub-task 3, and Sub-task 4.

[0107] The system processes each subtask sequentially according to their dependency order. Taking subtask 1 as an example, it retrieves the agent (such as...). Figure 4 The intelligent search engine shown generates an initial search query based on the current subtask "Collecting progress on solid-state battery energy density technology parameters in 2024" and calls the tool service to perform information retrieval. The tool service intelligent agent (such as...) Figure 4 The tool services shown integrate multiple search methods: it retrieves technical documents from the local knowledge base through native search enhancement services; it calls webpage parsing services to perform targeted crawling and content parsing of industry media and research institution websites; and it uses web search services to call search engine APIs to obtain real-time dynamic information. The search agent evaluates information sufficiency after each search. For example, the initial search might retrieve information such as manufacturer A announcing an energy density of 500Wh / kg and manufacturer B's semi-solid-state battery energy density exceeding 400Wh / kg, but lacks specific technical roadmap descriptions. The intelligent search engine determines the information is insufficient and generates an optimized search query, "Details of manufacturer A's solid-state battery technology roadmap sulfide electrolyte energy density," for iterative searching until detailed information including the technology roadmap, testing conditions, and release date is obtained, satisfying the information sufficiency requirement.

[0108] After each search, the system outputs the search results in a standardized format, using the <Important Information> tag to mark core data such as "Manufacturer A's sulfide system has an energy density of 500Wh / kg, and plans to mass-produce in 2027," along with document information and search status markers. After executing subtask 1, the system checks for dependencies. Subtask 1, as an independent subtask, has no prerequisite dependencies and is directly marked as complete, with the search results stored. The system processes subtasks 2, 3, and 4 sequentially, each executing a similar iterative search process. For subtask 3, which has dependencies, the system first obtains cost-related search results from subtasks 1 and 2 as input before starting execution.

[0109] After all subtasks have been executed, the system determines that all subtasks are complete and calls the generation of the intelligent agent (i.e. Figure 4 The generative large language model shown integrates, analyzes, and restructures the important information obtained from all subtasks. The generative agent, under fixed parameter conditions, synthesizes a structured research report based on a preset report template. This report includes an executive summary, extracting core conclusions such as breakthroughs in solid-state battery technology, company dynamics, cost challenges, and policy support; the main analysis section details progress in each dimension, presenting different manufacturers' technical indicators and mass production plans in comparative tables; the action recommendations section proposes high, medium, and low priority development suggestions based on the analysis conclusions; and finally, all cited sources are included. The system outputs the generated structured research report as the final result to the user.

[0110] Through the above process, this application embodiment realizes end-to-end automated processing from user solid-state battery research needs to structured analysis reports. The planning agent is responsible for overall decision-making, task decomposition and process scheduling. The retrieval agent is mainly used for sub-task execution, information acquisition and iterative optimization. The tool service agent provides diversified and highly accurate retrieval capabilities to support the entire process. The generation agent completes the final content integration, analysis and report output. These agents are centered on the task data platform, realizing the real-time transmission, updating and sharing of task requests, analysis results, retrieval data and progress status. Each agent focuses on its own task processing, data communication and cooperation, realizing efficient processing and high-quality results for complex research tasks.

[0111] In another embodiment of this application, a data processing apparatus is also provided, comprising:

[0112] The receiving unit is used to receive task requests input by the user.

[0113] The first generation unit is used to process the task request and generate a target task request, wherein the target task request represents a request that meets the processing requirements after intent analysis and complexity classification.

[0114] The second generation unit is used to generate a task execution plan based on the complexity classification information of the target task request. The task execution plan represents the execution plan of at least one subtask included in the target task request and the dynamic dependencies between the subtasks.

[0115] An execution unit is used to execute the task execution plan and obtain the target processing result.

[0116] It should be noted that the specific implementation of each unit in this embodiment can be referred to the corresponding content above, and will not be described in detail here.

[0117] In another embodiment of this application, a readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the data processing method described above.

[0118] In another embodiment of this application, an electronic device is also provided, see [link to relevant documentation]. Figure 5 ,include:

[0119] Memory 501 is used to store computer programs and data generated by the execution of said computer programs;

[0120] Processor 502 is configured to execute the computer program to achieve:

[0121] Receive task requests from user input;

[0122] The task request is processed to generate a target task request, which represents a request that meets the processing requirements after intent analysis and complexity classification.

[0123] Based on the complexity classification information of the target task request, a task execution plan is generated. The task execution plan represents the execution plan of at least one subtask included in the target task request and the dynamic dependencies between the subtasks.

[0124] Execute the task execution plan to obtain the target processing result.

[0125] In some possible implementations, processing the task request to generate a target task request includes:

[0126] The task request is subjected to intent recognition and complexity classification to obtain the classification result;

[0127] In response to the classification result being a complex classification, the task request is processed to complete the information, thereby obtaining the target task request.

[0128] Among the possible implementations are:

[0129] In response to the classification result being invalid, a rejection message corresponding to the task request is output.

[0130] In response to the classification result being a simple classification, the processing result corresponding to the task request is output.

[0131] In some possible implementations, the intent recognition and complexity classification of the task request is achieved through a first model; wherein, the first model is a language model based on training data labeled with intent categories and complexity levels, and optimized through supervised fine-tuning; the reward function used by the first model is constructed according to the classification accuracy and output normalization of the first model.

[0132] In some possible implementations, generating a task execution plan based on the complexity classification information of the target task request includes:

[0133] Based on the complexity classification information of the target task request, the target task request is split into at least one sub-task;

[0134] Analyze the semantic relationships and execution logic between each subtask to generate a dependency graph representing the execution order of the subtasks, wherein the dependency graph contains at least one non-fixed preset custom dependency relationship;

[0135] Based on the subtasks and the dependency graph, the task execution plan is constructed.

[0136] In some possible implementations, the generation of the task execution plan is achieved through a second model, wherein the second model receives the target task request as input and outputs the task execution plan containing the dependency graph; the second model represents a planning model trained by reinforcement learning, and the reward function of the reinforcement learning is constructed based at least on the completeness of the task execution plan's coverage of the original request, the execution efficiency reflected by the dependency graph, and the quality of the results obtained by executing the plan.

[0137] In some possible implementations, executing the task execution plan to obtain the target processing result includes:

[0138] For each subtask in the task execution plan, an information search and extraction operation is performed, wherein the information search and extraction operation includes: generating a search query based on the current subtask; calling a retrieval service to obtain retrieval results matching the search query; in response to the sufficiency condition not being met by the retrieval results, generating an optimized search query based on the retrieval results, and performing information search and extraction operations based on the optimized search query until the target retrieval result meets the sufficiency condition;

[0139] Based on the target retrieval results corresponding to each subtask, the target processing results are generated.

[0140] In some possible implementations, the execution of information search and extraction operations is implemented through a third model; the third model is a search model trained by reinforcement learning, which evaluates the sufficiency of information and generates or optimizes search queries based on the current subtask and the retrieved results; wherein, the reward function of the reinforcement learning is constructed based on the deviation of the retrieved results obtained under the guidance of the third model from the benchmark retrieved results in terms of the quality of the final generated target processing result.

[0141] In some possible implementations, the invocation of the retrieval service to obtain retrieval results matching the search query includes invoking a multi-source retrieval system; the multi-source retrieval system integrates at least two of the following retrieval services: web page content crawling service based on condition parsing, online search engine application programming interface service, and local document retrieval service based on semantic understanding of a large language model.

[0142] It should be noted that the specific implementation of the processor in this embodiment can be referred to the corresponding content above, and will not be described in detail here.

[0143] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0144] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0145] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0146] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A data processing method, comprising: Receive task requests from user input; The task request is processed to generate a target task request, which represents a request that meets the processing requirements after intent analysis and complexity classification. Based on the complexity classification information of the target task request, a task execution plan is generated. The task execution plan represents the execution plan of at least one subtask included in the target task request and the dynamic dependencies between the subtasks. Execute the task execution plan to obtain the target processing result.

2. The method according to claim 1, wherein processing the task request to generate a target task request includes: The task request is subjected to intent recognition and complexity classification to obtain the classification result; In response to the classification result being a complex classification, the task request is processed to complete the information, thereby obtaining the target task request.

3. The method according to claim 2, further comprising: In response to the classification result being invalid, a rejection message corresponding to the task request is output. In response to the classification result being a simple classification, the processing result corresponding to the task request is output.

4. The method according to claim 2, wherein the intent recognition and complexity classification of the task request are implemented through a first model; wherein, The first model is a language model that is trained on training data labeled with intention categories and complexity levels and optimized through supervised fine-tuning; the reward function used by the first model is constructed based on the classification accuracy and output normalization of the first model.

5. The method according to claim 1, wherein generating a task execution plan based on the complexity classification information of the target task request includes: Based on the complexity classification information of the target task request, the target task request is split into at least one sub-task; Analyze the semantic relationships and execution logic between each subtask to generate a dependency graph representing the execution order of the subtasks, wherein the dependency graph contains at least one non-fixed preset custom dependency relationship; Based on the subtasks and the dependency graph, the task execution plan is constructed.

6. The method according to claim 5, wherein the generation of the task execution plan is implemented through a second model, wherein, The second model receives the target task request as input and outputs the task execution plan containing the dependency graph; the second model represents a planning model trained by reinforcement learning, and the reward function of the reinforcement learning is constructed based on at least the completeness of the task execution plan in covering the original request, the execution efficiency reflected by the dependency graph, and the quality of the results obtained by executing the plan.

7. The method according to claim 1, wherein executing the task execution plan to obtain the target processing result includes: For each subtask in the task execution plan, an information search and extraction operation is performed, wherein the information search and extraction operation includes: generating a search query based on the current subtask; calling a retrieval service to obtain retrieval results matching the search query; in response to the sufficiency condition not being met by the retrieval results, generating an optimized search query based on the retrieval results, and performing information search and extraction operations based on the optimized search query until the target retrieval result meets the sufficiency condition; Based on the target retrieval results corresponding to each subtask, the target processing results are generated.

8. The method according to claim 7, wherein the information search and extraction operation is implemented through a third model; the third model is a search model trained by reinforcement learning, which evaluates the sufficiency of information and generates or optimizes the search query based on the current subtask and the obtained search results; wherein, The reward function of the reinforcement learning is constructed based on the deviation of the retrieval results obtained under the guidance of the third model from the benchmark retrieval results in terms of the quality of the final generated target processing results.

9. The method according to claim 8, wherein calling the retrieval service to obtain retrieval results matching the search query includes calling a multi-source retrieval system; the multi-source retrieval system integrates at least two of the following retrieval services: web page content crawling service based on condition parsing, network search engine application programming interface service, and local document retrieval service based on large language model semantic understanding.

10. An electronic device, comprising: A memory for storing computer programs and the data generated by the execution of said computer programs; A processor for executing the computer program to achieve: Receive task requests from user input; The task request is processed to generate a target task request, which represents a request that meets the processing requirements after intent analysis and complexity classification. Based on the complexity classification information of the target task request, a task execution plan is generated. The task execution plan represents the execution plan of at least one subtask included in the target task request and the dynamic dependencies between the subtasks. Execute the task execution plan to obtain the target processing result.

Images