A data generation method and device based on a simulated search environment, a storage medium, and an apparatus

Abstract

The application discloses a data generation method and device based on a simulation search environment, a storage medium and equipment, relates to the technical field of communication, and mainly aims to solve the problem of poor data effectiveness in the existing search environment model training process. Including: obtaining the task content of a data generation task, and determining the task type based on the search complexity of the task content, the task type including an expansion type and a refinement type; processing the task content based on the task type to obtain a plurality of candidate tasks, and performing quality detection on the candidate tasks based on a first intelligent agent; when the quality detection is passed, performing multi-round solving on the candidate tasks in a simulation search environment based on a second intelligent agent to generate a solving track, and evaluating the solving track based on a third intelligent agent, the simulation search environment including a simulation webpage search environment and a simulation webpage browsing environment; and if the evaluation result matches a preset task constraint condition, generating target data of the data generation task.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a data generation method, apparatus, storage medium, and device based on a simulated search environment. Background Technology

[0002] With the rapid development of large language models, information retrieval, and intelligent agent technologies, training search-oriented intelligent agents for solving complex problems has become an important research direction in the field of artificial intelligence. In particular, for deep search tasks involving multi-turn retrieval, cross-page browsing, and progressive evidence integration, some machine learning models cannot complete the answer solely based on their static parameter knowledge. They must obtain external information by repeatedly calling search and browsing tools and combine this with multi-step reasoning to arrive at the result. Therefore, constructing high-quality task data suitable for training deep search intelligent agents is crucial to improving the deep search and reasoning capabilities of these agents.

[0003] Currently, the data required for model training mainly relies on real search environment data. However, training based on real search environments is costly, unstable, and lacks reproducibility. Continuous API calls lead to high training costs, and dynamic changes in the external environment cause fluctuations in training data, making it difficult to stabilize the data and manage the training process. This results in high system pressure when using real data for model training, thus affecting the processing effectiveness of search tools. Summary of the Invention

[0004] In view of this, this application provides a data generation method, apparatus, storage medium and device based on a simulated search environment, the main purpose of which is to solve the problem of poor data validity in the training process of existing search environment models.

[0005] According to one aspect of this application, a data generation method based on a simulated search environment is provided, comprising: The task content of the data generation task is obtained, and the task type is determined based on the search complexity of the task content. The task type includes extended type and refined type. The task content is processed based on the task type to obtain multiple candidate tasks, and the candidate tasks are quality checked based on the first intelligent agent. If the quality check is passed, the candidate task is solved in multiple rounds in a simulated search environment based on the second agent to generate a solution trajectory, and the solution trajectory is evaluated based on the third agent. The simulated search environment includes a simulated web search environment and a simulated web browsing environment. If the evaluation result matches the preset task constraints, the target data for the data generation task is generated.

[0006] Furthermore, determining the task type based on the search complexity of the task content includes: The search complexity is determined based on the task content, and the search complexity includes difficulty information and solution trajectory summary information; When the difficulty information is lower than the preset complexity requirement, or when the solution trajectory summary information does not match the search path conditions, the task type is determined to be an extended type. When the difficulty information is higher than the preset complexity requirement, and the solution trajectory summary information matches the search path conditions, the task type is determined to be a refinement type.

[0007] Furthermore, the process of processing the task content based on the task type to obtain multiple candidate tasks includes: When the task type is an extended type, the first preset prompt word template and the first structured input / output constraints are retrieved; Based on the preset prompt word template and the structured input and output constraints, the fourth agent is searched and configured to generate the target entity, intermediate bridging entity, and bridging relationship between the target entity and the intermediate bridging entity in the task content. Under entity constraints, the task content is expanded based on the target entity, intermediate bridging entities, and the bridging relationship to obtain multiple candidate tasks.

[0008] Furthermore, the process of processing the task content based on the task type to obtain multiple candidate tasks includes: When the task type is a refinement type, the second preset prompt word template, the refinement template, and the second structured input / output constraints are retrieved. The fourth agent is refined and configured based on the second preset prompt word template, the refined template, and the structured input and output constraints to generate a refined strategy. Based on the refinement strategy, the task content is refined and extracted to obtain multiple candidate tasks.

[0009] Furthermore, before generating the solution trajectory by performing multiple rounds of solving the candidate task in a simulated search environment based on the second intelligent agent, the method further includes: Based on preset search prompts, a fifth agent is invoked to perform simulated searches for different search inputs, resulting in candidate search results. Simulated search webpage content is then constructed based on the candidate search results and search input information, and this is determined as the simulated webpage search environment. Based on preset browsing prompts, a sixth agent is invoked to simulate browsing different browsing inputs, resulting in candidate browsing results. Based on the candidate browsing results and browsing input information, simulated webpage content is constructed and determined as the simulated webpage browsing environment.

[0010] Furthermore, before acquiring the task content of the data generation task, the method further includes: Obtain the initialization task content, which is used to characterize the target generation content required for different data generation tasks; When a task status update instruction is received, the initial task content is updated based on the task content carried in the task status update instruction.

[0011] Furthermore, the quality detection of the candidate tasks based on the first intelligent agent includes: Based on the detection prompts, the first agent is invoked to perform quality detection on the candidate tasks. The detection prompts include uniqueness prompts, pseudo-multi-hop prompts, and leakage prompts. The method further includes: If the quality check fails, a rollback operation is performed to generate a task status update instruction.

[0012] According to another aspect of this application, a data generation apparatus based on a simulated search environment is provided, comprising: The acquisition module is used to acquire the task content of the data generation task and determine the task type based on the search complexity of the task content. The task type includes extended type and refined type. The processing module is used to process the task content based on the task type to obtain multiple candidate tasks, and to perform quality detection on the candidate tasks based on the first intelligent agent. An evaluation module is used to, when a quality check is passed, perform multiple rounds of solving the candidate task in a simulated search environment based on a second agent to generate a solution trajectory, and evaluate the solution trajectory based on a third agent. The simulated search environment includes a simulated web search environment and a simulated web browsing environment. The generation module is used to generate the target data for the data generation task if the evaluation result matches the preset task constraints.

[0013] Furthermore, the acquisition module is specifically used to determine the search complexity based on the task content, the search complexity including difficulty information and solution trajectory summary information; when the difficulty information is lower than the preset complexity requirement, or the solution trajectory summary information does not match the search path conditions, the task type is determined to be an extended type; when the difficulty information is higher than the preset complexity requirement, and the solution trajectory summary information matches the search path conditions, the task type is determined to be a refined type.

[0014] Furthermore, the processing module is specifically configured to, when the task type is an extended type, retrieve a first preset prompt word template and a first structured input / output constraint; perform search configuration on the fourth agent based on the preset prompt word template and the structured input / output constraint, generate the target entity, intermediate bridging entity, and bridging relationship between the target entity and the intermediate bridging entity of the task content; and, under entity constraint conditions, extend the task content based on the target entity, intermediate bridging entity, and the bridging relationship to obtain multiple candidate tasks.

[0015] Furthermore, the processing module is specifically used to, when the task type is a refinement type, retrieve the second preset prompt word template, the refinement template, and the second structured input-output constraints; refine the configuration of the fourth agent based on the second preset prompt word template, the refinement template, and the structured input-output constraints to generate a refinement strategy; and refine and extract the task content based on the refinement strategy to obtain multiple candidate tasks.

[0016] Furthermore, the device also includes: A creation module is used to invoke a fifth intelligent agent based on preset search prompts to simulate searches for different search inputs, obtain candidate search results, and construct simulated search webpage content based on the candidate search results and search input information, thus determining the simulated webpage search environment; and to invoke a sixth intelligent agent based on preset browsing prompts to simulate browsing for different browsing inputs, obtain candidate browsing results, and construct simulated browsing webpage content based on the candidate browsing results and browsing input information, thus determining the simulated webpage browsing environment.

[0017] Furthermore, the acquisition module is also used to acquire initial task content, which is used to characterize the target generation content required for different data generation tasks; when a task status update instruction is received, the initial task content is updated based on the task content carried in the task status update instruction.

[0018] Furthermore, the processing module is also used to call the first intelligent agent to perform quality detection on the candidate task based on the detection prompt words, wherein the detection prompt words include uniqueness prompt words, pseudo-multi-hop prompt words, and leakage prompt words; if the quality detection fails, a rollback operation is performed to generate a task status update instruction.

[0019] According to another aspect of this application, a storage medium is provided, wherein at least one executable instruction is stored therein, the executable instruction causing a processor to perform operations corresponding to the data generation method based on the simulated search environment described above.

[0020] According to another aspect of this application, a device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the data generation method based on the simulated search environment described above.

[0021] By employing the above technical solutions, the technical solutions provided in the embodiments of this application have at least the following advantages: This application provides a data generation method, apparatus, storage medium, and device based on a simulated search environment. Compared with the prior art, the embodiments of this application obtain the task content of the data generation task and determine the task type based on the search complexity of the task content. The task type includes extended type and refined type. Based on the task type, the task content is processed to obtain multiple candidate tasks, and the candidate tasks are quality-checked based on a first intelligent agent. If the quality check is passed, the candidate tasks are solved multiple times in the simulated search environment based on a second intelligent agent to generate a solution trajectory, and the solution trajectory is evaluated based on a third intelligent agent. The simulated search environment includes a simulated web search environment and a simulated web browsing environment. If the evaluation result matches the preset task constraints, the target data of the data generation task is generated, realizing the purpose of data generation based on simulated search and browsing environments, reducing training costs, improving the stability and reproducibility of data training, and making the target data generated by browsing and searching simulated reduce the fluctuation of training data caused by dynamic changes in the external environment. This stabilizes the data task management training process, thereby improving the processing effectiveness of the search tool.

[0022] 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 and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0023] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This paper illustrates a flowchart of a data generation method based on a simulated search environment, according to an embodiment of this application. Figure 2This illustration shows a schematic diagram of collaborative processing of various intelligent agents according to an embodiment of this application; Figure 3 This illustration shows an interactive diagram of a simulated browsing scenario and a simulated search scenario provided in an embodiment of this application; Figure 4 This paper shows a schematic diagram of the overall structure of a data generation system provided in an embodiment of this application; Figure 5 This application provides a flowchart of a target data generation process according to an embodiment of the present application. Figure 6 This illustration shows a block diagram of a data generation device based on a simulated search environment, as provided in an embodiment of this application. Figure 7 A schematic diagram of the structure of a device provided in an embodiment of this application is shown. Detailed Implementation

[0024] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0025] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. 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 apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] The embodiments of this invention can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results. Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0027] Based on this, in one embodiment, the present invention provides a data generation method based on a simulated search environment. Taking the application of this method to computer devices such as servers as an example, the server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0028] This application provides a data generation method based on a simulated search environment, such as... Figure 1 As shown, the method includes: 101. Obtain the task content of the data generation task, and determine the task type based on the search complexity of the task content.

[0029] In this embodiment, the current execution end acts as the execution subject for generating execution data. To generate effective and accurate sample data for training, for search APIs or web browsing services, the data generation task can be triggered by the task manager, thereby obtaining the task content corresponding to the data generation task. The task content includes, but is not limited to, question content, reference answer, task level, historical operation records, historical candidate versions, intermediate entity sets, and verification result records, representing the required content for the target data to be generated. Furthermore, the task type is determined based on the search complexity of this task content. Here, search complexity is used to characterize the difficulty or complexity of searching or browsing based on the task content, thereby determining the task type. The task type includes extended types and refined types. Extended types represent task processing types that require expanding the task content, while refined types represent task processing types that require refining the task content.

[0030] In another embodiment of this application, for further definition and explanation, the step of determining the task type based on the search complexity of the task content includes: The search complexity is determined based on the task content. When the difficulty information is lower than the preset complexity requirement, or when the solution trajectory summary information does not match the search path conditions, the task type is determined to be an extended type. When the difficulty information is higher than the preset complexity requirement, and the solution trajectory summary information matches the search path conditions, the task type is determined to be a refinement type.

[0031] To expand and refine the data, thereby improving the effectiveness of target data generation, the current execution end determines the search complexity based on the task content. This search complexity includes difficulty information and solution trajectory summary information. The difficulty information is generated in report form, including but not limited to the number of solution steps, answer correctness, and difficulty level. The solution trajectory summary information includes but is not limited to the number of search calls, browsing calls, search query content, target links, and target answers, and can be obtained based on the agent's identification of the task content. This embodiment of the application does not impose specific limitations on this.

[0032] In one specific embodiment, the preset complexity requirement can be configured based on the difficulty requirements corresponding to the number of solution steps, the correctness of the answer, and the difficulty level, respectively. Simultaneously, the search path conditions can be configured based on the search requirements of the number of search calls, the number of browsing calls, the search query content, the target link, and the target answer. This embodiment does not impose specific limitations. When the difficulty information is lower than the preset complexity requirement, or when the solution trajectory summary information does not match the search path conditions, the task type is determined to be an extended type. Conversely, when the difficulty information is higher than the preset complexity requirement, and the solution trajectory summary information matches the search path conditions, the task type is determined to be a refined type.

[0033] In another embodiment of this application, to further define and illustrate, before the step of obtaining the task content of the data generation task, the method further includes: Get the initialization task content; When a task status update instruction is received, the initial task content is updated based on the task content carried in the task status update instruction.

[0034] To achieve diversified management of data tasks, thereby maintaining task content and improving the execution efficiency of data tasks, the current execution end pre-acquires initial task content. The initial task content is used to characterize the target generation content required for different data generation tasks, which may include, but is not limited to, problem content that requires external information retrieval, web browsing, and evidence location to solve. In this case, the initial task content may come from open-domain question-answering datasets, manually constructed samples, or existing task libraries. This application embodiment does not make specific limitations.

[0035] In a specific embodiment, the current execution end can maintain the task content through the task manager, that is, maintain the initialized or completed task content, historical operation records, candidate intermediate entities, task levels and evaluation results, etc., where intermediate entities refer to the object information units identified in the process of generating, expanding, solving or verifying candidate tasks, which are used to connect questions and answers. Object information units include people, places, organizations, works, events, institutions, time markers, proper nouns or combinations thereof, which are not specifically limited in this embodiment.

[0036] In one specific embodiment, within the task manager, regardless of whether it's initial task content or completed task content, the task manager can also manage task content through received task status update instructions. That is, when a task status update instruction is received, the initial task content is updated based on the task content carried in the task status update instruction. In this case, the carried task content can be entered along with the instruction, or it can be generated from a webpage or browser; this embodiment does not impose specific limitations.

[0037] In one specific embodiment, the task manager can also set seed tasks to generate task content. In this case, seed tasks refer to initial question samples, initial question-answer pairs, or basic task entries that serve as the starting point for subsequent task evolution. For example, seed tasks could be initial questions that can serve as the starting point for subsequent expansion, such as "Who is the author of a certain work?", "Which country did a certain athlete represent when winning a certain competition?", or "Which school did the founder of a certain organization graduate from?". The task manager can initialize seed tasks to form initial task content; this embodiment of the application does not impose specific limitations on this.

[0038] 102. Process the task content based on the task type to obtain multiple candidate tasks, and perform quality detection on the candidate tasks based on the first intelligent agent.

[0039] In this embodiment of the application, the current execution end processes the task content according to the task type. At this time, multiple candidate tasks can be obtained by processing according to the task type based on the intelligent agent. Then, the intelligent agent is used again to perform quality detection on the candidate tasks to ensure that the candidate tasks are quickly detected once.

[0040] It should be noted that the intelligent agent in the embodiments of this application refers to an artificial intelligence system constructed using a large language model, capable of perceiving the environment, making autonomous decisions, and executing actions to achieve a goal. The large language model is preferably an Agent large model, but this application embodiment does not impose specific limitations. Furthermore, the intelligent agent in the embodiments of this application can be the same or different pre-trained intelligent agents, such as... Figure 2 The diagram shown illustrates the collaborative processing of various intelligent agents, without specific limitations.

[0041] In another embodiment of this application, for further definition and explanation, the step processes the task content based on the task type to obtain multiple candidate tasks, including: When the task type is an extended type, the first preset prompt word template and the first structured input / output constraints are retrieved; Based on the preset prompt word template and the structured input and output constraints, the fourth agent is searched and configured to generate the target entity, intermediate bridging entity, and bridging relationship between the target entity and the intermediate bridging entity in the task content. Under entity constraints, the task content is expanded based on the target entity, intermediate bridging entities, and the bridging relationship to obtain multiple candidate tasks.

[0042] To expand the task content and improve the accuracy of target data generation, the current execution end first retrieves a first preset prompt word template and a first structured input / output constraint for the expanded task content. Since processing using an intelligent agent requires the use of prompt words, the first preset prompt word template is the expanded processing text containing prompt words configured for the expanded requirements; this embodiment does not specifically limit its scope. The first structured input / output constraint refers to structured constraints that meet requirements such as semantic fluency, unique answer, prohibition of directly revealing the target entity, and prohibition of containing prohibited entities. These constraints can be configured based on data processing needs; this embodiment does not specifically limit their scope.

[0043] To ensure consistency in the processed candidate tasks, entity constraints include, but are not limited to, the answer to the question in the initial task content must be equal to the target entity, the question in the initial task content must not directly contain the target entity, the question in the initial task content should identify at least one intermediate bridging entity, and the bridging relationship and bridging facts between the target entity and the intermediate bridging entity should be given. This ensures that the rewritten question forms a clear dependency chain, that is, the target entity can be located based on the bridging clues, and then the solution can be returned to the main body of the initial question.

[0044] To add multi-round search requirements to the current task and combine the processing characteristics of the intelligent agent, namely, using a large language model as the underlying semantic generation and rewriting engine, and through preset prompt word templates and structured input-output constraints, the large language model is configured as an extended intelligent agent for candidate task expansion. Preferably, the large language model can be a general-purpose large-parameter language model; however, this application embodiment does not impose specific limitations. The target entity refers to a specific entity phrase that appears word-for-word in the current question and is suitable as a subsequent expansion object. Preferably, the target entity is a person, place, event, organization, work title, award, year, or other proprietary object, and should not be the final answer itself. The extended intelligent agent is preferably a target entity with rich associations with other entities, so that new bridging clues can be discovered through subsequent searches. Intermediate bridging entities refer to entities identified in the search result summary after performing a simulated search on the target entity, which can form a factual association with the target entity and can be used to construct intermediate retrieval paths. For example, if the target entity is a competition, the intermediate bridging entity could be the venue, city, organization, or person associated with the competition. If the target entity is a person, the intermediate bridging entity could be their workplace, birthplace, award-winning organization, or related events. The role of the intermediate bridging entity is to prevent the solution process from jumping directly from the original problem to the final answer. Instead, the intermediate bridging entity must be located first, and then the search or reasoning can continue through it. The bridging relationship refers to the factual association between the target entity and the intermediate bridging entity. Preferably, the bridging relationship includes, but is not limited to, relationships such as venue, affiliated organization, author, award recipient, educational background, employment, biographical events, and file affiliation. Entity constraints refer to the limiting descriptive information introduced during the problem rewriting process to replace the original target entity. Preferably, the entity constraints originate from the problem clues, bridging facts, or bridging entity-related descriptions in the initial task content question, and are embedded into the rewritten question in the form of relative clauses, limiting phrases, event descriptions, or attribute descriptions. That is, the entity constraints are not randomly added, but are transformed from clues in the initial task content question that uniquely point to the target entity. The aforementioned entity constraints, target entity, intermediate bridging entity, and bridging relationship can be output by the large language model when generating the initial task content question, used to characterize how the bridging entity relates to the target entity.

[0045] In a specific embodiment, to ensure the effectiveness of the extension and avoid the agent directly rewriting the initial question freely, the current execution end, under entity constraints, extends the task content based on the target entity, intermediate bridging entities, and the bridging relationship to obtain multiple candidate tasks. Specifically, during the extension process, the agent first identifies the target entity, performs a simulated search for the target entity to obtain a search result summary, and then generates an extended question-and-answer format for the task content with the target entity as the answer based on the search result summary. The extended question-and-answer format is then used to rewrite the original question in the task content to generate new candidate questions, i.e., candidate tasks.

[0046] In another embodiment of this application, for further definition and explanation, the step processes the task content based on the task type to obtain multiple candidate tasks, including: When the task type is a refinement type, the second preset prompt word template, the refinement template, and the second structured input / output constraints are retrieved. The fourth agent is refined and configured based on the second preset prompt word template, the refined template, and the structured input and output constraints to generate a refined strategy. Based on the refinement strategy, the task content is refined and extracted to obtain multiple candidate tasks.

[0047] To expand the task content and improve the accuracy of target data generation, the current execution end, for refined task content, first retrieves the second preset prompt word template, the refined template, and the second structured input / output constraints. Since processing with an intelligent agent requires the use of prompt words, the second preset prompt word template is the refined text containing prompt words configured for expanded needs; this embodiment does not specifically limit its scope. The first structured input / output constraints refer to structured constraints that meet requirements such as semantic fluency, unique answers, no direct disclosure of target entities, and no inclusion of prohibited entities. These constraints can be configured based on data processing needs; this embodiment does not specifically limit their scope. The refined template refers to the user-input requirement template for refined processing. For example, the content of the refined template may include the current question, the correct answer, known facts, problem diagnosis information output by the decision-making intelligent agent, suggested repair methods output by the decision-making intelligent agent, and the number of historical refined attempts; this embodiment does not specifically limit its scope.

[0048] In one specific embodiment, the second preset prompt word template is used to set the underlying large language model as a question difficulty enhancement agent or a question refinement agent, and requires it to perform fuzzification processing on the current question without increasing the number of hops, thereby increasing the question difficulty. In this case, the large language model can be an existing general-purpose large-parameter language model to obtain the refined agent. In addition, constraints can be set on the second preset prompt word template, such as refining only one information segment of the question, ensuring that the refined question remains clear and has a unique answer, not retaining direct prompt information that makes the answer obvious, not introducing information that contradicts known facts, and not changing the correct answer, etc. The embodiments of this application do not impose specific limitations.

[0049] To adjust the expression method to weaken explicit cues, increase reasonable ambiguity, or add distractors, a refinement strategy is generated when the fourth agent is refined based on a second preset prompt word template, a refinement template, and structured input-output constraints. This refinement strategy characterizes the specific methods that can be refined, and the task content is refined and extracted based on the refinement strategy to obtain multiple candidate tasks. Specifically, the refinement strategy may include removing direct prompt information, attribute replacement (i.e., replacing the original name with a real attribute, feature, or description), and generalization (i.e., replacing the original expression with a broader but still accurate description). Attribute replacement and generalization are preferably performed based on known facts; when no known facts are available, it is preferable to only remove direct prompt information or perform generalization without fabricating new factual content.

[0050] In a specific embodiment, after invoking the simplified agent, the returned result can be structured and parsed. If the returned result contains preset fields, the refined problem, refined technology type, facts used, and reasoning are extracted. If the returned result does not meet the preset JSON format, the returned text can still be output as the refined result. Through refinement, candidate tasks can no longer rely solely on a single direct query to obtain the answer, but require further search, judgment, and browsing operations. This allows for controlled rewriting of the candidate task's expression while maintaining the answer, avoiding fabricated information, and not increasing the number of hops, thereby increasing the search difficulty and solution value of the candidate task.

[0051] It should be noted that in the embodiments of this application, when performing expansion or refinement processing on the extended and refined intelligent agents, multiple rounds of processing are required to generate multiple candidate tasks.

[0052] In another embodiment of this application, for further definition and explanation, the step of performing quality detection on the candidate task based on the first intelligent agent includes: Based on the detection prompt, the first agent is invoked to perform quality detection on the candidate task.

[0053] To ensure that candidate tasks are high-quality data and to achieve rapid quality inspection, thus avoiding sending candidate tasks with obvious defects into the subsequent solving and evaluation process, the current execution end first calls the intelligent agent to perform quality inspection according to detection prompts during quality inspection. These detection prompts include uniqueness prompts, pseudo-multi-hop prompts, and leakage prompts.

[0054] In a specific embodiment, the uniqueness prompt is used to indicate whether the question in the candidate task has a unique and unambiguous verifiable answer, including but not limited to determining whether the current question may correspond to multiple candidate answers, determining whether the question description is too broad, has insufficient constraints or conflicting descriptions, whether the output in the detection result is unique, possible alternative answers and explanations, etc. In this case, if the detection result shows that the candidate task has multiple reasonable answers, or the answer boundary is unclear, then the uniqueness detection is determined to fail.

[0055] In a specific embodiment, pseudo-multi-hop cue words are used to indicate whether there is a structural shortcut in the candidate task, that is, whether the answer can be obtained directly without actually going through intermediate bridging parsing. This includes, but is not limited to, judging whether the target entity, answer entity, or a direct equivalent expression between the two directly appears in the current question text; judging whether the current question directly reveals the bridging target in the question text through appositives, explicit definitions, direct naming, etc.; judging whether the current question still retains the intermediate dependency chain that requires locating the bridging entity first and then returning to the main question to complete the solution. In this case, the judgment is not based on whether a certain clue is widely known, but on whether the question text itself explicitly gives a shortcut. Therefore, if the intermediate description cannot be directly parsed from the question text alone, it is not considered pseudo-multi-hop and is judged to pass the quality check.

[0056] In one specific embodiment, the leakage warning word refers to determining whether a question in a candidate task directly leaks the standard answer or its obvious synonyms. This includes, but is not limited to, detecting whether the current question directly contains the standard answer string, detecting whether the current question contains obvious synonyms, substitutions, or approximate references to the standard answer, outputting the leakage location, leakage fragment, and explanation of the reason. If an answer leakage is detected, the answer leakage detection is deemed to have failed.

[0057] It should be noted that the candidate tasks obtained in this embodiment may include the initial question, the expanded or refined question, the standard answer, and the bridging context corresponding to the most recent task evolution. The bridging context preferably includes the target entity, the bridging entity, the bridging facts, and replacement description information, used to assist in determining whether the current question still maintains a reasonable multi-hop dependency structure. Furthermore, the original question, the current question, the standard answer, and the bridging context are filled into a preset quality access control prompt template to form the quality detection input. The prompt template is used to configure the first agent as a quality access control agent in a multi-hop question generation scenario, and requires it to output only structured JSON detection results.

[0058] Correspondingly, the method also includes: If the quality check fails, a rollback operation is performed to generate a task status update instruction.

[0059] In one specific embodiment, during quality inspection, the uniqueness detection results, pseudo-multi-hop detection results, and answer leakage detection results are aggregated to generate a structured quality access control result. Preferably, in addition to whether the quality access control is passed, the detection result may also include failed detection items, sub-results of each detection item, and rollback reasons. In this case, the structured result is represented in JSON format for subsequent module parsing and invocation.

[0060] It should be noted that if a candidate task passes the quality gate, it enters the solution verification stage. If a candidate task fails the quality gate, it does not enter the solution verification stage. Instead, the corresponding reason for failure is written into the rollback reason field, and the rollback control module restores the task to its previous valid state, or it is handed over to the decision-making agent module for further processing. Quality detection can use an existing large language model as the underlying detection engine, and configure the large language model as the question quality detection agent (i.e., the first agent) through preset system prompts and user input templates. Furthermore, quality detection is not performed in isolation, but forms a closed-loop constraint with the preceding expansion and refinement processes. The expansion process requires that the rewritten question has a unique and unambiguous answer, and avoids pseudo-multi-hops and answer near-synonyms. The refinement process also requires that the refined question maintain a unique answer and not introduce unsolvable ambiguity. Therefore, quality detection further independently verifies these constraints, thereby improving the stability of candidate task quality.

[0061] 103. When the quality test is passed, the candidate task is solved in multiple rounds in a simulated search environment based on the second intelligent agent to generate a solution trajectory, and the solution trajectory is evaluated based on the third intelligent agent.

[0062] In this embodiment, after the agent performs quality detection on the processed candidate tasks, the second agent performs multiple rounds of solving the candidate tasks in a simulated search environment to generate a solution trajectory. This means that the agent performs multiple rounds of web page search and web page browsing on the candidate tasks in the simulated search environment to obtain data from the search and browsing process, which serves as the solution trajectory. The simulated search environment includes a simulated web page search environment and a simulated web page browsing environment. Both are pre-generated by the agent under the constraint of boosting words, serving as the simulated environment for the data solving process.

[0063] It should be noted that the obtained solution trajectory can be used as the initial data result of the data generation task. In order to ensure the final validity of the data, the current execution end evaluates the obtained solution trajectory based on the agent to determine whether to roll back the data or process the central execution task content.

[0064] In a specific embodiment, when a candidate task is solved multiple times in a simulated search environment based on a second agent to generate a solution trajectory, the second agent, acting as the solving agent, simultaneously invokes a simulated web search environment and a simulated web browsing environment to perform multiple rounds of solving the candidate task. Each round of the agent's solving process includes: 1. Generating a search query based on the current task; 2. Calling the simulated web search module to obtain candidate search results; 3. Selecting a target link; 4. Calling the simulated web browsing module to obtain web page content; 5. Continuing to initiate subsequent searches or browsing based on the web page content and intermediate results; 6. Until a final answer is formed, which serves as the solution trajectory. Here, the solution trajectory characterizes the actual solving process of the candidate task in the simulated search environment and is an important basis for subsequent task evaluation. The solution trajectory may include, but is not limited to, each round of search query, search result list, selected link, browsing returned content, intermediate entities, intermediate inference results, and the final answer.

[0065] In another embodiment of this application, for further definition and explanation, before the step of generating a solution trajectory by having the second agent solve the candidate task in a simulated search environment for multiple rounds, the method further includes: Based on preset search prompts, a fifth agent is invoked to perform simulated searches for different search inputs, resulting in candidate search results. Simulated search webpage content is then constructed based on the candidate search results and search input information, and this is determined as the simulated webpage search environment. Based on preset browsing prompts, a sixth agent is invoked to simulate browsing different browsing inputs, resulting in candidate browsing results. Based on the candidate browsing results and browsing input information, simulated webpage content is constructed and determined as the simulated webpage browsing environment.

[0066] In order to improve the accuracy of data generation by performing data processing in simulated web search and web browsing environments, the current execution end calls the corresponding intelligent agents to simulate based on preset search and browsing prompts, so as to pre-build simulated web search and web browsing environments.

[0067] In one specific embodiment, when creating a simulated search environment, different candidate tasks are first received, and a fifth agent is invoked based on preset search prompts to generate candidate search results. The candidate search results are then parsed into structured search results containing titles, link addresses, and summary information. Subsequently, when invoking a simulated webpage browsing agent, the current execution end matches the corresponding summary information in historical search results based on the target link address and extracts the corresponding query statement from the historical search call record. The target link address, summary information, and historical query statement are combined according to preset rules to form browsing input. Then, based on the browsing input, the agent is invoked to generate the webpage title and main content corresponding to the target link, thereby obtaining simulated webpage content that is consistent with the previous search context, which serves as the simulated webpage search environment.

[0068] Specifically, in a scenario utilizing a simulated search environment, candidate tasks are used as search queries. These queries are then populated into a pre-defined search suggestion template to form the search input that drives the agent to generate search results. The search suggestions constrain the agent to output results as a search engine and limit the output to a pre-defined structured format. The search input is sent to a large language model, which generates one or more candidate search result entries semantically related to the search query. The agent's returned content is then parsed into structured data, with each search result entry including at least a title, link address, and summary information. Furthermore, the parsed search results undergo format validity and content integrity checks. Results that do not meet the pre-defined data format requirements can be retried, have fields completed, or be removed. Preferably, the search results can also be deduplicated, filtered for relevance, pruned, and filtered for abnormal links.

[0069] Finally, structured search results conforming to preset format requirements are output for subsequent use by the simulated web browsing module or the intelligent agent solving module. Preferably, the structured search results adopt a unified data format, for example including the following structure: { "results": [ { "title": "..." "url": "https: / / ...", "snippet": "..."} ] }

[0070] The title represents the topic of the search results, the link address uniquely identifies the candidate web page resource, and the summary information summarizes the web page content related to the query.

[0071] In addition, the preset search suggestions may include the following statements: "You are a simulated web browser and content crawler"; "Given a URL (and some optional search context, such as the original query and summary fragments), you need to simulate a visit to that page and return its title and main body content." "You must output only one JSON object and not any other content."

[0072] The output JSON format is as follows: { "url": "<the same URL you received>" "title": "<page title>" "content": "<Main content of the page>" }

[0073] Here, "content" instructs the agent to process a coherent, complete text similar to the main body of an article. "max_chars" instructs the agent to impose a length limit, truncating "content" to a maximum of this number of characters. If truncation occurs, "...[TRUNCATED]" should be appended to the end.

[0074] In one specific embodiment, the same creation method as the simulated search environment can be used to create the simulated browsing environment, and the simulated browsing scenario can interact with the simulated search scenario, such as... Figure 3As shown, this is used to generate corresponding webpage content based on the target link and related context, simulating the process of opening a webpage and obtaining its main content. First, different browsing inputs are received. These inputs include the target link address, and preferably also include historical query statements, summary information corresponding to the target link, and the current task context. The target link address, historical query statements, summary information, and current task context are then filled into a preset browsing prompt template to form the browsing input driving the large language model to generate webpage content. The browsing prompt template constrains the agent to output structured content of the target webpage as a web browser or content crawler. The browsing input is sent to the agent, which generates candidate webpage content based on the webpage topic, query semantics, and summary clues corresponding to the target link. The candidate webpage content includes: link address, webpage title, and main page content. Subsequently, the content returned by the agent is parsed into a preset structured format, and the main page content is cleaned, spliced, truncated, or normalized. The main page content is continuous natural language text used to represent the main content of the webpage. When the main content of the page exceeds a preset length threshold, the excess portion is truncated, and a preset truncation marker is appended at the truncation location. The truncation marker is "...[TRUNCATED]" or other identifiable truncation description. Finally, the structured webpage content is output for the solving agent module to read, reason, and decide on subsequent search or browsing operations. Preferably, the simulated browsing results adopt the following structured format: { "url": "..." "title": "..." "content": "..." }

[0075] In this context, url identifies the currently viewed webpage, title represents the webpage's theme, and content represents the page content generated based on the target link and context.

[0076] In addition, the preset browsing prompts may include the following statements: "You are a browser, a web search engine"; "Given a user's query, you need to return a list of search results"; "You must output only one JSON object and not any other content"; The JSON format is as follows: { "results": [ { "title": "..." "url": "https: / / ...", "snippet": "..." } ] }

[0077] In one specific embodiment, to improve the consistency between the results generated by the simulated web browsing module and the preceding search behavior, context completion is performed on the browsing input when the simulated web browsing module is invoked to construct a context-complete structure for the browsing input. Specifically, a browsing request for a target link is received. The browsing request includes at least the target link address, and preferably, also includes the current task question or the current browsing target description. Search results entries corresponding to the target link address are retrieved from the historical search results record. If a matching entry is found, the summary information corresponding to that entry is extracted. Furthermore, the query statement used to generate the search results corresponding to the target link is found in the historical search call record, and this query statement is extracted as part of the browsing context. Finally, the target link address, historical query statement, corresponding summary information, and optional current question description are combined according to a preset template to form the browsing module input. The browsing module input includes at least the target link address, query statement, summary information, and current question or task context. The browsing module input is sent to the simulated web browsing module so that the simulated web browsing module generates web page content consistent with the preceding search behavior based on the link information and search context. The matching methods include at least one of the following: exact matching by link address, matching after standardization by link address, and combined matching by title and link in search results entries. Additionally, historical query statements can be extracted by directly reading search request parameters from historical tool call records, extracting query fields from historical dialogue records, call logs, or intermediate messages, or extracting query statements through regular expression matching, field parsing, or structured log parsing. This application's embodiments do not impose specific limitations on these methods.

[0078] In this embodiment, when no historical summary information or historical query statement corresponding to the target link is found, browsing input can be constructed solely based on the target link address and the current task context. Compared to the scheme of generating webpage content based solely on the target link address, this embodiment further introduces historical query statements and corresponding summary information during the simulated webpage browsing stage. This allows the browsing agent to inherit the previous search context when generating webpage content, thereby improving the consistency, coherence, and verifiability between the simulated webpage content and the search results. Simultaneously, by constraining the output JSON structure and implementing truncation control on the text length, the stability of the simulation environment during data generation and subsequent solution verification processes is improved.

[0079] 104. If the evaluation result matches the preset task constraints, generate the target data for the data generation task.

[0080] In this embodiment of the application, after the evaluation result is obtained, it is matched with the preset task constraints. At this time, the preset task constraints may include the number of solution trajectories that pass the evaluation, the existence of specific words (such as location, name, URL), etc., as constraints for the final output target data. This embodiment of the application does not make specific limitations.

[0081] In one specific embodiment, when the decision-making agent determines that the current candidate task satisfies preset task constraints, it outputs the target data. The preset task constraints may include one or more of the following: data identifier, question, answer, intermediate entity set, search path, browsing path, evaluation result, task level, and historical construction information.

[0082] In a specific embodiment, when evaluating the solved trajectory, a large language model is used as the underlying evaluation engine. This large language model can be configured as the judging agent (i.e., the third agent) in a multi-hop problem generation scenario by using preset evaluation prompts, user input templates, and structured output constraints. The large language model can be a general-purpose large-parameter language model; this embodiment does not impose specific limitations. The specific evaluation process includes: 1. Obtain evaluation input, including but not limited to the initial problem, candidate problems, standard answer, and the complete solution trajectory corresponding to the solving agent. Preferably, the solution trajectory includes assistant messages from each round of the solution process, tool call content, tool return results, and the final answer.

[0083] 2. Trajectory Formatting: The solved trajectory is formatted to form trajectory text that can be read by the evaluation model. Preferably, the formatting process includes at least removing system messages, retaining the assistant's search, browsing, and answer output content, summarizing or truncating the tool's returned results, and extracting the final answer from the trajectory. Through the above processing, key information such as the search path, browsing path, and answer output can be retained while controlling the input length.

[0084] 3. Construct evaluation prompt input by filling the initial question, expanded or refined question, standard answer, formatted solution path, and final answer into a preset evaluation prompt template to form evaluation input. Preferably, the evaluation prompt words should at least require the underlying large language model to judge the correctness of the answer, the difficulty of the task, and whether there is a shortcut based on the solution path, and provide suggestions for the next step.

[0085] 4. Invoke the large language model to perform the evaluation. Send the evaluation input to the agent, and the large language model will output a structured evaluation result based on the candidate task and the solution trajectory. Preferably, the structured evaluation result includes at least the following evaluation results: whether the answer is correct, the total number of steps, the difficulty level, whether a shortcut exists, suggested actions, explanations of reasons, and suggestions for improvement.

[0086] 5. Standardize the evaluation results, which involves standardizing the evaluation results returned by the large language model. Specific steps may include: (1) Verify the total number of steps. If the model does not return a valid number of steps, calculate the number of search calls and browsing calls in the solved trajectory. (2) Standardize the difficulty level field and limit it to one of "too easy", "appropriate", "too difficult" or "invalid"; (3) Standardize the suggested action field and limit it to one of the following: expansion, refinement, rollback or termination; (4) Standardize the format of the suggestion list.

[0087] At this point, the evaluation content may include, but is not limited to, whether the final answer is correct, whether multiple rounds of searching and / or browsing actually occurred, whether a coherent search and browsing path was formed, whether a structural shortcut exists, whether the current task difficulty is suitable for outputting the target data, and whether the next step should be to continue expanding, refining, rolling back, or terminating the output. The difficulty level is determined based on at least whether the solution is correct and the number of steps. Preferably, when the solution is correct and the number of steps is small, it is judged as "too easy"; when the solution is correct and the number of steps reaches a preset threshold, it is judged as "appropriate"; when the number of steps is large but the answer is still incorrect, it is judged as "too difficult"; when the problem has logical or factual problems that make it unsolvable, it is judged as "failed". Correspondingly, the evaluation agent outputs a final evaluation report, which may include the difficulty level, the number of search and browsing steps, whether a shortcut exists, the type of defect, a description of the path quality, and suggestions for the next step. Based on the evaluation report, the decision-making agent can then determine whether to continue expanding, refining, rolling back, or terminating the output.

[0088] In one specific embodiment, rollback refers to restoring the current task content to the previous valid content when a candidate task fails the quality check, or when the evaluation agent determines that the current candidate task has problems such as being unsolvable, having logical errors, pseudo-multi-hops, or insufficient path quality. During the rollback process, the reason for the failure, the corresponding error type, and the evolution path of the failure can also be recorded to avoid repeatedly generating the same type of low-quality candidate task in the future.

[0089] It should be noted that the obtained target data can be applied to different training scenarios, such as search-type agent training scenarios, reinforcement learning scenarios based on simulated web page environments, and search task verification and evaluation scenarios.

[0090] In a specific search-based agent training scenario, the target data can serve as training data for a search-based agent, a multi-turn retrieval agent, or a deep search agent. This data is used to train the model to perform multi-turn searches, link selection, web browsing, evidence integration, and answer generation in an environment comprised of a simulated web search module and a simulated web browsing module. Since the data samples contain questions, answers, intermediate entity sets, search paths, browsing paths, and evaluation results, they can be used not only for result-supervised training but also for process-supervised training.

[0091] In a specific reinforcement learning scenario based on a simulated webpage environment, target data can be used to train an agent for reinforcement learning in simulated webpage search and browsing environments. Specifically, the agent can initiate search and browsing operations around the problem samples and obtain reward signals based on the correctness of the final answer, the number of search steps, path quality, intermediate entity coverage, repeated search cases, or invalid browsing cases, thereby learning better search and browsing strategies.

[0092] In a specific search task verification and evaluation scenario, the target data can be used for offline verification and capability evaluation of the search agent. Since the data sample includes standard answers, search paths, browsing paths, and evaluation reports, it can be used to test whether the model under evaluation can complete multi-step solutions in a simulated webpage environment, whether it can form a reasonable path, and whether it possesses intermediate bridging search capabilities.

[0093] In a specific scenario of an embodiment of this application, such as Figure 4 The diagram shown illustrates the overall structure of the data generation system. The data generation steps for each virtual module include: Step S1: Receive the seed task and initialize the task state; Step S2: The decision-making agent module determines the next operation based on the current task status; Step S3: If the next operation is expansion, then call the expansion agent module to bridge and expand the task; if the next operation is refinement, then call the refinement agent module to refine the task description. Step S4: Input the candidate tasks into the quality access control module for screening; Step S5: For candidate tasks that pass the screening, the solving agent module performs multiple rounds of search and browsing in the simulated search environment to obtain the solution trajectory; Step S6: The evaluation agent module evaluates the candidate tasks based on the solution trajectory; Step S7: Based on the evaluation results, the decision-making agent module decides whether to continue expanding, continue refining, execute a rollback, or terminate the output; Step S8: When the candidate task meets the output conditions, the data output module outputs the final data sample.

[0094] Steps S2 to S7 can be executed repeatedly until a data sample that meets the requirements is generated, or until a preset iteration termination condition is reached, such as... Figure 5 As shown.

[0095] This application provides a data generation method based on a simulated search environment. Compared with the prior art, this application obtains the task content of the data generation task and determines the task type based on the search complexity of the task content. The task type includes extended type and refined type. Based on the task type, the task content is processed to obtain multiple candidate tasks, and the candidate tasks are quality-checked based on a first intelligent agent. If the quality check is passed, the candidate tasks are solved multiple times in a simulated search environment based on a second intelligent agent to generate a solution trajectory. The solution trajectory is evaluated based on a third intelligent agent. The simulated search environment includes a simulated web search environment and a simulated web browsing environment. If the evaluation result matches the preset task constraints, the target data of the data generation task is generated. This achieves the purpose of data generation based on simulated search and browsing environments, reduces training costs, and improves the stability and reproducibility of data training. The target data generated through browsing and search simulation reduces the fluctuation of training data caused by dynamic changes in the external environment, stabilizes the data task management training process, and thus improves the processing effectiveness of the search tool.

[0096] Furthermore, as a response to the above Figure 1 To implement the method shown, this application provides a data generation device based on a simulated search environment, such as... Figure 6 As shown, the device includes: The acquisition module 21 is used to acquire the task content of the data generation task and determine the task type based on the search complexity of the task content. The task type includes extended type and refined type. The processing module 22 is used to process the task content based on the task type to obtain multiple candidate tasks, and to perform quality detection on the candidate tasks based on the first intelligent agent. The evaluation module 23 is used to, when the quality detection is passed, perform multiple rounds of solving the candidate task in a simulated search environment based on the second agent to generate a solution trajectory, and evaluate the solution trajectory based on the third agent. The simulated search environment includes a simulated web search environment and a simulated web browsing environment. The generation module 24 is used to generate the target data of the data generation task if the evaluation result matches the preset task constraints.

[0097] Furthermore, the acquisition module is specifically used to determine the search complexity based on the task content, the search complexity including difficulty information and solution trajectory summary information; when the difficulty information is lower than the preset complexity requirement, or the solution trajectory summary information does not match the search path conditions, the task type is determined to be an extended type; when the difficulty information is higher than the preset complexity requirement, and the solution trajectory summary information matches the search path conditions, the task type is determined to be a refined type.

[0098] Furthermore, the processing module is specifically configured to, when the task type is an extended type, retrieve a first preset prompt word template and a first structured input / output constraint; perform search configuration on the fourth agent based on the preset prompt word template and the structured input / output constraint, generate the target entity, intermediate bridging entity, and bridging relationship between the target entity and the intermediate bridging entity of the task content; and, under entity constraint conditions, extend the task content based on the target entity, intermediate bridging entity, and the bridging relationship to obtain multiple candidate tasks.

[0099] Furthermore, the processing module is specifically used to, when the task type is a refinement type, retrieve the second preset prompt word template, the refinement template, and the second structured input-output constraints; refine the configuration of the fourth agent based on the second preset prompt word template, the refinement template, and the structured input-output constraints to generate a refinement strategy; and refine and extract the task content based on the refinement strategy to obtain multiple candidate tasks.

[0100] Furthermore, the device also includes: A creation module is used to invoke a fifth intelligent agent based on preset search prompts to simulate searches for different search inputs, obtain candidate search results, and construct simulated search webpage content based on the candidate search results and search input information, thus determining the simulated webpage search environment; and to invoke a sixth intelligent agent based on preset browsing prompts to simulate browsing for different browsing inputs, obtain candidate browsing results, and construct simulated browsing webpage content based on the candidate browsing results and browsing input information, thus determining the simulated webpage browsing environment.

[0101] Furthermore, the acquisition module is also used to acquire initial task content, which is used to characterize the target generation content required for different data generation tasks; when a task status update instruction is received, the initial task content is updated based on the task content carried in the task status update instruction.

[0102] Furthermore, the processing module is also used to call the first intelligent agent to perform quality detection on the candidate task based on the detection prompt words, wherein the detection prompt words include uniqueness prompt words, pseudo-multi-hop prompt words, and leakage prompt words; if the quality detection fails, a rollback operation is performed to generate a task status update instruction.

[0103] This application provides a data generation device based on a simulated search environment. Compared with the prior art, this application obtains the task content of the data generation task and determines the task type based on the search complexity of the task content. The task type includes extended type and refined type. Based on the task type, the task content is processed to obtain multiple candidate tasks, and the candidate tasks are quality-checked based on a first intelligent agent. If the quality check is passed, the candidate tasks are solved multiple times in the simulated search environment based on a second intelligent agent to generate a solution trajectory. The solution trajectory is evaluated based on a third intelligent agent. The simulated search environment includes a simulated web search environment and a simulated web browsing environment. If the evaluation result matches the preset task constraints, the target data of the data generation task is generated. This achieves the purpose of data generation based on simulated search and browsing environments, reduces training costs, and improves the stability and reproducibility of data training. The target data generated through browsing and search simulation reduces the fluctuation of training data caused by dynamic changes in the external environment, stabilizes the data task management training process, and thus improves the processing effectiveness of the search tool.

[0104] According to one embodiment of this application, a storage medium is provided, the storage medium storing at least one executable instruction that can execute the data generation method based on a simulated search environment in any of the above method embodiments.

[0105] Figure 7 The diagram shows a structural schematic of a device according to one embodiment of the present application. The specific embodiments of the present application do not limit the specific implementation of the device.

[0106] like Figure 7 As shown, the device may include: a processor 302, a communications interface 304, a memory 306, and a communications bus 308.

[0107] The processor 302, communication interface 304, and memory 306 communicate with each other via communication bus 308.

[0108] Communication interface 304 is used to communicate with other network elements such as clients or other servers.

[0109] The processor 302 is used to execute program 310, specifically to execute the relevant steps in the above-described embodiment of the data generation method based on a simulated search environment.

[0110] Specifically, program 310 may include program code that includes computer operation instructions.

[0111] Processor 302 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or they may be processors of different types, such as one or more CPUs and one or more ASICs.

[0112] Memory 306 is used to store program 310. Memory 306 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0113] Specifically, program 310 can be used to cause processor 302 to perform the following operations: The task content of the data generation task is obtained, and the task type is determined based on the search complexity of the task content. The task type includes extended type and refined type. The task content is processed based on the task type to obtain multiple candidate tasks, and the candidate tasks are quality checked based on the first intelligent agent. If the quality check is passed, the candidate task is solved in multiple rounds in a simulated search environment based on the second agent to generate a solution trajectory, and the solution trajectory is evaluated based on the third agent. The simulated search environment includes a simulated web search environment and a simulated web browsing environment. If the evaluation result matches the preset task constraints, the target data for the data generation task is generated.

[0114] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0115] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A data generation method based on a simulated search environment, characterized in that, include: The task content of the data generation task is obtained, and the task type is determined based on the search complexity of the task content. The task type includes extended type and refined type. The task content is processed based on the task type to obtain multiple candidate tasks, and the candidate tasks are quality checked based on the first intelligent agent. If the quality check is passed, the candidate task is solved in multiple rounds in a simulated search environment based on the second agent to generate a solution trajectory, and the solution trajectory is evaluated based on the third agent. The simulated search environment includes a simulated web search environment and a simulated web browsing environment. If the evaluation result matches the preset task constraints, the target data for the data generation task is generated. The process of processing the task content based on the task type to obtain multiple candidate tasks includes: When the task type is an extended type, the first preset prompt word template and the first structured input / output constraints are retrieved; Based on the preset prompt word template and the first structured input-output constraints, the fourth agent is searched and configured to generate the target entity, intermediate bridging entity, and bridging relationship between the target entity and the intermediate bridging entity in the task content. Under entity constraints, the task content is expanded based on the target entity, intermediate bridging entities, and the bridging relationship to obtain multiple candidate tasks.

2. The method according to claim 1, characterized in that, The process of determining the task type based on the search complexity of the task content includes: The search complexity is determined based on the task content, and the search complexity includes difficulty information and solution trajectory summary information; When the difficulty information is lower than the preset complexity requirement, or when the solution trajectory summary information does not match the search path conditions, the task type is determined to be an extended type. When the difficulty information is higher than the preset complexity requirement, and the solution trajectory summary information matches the search path conditions, the task type is determined to be a refinement type.

3. The method according to claim 2, characterized in that, The process of processing the task content based on the task type to obtain multiple candidate tasks includes: When the task type is a refinement type, the second preset prompt word template, the refinement template, and the second structured input / output constraints are retrieved. Based on the second preset prompt word template, the refined template, and the second structured input-output constraints, the fourth agent is refined and configured to generate a refinement strategy. The refinement strategy is used to characterize the specific methods that can be refined. The refinement strategy includes removing direct prompt information, attribute replacement, and generalization processing. Based on the refinement strategy, the task content is refined and extracted to obtain multiple candidate tasks.

4. The method according to claim 1, characterized in that, Before generating the solution trajectory by performing multiple rounds of solving the candidate task in a simulated search environment based on the second intelligent agent, the method further includes: Based on preset search prompts, a fifth agent is invoked to perform simulated searches for different search inputs, resulting in candidate search results. Simulated search webpage content is then constructed based on the candidate search results and search input information, and this is determined as the simulated webpage search environment. Based on preset browsing prompts, a sixth agent is invoked to simulate browsing different browsing inputs, resulting in candidate browsing results. Based on the candidate browsing results and browsing input information, simulated webpage content is constructed and determined as the simulated webpage browsing environment.

5. The method according to claim 1, characterized in that, Before acquiring the data to generate the task content, the method further includes: Obtain the initialization task content, which is used to characterize the target generation content required for different data generation tasks; When a task status update instruction is received, the initial task content is updated based on the task content carried in the task status update instruction.

6. The method according to claim 1, characterized in that, The quality detection of the candidate tasks based on the first intelligent agent includes: Based on the detection prompts, the first agent is invoked to perform quality detection on the candidate tasks. The detection prompts include uniqueness prompts, pseudo-multi-hop prompts, and leakage prompts. The method further includes: If the quality check fails, a rollback operation is performed to generate a task status update instruction.

7. A data generation device based on a simulated search environment, characterized in that, include: The acquisition module is used to acquire the task content of the data generation task and determine the task type based on the search complexity of the task content. The task type includes extended type and refined type. The processing module is used to process the task content based on the task type to obtain multiple candidate tasks, and to perform quality detection on the candidate tasks based on the first intelligent agent. An evaluation module is used to, when a quality check is passed, perform multiple rounds of solving the candidate task in a simulated search environment based on a second agent to generate a solution trajectory, and evaluate the solution trajectory based on a third agent. The simulated search environment includes a simulated web search environment and a simulated web browsing environment. The generation module is used to generate the target data for the data generation task if the evaluation result matches the preset task constraints. The processing module is specifically used to retrieve the first preset prompt word template and the first structured input-output constraint when the task type is an extended type; to search and configure the fourth agent based on the preset prompt word template and the first structured input-output constraint, and to generate the target entity, intermediate bridging entity and the bridging relationship between the target entity and the intermediate bridging entity of the task content. Under entity constraints, the task content is expanded based on the target entity, intermediate bridging entities, and the bridging relationship to obtain multiple candidate tasks.

8. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

Images