Report generation method based on intelligent agent

Abstract

The invention discloses an agent-based report generation method, and relates to the technical field of artificial intelligence. The method comprises the following steps: generating a target execution plan according to a to-be-processed query; determining a specified sub-task based on the dependent task, and issuing the specified sub-task to a task bus; a specified result object is obtained on the task bus, and dimension scores of the specified result object in multiple dimensions are calculated in combination with the result content, the metadata and the object associated data; generating a report outline by utilizing the dimension score and all result objects in the object library, and issuing a draft task; and performing integration processing on the task first draft according to the report outline to obtain a query report corresponding to the to-be-processed query. According to the method, the task bus is used as message middleware, asynchronous parallel processing can be realized, and the report generation efficiency is improved; object association data is introduced to establish a relationship between different result objects, deep analysis and logical reasoning can be performed, and the quality of a report can be effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a report generation method based on an agent. BACKGROUND

[0002] A report is a text formed after in-depth analysis of the content under a specific topic, which has high professionalism and rigor. In existing report writing, professional personnel need to read a large amount of data and professionally analyze and organize a large amount of data.

[0003] With the rapid development of artificial intelligence technology, especially in the field of text generation, relying on the powerful text analysis and processing capabilities of the model can realize automatic collection of information, analysis of information and integration to generate a report. However, in the current report generation scheme based on the model, the mode of assigning tasks, waiting, and summarizing is usually adopted, which is low in efficiency, and only a variety of information is summarized when generating a report, resulting in poor quality of the report. SUMMARY

[0004] In view of the above problems, the purpose of the present application is to provide a report generation method based on an agent, which can improve the efficiency of report generation and improve the quality of the report.

[0005] To solve the above technical problems, the present application provides the following technical solutions:

[0006] A report generation method based on an agent, applied to an agent, the method comprising:

[0007] Generating a target execution plan according to a to-be-processed query, the target execution plan comprising a plurality of sub-tasks and a dependent task corresponding to each sub-task;

[0008] Determining a specified sub-task based on the dependent task, and publishing the specified sub-task to a task bus, so that a working agent obtains the specified sub-task and executes it, and returns a specified result object, the specified result object comprising result content, metadata, and object association data;

[0009] Obtaining the specified result object from the task bus, and determining a new specified sub-task based on the task state of the specified sub-task to publish;

[0010] For each specified result object, calculating a dimension score of the specified result object in multiple dimensions in combination with the result content, metadata, and object association data;

[0011] Generating a report outline and a corresponding draft task using the dimension score and all result objects in the object library, and publishing the draft task to the task bus so that the working agent returns a task draft after executing the draft task;

[0012] The task draft is integrated according to the report outline, and a query report corresponding to the to-be-processed query is obtained.

[0013] A report generation method based on an agent, applied to a working agent, the method comprising:

[0014] A specified subtask is obtained from a task bus and executed, and a task execution result is obtained, the specified subtask being determined by an orchestration agent from a target execution plan, the target execution plan being generated based on a to-be-processed query;

[0015] The task execution result is subjected to content integration processing, and a result content is obtained.

[0016] The task execution result is subjected to quality evaluation and content extraction processing by a large language model, and metadata is generated.

[0017] The logical relationship between the task execution result and other result objects is classified, and object association data of the task execution result is obtained.

[0018] The result content, the metadata, and the object association data are merged into a specified result object, and the specified result object is published to the task bus, so that the orchestration agent generates a report corresponding to the to-be-processed query based on the specified result object on the task bus.

[0019] A report generation device based on an agent, applied to an orchestration agent, the device comprising:

[0020] A plan generation module for generating a target execution plan according to a to-be-processed query, the target execution plan comprising a plurality of subtasks and a dependent task corresponding to each subtask;

[0021] A determination module for determining a specified subtask based on a dependent task and publishing the specified subtask to a task bus, so that a working agent obtains and executes the specified subtask and returns a specified result object, the specified result object comprising a result content, metadata, and object association data;

[0022] A task publishing module for obtaining the specified result object from the task bus and determining a new specified subtask based on the task state of the specified subtask for publishing;

[0023] A computing module for calculating, for each specified result object, a dimension score of the specified result object in multiple dimensions in combination with the result content, the metadata, and the object association data.

[0024] The outline generation module is configured to generate a report outline and corresponding drafting tasks by using the dimension scores and all result objects in the object library, and publish the drafting tasks to a task bus so that the working agent returns a task draft after executing the drafting tasks;

[0025] The report generation module is configured to integrate the task draft according to the report outline to obtain a query report corresponding to the to-be-processed query.

[0026] A report generation device based on an agent is applied to a working agent, and the device comprises:

[0027] The acquisition execution module is configured to acquire and execute specified sub-tasks from the task bus to obtain task execution results, wherein the specified sub-tasks are determined by the orchestration agent from a target execution plan, and the target execution plan is generated based on a to-be-processed query.

[0028] The integration module is configured to perform content integration processing on the task execution results to obtain result content.

[0029] The extraction module is configured to perform quality evaluation and content extraction processing on the task execution results by a large language model to generate metadata.

[0030] The classification module is configured to perform classification processing on a logical relationship between the task execution results and other result objects to obtain object association data of the task execution results.

[0031] The object publishing module is configured to merge the result content, the metadata and the object association data into a specified result object, and publish the specified result object to the task bus so that the orchestration agent generates a report corresponding to the to-be-processed query based on the specified result object on the task bus.

[0032] An electronic device comprises a processor and a memory, and the memory stores a plurality of instructions; the processor loads the instructions from the memory to execute the steps in any of the report generation methods based on an agent provided by the present application.

[0033] A computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor to execute the steps in any of the report generation methods based on an agent provided by the present application.

[0034] A computer program product comprises a computer program / instruction, and the computer program / instruction is executed by a processor to implement the steps in any of the report generation methods based on an agent provided by the present application.

[0035] The technical solutions provided by the present application have at least the following beneficial effects:

[0036] In the embodiment of the present application, a target execution plan is generated according to a to-be-processed query, and a specified subtask is determined from the target execution plan and issued to a task bus, so that a working intelligent agent can obtain an execution subtask implementation from the task bus to implement parallel execution, and a specified result object obtained after execution is transmitted to a scheduling intelligent agent through the task bus; after the scheduling intelligent agent obtains the specified result object, a dimension score of the specified result object in multiple dimensions is calculated, and then a report outline is generated by using the dimension score and a result object in an object library, and a draft task is created and transmitted to the working intelligent agent through the task bus for execution, and a task draft is obtained to generate a query report. By taking the task bus as a message middleware, asynchronous parallel processing can be implemented, and the generation efficiency of the report is improved; the object association data is introduced to establish a connection between different result objects, in-depth analysis and logical reasoning can be performed, and thus the quality of the report can be effectively improved. BRIEF DESCRIPTION OF DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0038] Figure 1 is an application scenario diagram of the report generation method based on intelligent agents provided by the embodiment of the present application;

[0039] Figure 2 is a flow diagram of the report generation method based on intelligent agents provided by the embodiment of the present application;

[0040] Figure 3 is a diagram for generating a report outline based on all result objects provided by the embodiment of the present application;

[0041] Figure 4 is a flow diagram of the report generation method based on intelligent agents provided by another embodiment of the present application;

[0042] Figure 5 is a flow diagram of the report generation method based on intelligent agents provided by another embodiment of the present application;

[0043] Figure 6 is an interaction diagram of the report generation method based on intelligent agents provided by the embodiment of the present application;

[0044] Figure 7 is a structural diagram of the report generation device based on intelligent agents provided by the embodiment of the present application;

[0045] Figure 8is a structural schematic diagram of an agent-based report generation device provided by another embodiment of the present application.

[0046] Figure 9 is a structural schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0048] It can be understood that, in the specific embodiments of the present application, the data related to user information needs to be obtained with the permission or consent of the user, and the collection, use and processing of the related data need to comply with the relevant laws, regulations and standards of the country and region.

[0049] For reference Figure 1 , a schematic diagram of an application scenario of an agent-based report generation method is shown. The application scenario can include a terminal 101 and a server 102, and the terminal 101 and the server 102 can exchange data through a network. The terminal 101 can be a mobile phone, a tablet computer, a smart Bluetooth device, a computer, a large screen device, a robot, etc. The server 102 can be a single server or a server cluster composed of multiple servers.

[0050] The user can send a to-be-processed query to the server 102 through the terminal 101. The report generation system can be deployed on the server 102, and the system can include an orchestration agent.

[0051] The orchestration agent in the server 102 can generate a target execution plan according to the to-be-processed query, the target execution plan including a plurality of sub-tasks and a corresponding dependent task of each sub-task; determine a specified sub-task based on the dependent task, and publish the specified sub-task to the task bus, so that the working agent obtains the specified sub-task and executes it, and then returns a specified result object, wherein the specified result object can include result content, metadata, and object association data. The orchestration agent obtains the specified result object from the task bus, and determines a new specified sub-task based on the state of the specified sub-task to publish; for each specified result object, the result content, metadata, and object association data are combined to calculate the dimension score of the specified result object in multiple dimensions; using the dimension score and all result objects in the object library, a report outline and a corresponding drafting task are generated, and the drafting task is published to the task bus so that the working agent returns a task draft after executing the drafting task; the orchestration agent obtains the task draft from the task bus and integrates it to generate a query report corresponding to the to-be-processed query.

[0052] The server 102 can send the query report to the terminal 101, and the terminal 101 can display the query report to the user.

[0053] In this embodiment, an agent-based report generation method is provided, which can be applied to an orchestration agent, as shown in the figure, the specific process of the agent-based report generation method can be as follows: Figure 2

[0054] S110, generating a target execution plan according to a to-be-processed query.

[0055] The to-be-processed query is input by the user to describe the content of the demand, and the form of the to-be-processed query can be text, image, video, audio, etc., which can be selected according to actual needs. The to-be-processed query can finally be converted into a text form for subsequent processing. For example, when the to-be-processed query is an image, a video, or an audio, image recognition, voice recognition, etc. can be used to convert the image, the video, or the audio into text.

[0056] The to-be-processed query can include basic information of the report to be generated, such as the report theme, the report format, and special requirements. The specific content in the to-be-processed query can be set according to actual needs, which is not limited here.

[0057] ​The target execution plan is a plan needed to generate a query report corresponding to the to-be-processed query. The target execution plan can include multiple sub-tasks and dependent tasks corresponding to each sub-task. The sub-task refers to a task obtained by decomposing the task of generating a query report, which can be completed step by step. The dependent task refers to a preceding task of the sub-task, and the sub-task can be executed only when the preceding execution is completed. For example, to analyze the sales situation of product A in 25 years, the corresponding sales data needs to be obtained first. As an implementation manner, a large language model can be directly used to analyze and decompose the to-be-processed query to generate a corresponding target execution plan. For example, the to-be-processed query can be fused with a decomposition template to generate a decomposition prompt word, and the decomposition prompt word includes decomposition requirements. The decomposition prompt word is used to guide the large language model to perform decomposition processing on the to-be-processed query according to the decomposition requirements to generate a target execution plan.

[0058] The decomposition template refers to a prompt word template pre-set for decomposing a user query. The to-be-processed query can be fused with the decomposition template to obtain a decomposition prompt word. Specifically, the decomposition template can include decomposition requirements and query slots, wherein the decomposition requirements can include a role needed to be played by the large language model, a task needed to be processed, an output requirement, an example, and the like.

[0059] The decomposition template can be set according to actual needs. In the embodiment of the present application, the decomposition template can be:

[0060] "#Role

[0061] You are a world-class AI project research director and system architect. Your expertise is to break down a complex, high-level research goal into a detailed, structured, and executable task plan. You think carefully and logically, ensuring that each step in the plan is operational.

[0062] ]

[0064] Now, please generate a task plan for the following user query according to the above rules:

[0065] User original query: {{USER_QUERY}}

[0066] The USER_QUERY position can be used to fill in the to-be-processed query, so as to fuse the to-be-processed query into the decomposition template to generate a decomposition prompt word.

[0067] In some embodiments, the target execution plan can further include the task type corresponding to each subtask, for example, whether it is a search type or an analysis type, etc. In order to improve the accuracy and feasibility of the target execution plan, when generating the target execution plan according to the to-be-processed query, the to-be-processed query can be fused with the decomposition template to generate a decomposition prompt word, the decomposition prompt word including a decomposition requirement; the decomposition prompt word is used to guide the large language model to perform decomposition processing on the to-be-processed query according to the decomposition requirement, to generate a preliminary execution plan; if the format of the preliminary execution plan does not meet the preset format condition, a new preliminary execution plan is generated until the preset number of times is reached or the preset format condition is met; if the format of the preliminary execution plan meets the preset format condition, the preliminary execution plan is logically verified according to the dependent tasks and the task type of the subtasks; if the preliminary execution plan passes the logical verification, the preliminary execution plan is determined as the target execution plan.

[0068] The execution plan output by the aforementioned large language model is used as the preliminary execution plan, and the preliminary execution plan is subjected to format verification and logical verification to ensure the accuracy and feasibility of the plan. According to the aforementioned decomposition template, the large language model will output the preliminary execution plan in JSON format, and if the preliminary execution plan is a valid JSON, it is determined that the format of the preliminary execution plan meets the preset format condition; if the preliminary execution plan is not a valid JSON, it is determined that the format of the preliminary execution plan does not meet the preset format condition.

[0069] Specifically, the scheduling intelligent agent can parse the string returned by the large language model, and if the parsing fails, it indicates that the large language model does not follow the instructions in the decomposition prompt word, and it is determined that the preliminary execution plan is not a valid JSON; if the parsing is successful, it indicates that the large language model follows the instructions in the decomposition prompt word, and it is determined that the preliminary execution plan is a valid JSON.

[0070] If the format of the preliminary execution plan does not meet the preset format condition, corresponding failure processing can be performed to generate a new preliminary execution plan until the preset number of times is reached or the preset format condition is met. The failure processing can include retrying and replacing the model. Retrying means sending the error information that does not meet the preset format condition along with the decomposition prompt word to the large language model again and requiring the large language model to correct it. Replacing the model means replacing other large language models with stronger capabilities for processing. Whether it is retrying or replacing the model, it is considered as one failure processing, and a new preliminary execution plan can be obtained each time the failure processing is performed. The new execution plan is still determined according to the aforementioned manner whether it meets the preset format condition, and if it meets, the cycle is ended; if the number of failure processing reaches the preset number of times, it is considered that the error cannot be repaired, and the task decomposition failure is directly reported and ended.

[0071] Only when the format of the preliminary execution plan meets the preset format condition, the logical verification can be continued on the preliminary execution plan. The logical verification is to ensure that the dependent tasks in the preliminary execution plan are truly existent and logically reasonable, so as to ensure the feasibility of the entire preliminary execution plan.

[0072] Optionally, the logical verification can include existence verification, cycle verification and compliance verification, and the preliminary execution plan further includes a task type corresponding to each subtask. When the logical verification is performed on the preliminary execution plan, the existence verification can be performed on each dependent task according to the existence of the dependent task in the preliminary execution plan; the cycle verification is performed based on all subtasks and corresponding dependent tasks; the compliance verification is performed according to the relationship between the task type of each subtask and a preset type; and if the preliminary execution plan passes the existence verification, the cycle verification and the compliance verification, it is determined that the preliminary execution plan passes the logical verification.

[0073] For each subtask in the preliminary execution plan, the dependent task corresponding to the subtask is obtained, and the dependent task is usually represented by a task identifier. The task identifier corresponding to all dependent tasks is obtained, and the subtask corresponding to the task identifier is detected in the initial execution plan. If all task identifiers corresponding to the subtasks are detected, it is determined that the preliminary execution plan passes the existence verification. If there is any task identifier corresponding to an undetected subtask, it is determined that the preliminary execution plan does not pass the existence verification. For example, the preliminary execution plan includes three subtasks, and the task identifiers are task_1, task_2 and task_3. The dependent tasks of a subtask are task_1, task_2 and task_5. Since the preliminary execution plan does not include a subtask with a task identifier of task_5, it is determined that it does not pass the existence verification.

[0074] The cycle verification is a detection of the cycle relationship between the subtasks and the dependent tasks. For example, task_2 depends on task_1, and task_1 depends on task_2, which forms a dead loop, resulting in the preliminary execution plan being unable to be executed. Specifically, all subtasks and dependent tasks in the preliminary execution plan can be obtained, and a directed graph is established between each subtask and its corresponding dependent task, so that the task dependency relationship in the entire preliminary execution plan can be converted into a directed graph. If a loop is detected in the directed graph, it is determined that the preliminary execution plan does not pass the cycle verification. If no loop is detected in the directed graph, it is determined that the preliminary execution plan passes the cycle verification.

[0075] The conformity check is used to detect whether the task type corresponding to the subtask is one of the preset types, the preset types being predefined types, and the task type being selected from the preset types. It can be detected whether each task type is included in the preset types, and the uniqueness of the task identifier of each subtask is detected; if all task types are included in the preset types, and the task identifier of the subtask is unique, it is determined that the preliminary execution plan passes the conformity check.

[0076] If the preliminary execution plan passes the existence check, the loop check and the conformity check, it is determined that the preliminary execution plan passes the logic check, and the preliminary execution plan can be directly used as the target execution plan; if the preliminary execution plan does not pass any one of the existence check, the loop check and the conformity check, it is determined that the preliminary execution plan does not pass the logic check.

[0077] If it does not pass the logic check, a new preliminary execution plan can be regenerated, and the format and logic check can be performed again. If the logic check still fails after a certain number of retries, an error message can be generated and the process ends. If the logic check passes, it indicates that the preliminary execution plan has passed all checks, and the preliminary execution plan is reliable and executable, and can be directly used as the target execution plan.

[0078] In S120, a specified subtask is determined based on the dependent task, and the specified subtask is published to the task bus.

[0079] After the target execution plan is determined, the scheduling agent can convert each object in the JSON array into an internal task instance or entity, and identify the startable subtask to obtain the specified subtask.

[0080] The specified subtask refers to the current executable subtask, which can be determined based on the dependent task corresponding to each subtask. Alternatively, in the target execution plan, the dependent task corresponding to the subtask can be represented by an array, and the scheduling agent can traverse all subtasks in the target execution plan, obtain the array corresponding to the dependent task of each subtask, and determine the subtask with an empty array as the specified subtask.

[0081] For example, there are four subtasks, task_1: (type: search) searching for the latest research papers on L4 / L5 level autonomous driving technology; task_2: (type: search) searching for mainstream city planning white papers on smart traffic; task_3: (type: search) searching for economic benefit analysis reports on shared travel modes (such as robotaxi); and task_4 (type: extract, dependencies: [“task_1”]) extracting key technical parameters from the papers.

[0082] In this example, only the dependent task of task_4 is not empty, and task_1, task_2 and task_3 are all specified subtasks.

[0083] The number of specified subtasks can be multiple, and the scheduling agent can publish multiple specified subtasks on the task bus. Alternatively, each specified subtask can be encapsulated as a message, and then according to the task type of the specified subtask, the target topic is determined, and the message corresponding to the specified subtask is published to the target topic of the task bus.

[0084] Among them, when the scheduling agent publishes the specified subtask, the specified subtask can be encapsulated as a JSON format message, and the message structure contains subtask identification, task type, priority, dependent task corresponding task identification, execution context, success standard and quality requirement, timeout and retry policy, resource constraint, etc. For example, the specific structure of the message is as follows:

[0085] -task_id: task unique identifier

[0086] -task_type: task type (search / extract / analyze / draft / cite, etc.)

[0087] -priority: dynamic priority (0-100, real-time adjustment is supported)

[0088] -parent_task_id: parent task ID (used to build dependency relationship and backtracking)

[0089] -context: execution context (contains existing references, related concept list)

[0090] -success_criteria: success criteria and quality requirements

[0091] -timeout: timeout setting and retry policy

[0092] resource_constraints: resource constraints (such as API call limits)”. The task type and the topic on the task bus have a mapping relationship, that is, based on the task type of the specified subtask, the corresponding bus topic can be determined. As an implementation manner, a correspondence table of each preset type and the bus topic can be stored in advance, and when the task type is obtained, the target topic can be obtained by querying the correspondence table. As another implementation manner, a mapping rule can be set in advance, such as the name of the target topic corresponding to the task type contains the task type, that is, a task type is task_type, and the target topic is task.type; if the task type task_type is search, the target topic is task.search, and if the task type task_type is analyze, the target topic is task.analyze.

[0093] The specified subtask is published to the task bus, so that the working agent can obtain the corresponding specified subtask from the task bus and execute it. The working agent can return a specified result object after execution. The specified result object is published to the task bus after the working agent standardizes the task execution result of the specified subtask. The specified result object can contain result content, metadata, and object association data.

[0094] The result content is the content obtained after the task execution result is integrated, the metadata is the content obtained after the quality of the task execution result is evaluated and the key information is extracted, and the object relationship data can be used to represent the association relationship between the specified result object and other result objects, which can include an associated object and a corresponding association type. The associated object is another result object associated with the current specified result object, and the association type is the type to which the relationship between the two belongs.

[0095] S130, obtaining the specified result object from the task bus, and determining a new specified subtask based on the task state of the specified subtask to publish.

[0096] The working agent can obtain the specified subtask through the task bus, and publish the specified result object obtained after executing the specified subtask to the task bus, so that the orchestration agent can obtain the specified result object from the task bus. After the orchestration agent obtains the specified result object, it can be considered that the corresponding specified subtask has been executed, so that the state of all subtasks can be updated, and based on the state of each subtask, a new specified subtask is determined and published. It should be noted that after the target execution plan is generated, the orchestration agent can manage the task state of each subtask. The initial state of each subtask can be "not executed", when the orchestration agent publishes the specified subtask to the task bus, the task state of the specified subtask can be updated to "published / running", and according to the dependent task, the subtask that needs to depend on the specified subtask is marked as "waiting for dependency". Then, after the orchestration agent listens to the specified result object on the task bus, the state of the subtask dependent on the specified subtask can be marked as "ready", and the subtask with the task state "ready" is published to the task bus as a new specified subtask. For example, the aforementioned task_1, task_2 and task_3 are published to the task bus, and the dependent task of task_4 is task_1, if the specified result object of task_1 is obtained, the task state of task_4 can be updated and published to the task bus.

[0097] In S140, for each specified result object, the dimension score of the specified result object in multiple dimensions is calculated in combination with the result content, the metadata and the object association data.

[0098] As mentioned above, the specified result object includes result content, metadata and object association data, the metadata includes confidence and semantic label, using these data, the value of the specified result object can be quantitatively scored in multiple dimensions to obtain the corresponding dimension score. When calculating the dimension score, a plurality of dimensions including a relevant dimension, a novel dimension, a credible dimension and an impact dimension can be obtained in advance; the semantic similarity between the result content and the to-be-processed query is calculated to obtain the relevance score corresponding to the relevant dimension; the semantic label is compared with the semantic label of other result objects in the object library to obtain the novelty score corresponding to the novel dimension; the confidence is taken as the credibility corresponding to the credible dimension; and the impact of the specified result object on other subtasks is calculated to obtain the impact degree corresponding to the impact dimension.

[0099] The plurality of dimensions can be set according to actual needs, and the present embodiment sets relevant dimensions, novel dimensions, credibility dimensions and influence dimensions. The relevance score corresponding to the relevant dimensions can be used to measure the degree of fit of the specified result object and the original query and the global context, and the higher the relevance score, the higher the degree of fit. Specifically, the embedding vectors of the to-be-processed query and the global context can be spliced, and the cosine similarity between the spliced result and the result content is calculated as the semantic similarity, thereby obtaining the relevance score.

[0100] The novelty score corresponding to the novel dimension can be used to measure the degree of new knowledge brought by the specified result object. The specified result object contains a semantic label, and the semantic label is compared with the semantic labels of other result objects in the object library to calculate the novelty score. The higher the novelty, the more scarce the semantic label is. Specifically, the semantic labels of other result objects in the object library can be obtained as reference semantic labels; the semantic similarity between the semantic label and each reference semantic label is calculated; the number of reference semantic labels with a semantic similarity less than a preset similarity is counted to obtain a novelty number; and the ratio between the novelty number and the total number of result objects in the object library is calculated to obtain the novelty score.

[0101] The credibility corresponding to the credibility dimension can be used to measure the credibility of the specified result object. The higher the credibility, the more reliable the specified result object. Since the specified result object itself contains a confidence degree, the confidence degree can be directly used as the credibility.

[0102] The influence corresponding to the influence dimension can be used to measure the potential influence of the specified result object on other sub-tasks. The greater the influence value, the greater the potential influence. Specifically, the specified result object and all other sub-tasks can be input into a large language model, and the large language model can be used to determine whether the specified result object causes other sub-tasks to be redundant or need to be modified. According to the ratio of the number of redundant or modified other sub-tasks to the total number of other sub-tasks, the influence degree can be obtained. Of course, for the two cases of redundancy and modification, corresponding weights can also be set, for example, the weight of redundancy is greater than the weight of modification, and the ratio based on the weight is calculated as the influence degree.

[0103] As an implementation manner, the influence degree can be evaluated for each other sub-task, that is, the influence degree here is an array composed of multiple numbers, and each number represents the influence degree of the specified result object on one other sub-task. In the present embodiment, this way is used as the influence degree, that is, the influence degree of the specified result object compared with each other sub-task can be obtained by calculation, and similarly, the large language model can be directly used for determination.

[0104] After the dimension scores of the specified result object in multiple dimensions are obtained, the orchestration agent can determine whether real-time and structural adjustment of the target execution plan is needed based on the dimension scores. Optionally, if the novelty score is greater than the first score, a new subtask is created according to the semantic label and added to the target execution plan; if the specified type is included in the association type, a generated subtask is generated based on the specified type and added to the target execution plan; and if the influence degree is greater than the second score, a target subtask is determined according to the score basis of the influence degree and removed from the target execution plan.

[0105] The scores of the aforementioned multiple dimensions include the correlation score, the novelty score, the credibility, and the influence degree. The first score is a pre-set score value related to the novelty dimension. If the novelty score is greater than the first score, it indicates that a key research direction outside the target execution plan appears. In order to report more perfect and comprehensive, a new subtask is created according to the semantic label and added to the target execution plan.

[0106] Since the specified result object includes object association data, the object association data is other result objects having support, contradiction, and extension association types with the specified result object. The specified type can be pre-set according to actual needs. In the embodiment of the present application, the specified type is the contradiction type. The contradiction type is that the arguments of the specified result object and other result objects directly conflict or are contrary. In order to solve the conflict between information, a subtask is generated according to the specified type and added to the target execution plan. The newly generated subtask can have a higher priority, which is used to investigate and solve the information conflict between the two result objects, thereby ensuring the accuracy of the subsequent generated report.

[0107] The influence degree of the specified result object on each other subtask has been calculated in the aforementioned embodiment. The second score can be a pre-set score value related to the influence degree. For each other subtask, if its influence degree exceeds the second score, it can be considered that the specified result object may have included the task target of the subtask, and the subtask does not need to be executed again. Therefore, the subtask with the influence degree greater than the second score can be marked, and the subtask with the mark is shielded or directly removed from the target execution plan. At the same time, the orchestration agent can also adjust the priority of the remaining subtasks in the target execution plan according to the actual situation, so that the entire target execution plan is more reasonable.

[0108] S150, using the dimension scores and all result objects in the object library, generating a report outline and corresponding drafting tasks, and publishing the drafting tasks to the task bus.

[0109] The object library can be used to store the result objects generated by the working agent after executing the specified sub-tasks, each result object represents a knowledge that can be used to generate a report, and the orchestration agent continuously collects these knowledge for generating a report. When the research exploration reaches a saturation state, the orchestration agent can automatically switch modes to integrate the scattered knowledge nodes into a report.

[0110] Among them, whether the exploration reaches a saturation state can be determined based on the dimension score, the dimension score can measure the value of the specified result object, when the average value of all result objects in the entire object library no longer increases, or is lower than a certain value, it means that further exploration cannot bring greater benefits, that is, it is considered to reach a saturation state, and a report outline is generated using all the result objects.

[0111] As an implementation, using the dimension score and all result objects in the object library, a report outline and corresponding draft tasks are generated, and the draft tasks are published to the task bus. When the average dimension score of the result objects is less than a preset value, a result graph is generated using the result objects, the result graph includes nodes representing result objects, edges representing object association data, and weights representing the connection tightness of the edges; subgraphs are extracted from the result graph according to the weights of the edges in the result graph, the weights of the edges within the subgraphs are greater than the weights of the edges between the subgraphs, and the subgraphs include multiple result objects; for each subgraph, a subgraph topic is generated based on all result objects in the subgraph to obtain a report outline; draft tasks are generated according to the report outline, and the draft tasks are published to the task bus.

[0112] Specifically, the orchestration agent can determine whether the exploration reaches a saturation state according to the dimension score. For example, for each obtained specified result object, after calculating the dimension score, the dimension scores corresponding to multiple dimensions are fused to obtain a value score, and then the average value of the value score of the specified result object and the value scores of other result objects in the object library is calculated. If the average value is lower than a preset value, it is considered that the exploration reaches a saturation state, and further exploration of new content is difficult to bring greater knowledge gain, so all result objects can be directly used for report generation. For example, the dimension scores of multiple dimensions can be combined into a numerical value as a value score in a weighted manner, each result object in the object library has a corresponding value score, and the orchestration agent calculates the average value and compares it with the preset value after obtaining each specified result object, and then determines whether all result objects can be used for report generation. For example, refer to Figure 3 , which shows a schematic diagram of generating a report outline based on all result objects.

[0113] The result graph can include nodes, edges, and weights of the edges, and each node represents a result object. The object association data in the result object can include an associated object corresponding to the result object and a specific association type. Thus, for each result object, an edge between the result object and its corresponding associated object can be established. For example, the associated objects of the result object 1 are the result object 2, the result object 4, and the result object 5. The node of the result object 1 can establish edges with the nodes of the result object 2, the result object 4, and the result object 5, respectively, to represent the association relationship. The edge can represent the association relationship between two result objects. To further determine the closeness of the association, the weight of the edge can be calculated.

[0114] The object association data can include an associated object and an association type. Generating a result graph from a result object can include generating a node representing the result object for each result object, establishing an edge between the result object and each associated object, and calculating the weight of each edge according to the association type, the cosine similarity of the object group, and the confidence of the object group, where the object group is the two result objects connected by the edge.

[0115] The object group refers to the two result objects connected by an edge. When calculating the weight of the edge, the association type, the cosine similarity of the object group, and the confidence can be used for calculation. For example, the association type of the object group can be obtained, and the type score corresponding to the association type of the object group can be obtained according to the mapping relationship between the preset type and the preset score. The cosine similarity of the result content in the two result objects is calculated. The confidence of the two result objects is obtained, respectively, and is denoted as the first confidence and the second confidence. The type score, the cosine similarity, the first confidence, and the second confidence are fused to obtain a fusion result. The fusion result is normalized to obtain the weight of the edge. The weight of the edge can be used to represent the association strength between the result objects. The greater the weight, the tighter the association. After creating nodes and edges and calculating the weight of the edge, the result graph can be obtained.

[0116] The community detection algorithm can be used in the result graph to identify node clusters that are closely connected in the result graph. These clusters are subgraphs. For example, each node can be regarded as a separate community. For each node, the node can be redistributed from the current community to the community in which the neighbor node is located, and the modularity gain can be calculated based on the weight of the edge. If the modularity gain is greater than a positive threshold, the node is placed in the community with the maximum modularity gain. If the modularity gain is not as expected, the node is kept in the original community. This process is repeated until no movement of any node can improve the modularity or a specified number of cycles is reached, and a number of communities are output. Each community is a list of identifiers of result objects, and each community is a subgraph.

[0117] For each subgraph, a subgraph topic can be generated according to all result objects in the subgraph, and then a report outline is obtained. Alternatively, for each subgraph, semantic tags in all result objects can be extracted as core semantic tags; according to the connection degree of the result objects in the subgraph, a key object is determined from the subgraph and the summary information of the result content of the key object is extracted to obtain a key summary; the core semantic tags, the key summary and a topic template are collected to generate a topic prompt word; the topic prompt word is used to guide a large language model to refine a subgraph topic corresponding to the subgraph; and an association relationship between the subgraph and the subgraph topic is established to obtain a report outline.

[0118] Each subgraph contains the identities of multiple result objects, and the specific result object can be located through the identity of the result object. The result object contains result content, confidence, semantic tags and object association data, and the semantic tags of all result objects in the subgraph can be collected as core semantic tags. Each node in the subgraph has a corresponding edge, and the connection degree of each node in the subgraph can be measured by the number of edges, for example, a node has 3 edges, and its connection degree is 3. The more edges, the greater the connection degree. The node with the greater connection degree can be considered to be more critical in the subgraph, and the node with the highest connection degree can be determined as a key object.

[0119] For the key object, the result content corresponding to the key object can be obtained, and the corresponding summary information is extracted as a key summary.

[0120] Then, the core semantic tags, key summaries and other data can be filled into the topic template to generate a topic prompt word. The topic template is a prompt word template for extracting a subgraph topic, which can contain roles, tasks and specific output requirements. The topic template can be set according to actual needs, which is not limited here. In the embodiments of the present application, the topic template can be:

[0121] "#Role You are a top-notch research analyst, skilled at accurately extracting core topics from a series of related research materials.

[0122] #Task Please generate a concise and highly summarized Chinese topic name based on the following provided information cluster (representing a strongly associated research topic), which will be used as the title of the report chapter.

[0123] #Information cluster details

[0124] -Contains object ID: [art_123, art_456, art_789]

[0125] - Core semantic tags: ["MCTS", "Deep Reinforcement Learning", "Policy Network", "Value Network", "AlphaGo"]

[0126] - Key object summary (art_123): "This article details the basic principles of Monte Carlo Tree Search (MCTS) and its application in chess games..."

[0127] - Key object summary (art_456): "The success of AlphaGo lies in the combination of deep neural networks and MCTS, with its policy network and value network significantly improving search efficiency..."

[0128] # Output requirements

[0129] Please only output one most appropriate Chinese topic name without any explanation.

[0130] Among them, the information cluster can contain the extracted core semantic tags, key summaries, and object identifiers. Fill in the corresponding slots of this template with these contents to generate a topic prompt word.

[0131] Input the topic prompt word into the large language model to guide the large language model to analyze the core semantic tags and key summaries according to the requirements of the topic prompt word, and extract the most appropriate Chinese topic name as the subgraph topic. Each subgraph is processed in the above manner, and the subgraph topic corresponding to each subgraph is obtained. Associating the subgraph topic with its corresponding subgraph can generate a report outline.

[0132] It should be noted that the report outline is structured data, and the specific structure is as follows:

[0133] After generating the report outline, the corresponding drafting tasks can be generated based on the report outline, and the drafting tasks can be published to the task bus. The drafting task is a task created by the intelligent agent to generate content for each chapter or subchapter based on the report outline. The drafting task can include the title of the corresponding chapter, the identifier of the result object to be referred to, etc. After the drafting task is created, it can be sent to the corresponding topic on the task bus, so that the working intelligent agent can retrieve the task draft returned after executing the drafting task from the task bus.

[0134] S160, integrating the task draft according to the report outline to obtain the query report corresponding to the query to be processed.

[0135] The orchestration agent obtains the task draft generated by the work agent from the task bus, splices and integrates the task draft according to the report outline, and polishes the necessary task draft, so as to obtain the query report. Alternatively, when generating the query report, the splicing order of each task draft can be determined according to the report outline; each task draft is spliced according to the splicing order to obtain a report draft; and the style and logic of the report draft are adjusted and optimized by using a large language model to obtain the query report.

[0136] According to the order in the report outline, the splicing order of each task draft can be determined, and each task draft is spliced according to the splicing order, so that each task draft can be integrated into a report draft. Since the task drafts are usually generated by different work agents, there are differences in writing style, logic, etc. In order to make the readability of the final report stronger, a large language model can be called to polish, unify the style, adjust the logic of the report draft, and add the corresponding introduction or conclusion, etc., so that the report is more perfect, and the content processed by the large language model is used as the query report. The orchestration agent can send the query report to the client to display the query report to the user through the client.

[0137] The report generation method based on the agent provided in the embodiment of the application can be applied in various scenes requiring report, document, article generation. The orchestration agent in the embodiment of the application can generate a target execution plan based on a to-be-processed query, and determine to publish a specified subtask to a task bus from the target execution plan, so that the work agent executes. When the specified result object is obtained, the dimension score of the result object in multiple dimensions is calculated to generate a report outline and establish a drafting task based on all the result objects. The obtained task draft is integrated according to the report outline to generate a final query report. By using the task bus as a message middleware, asynchronous parallel processing can be realized, and the generation efficiency of the report is improved. The object association data is introduced to establish the connection between different result objects, so that deep analysis and logical reasoning can be performed, and the quality of the report can be effectively improved.

[0138] In the embodiment, a report generation method based on an agent is provided, which can be applied to a work agent, such as Figure 4 As shown in the figure, the specific process of the report generation method based on the agent can be as follows:

[0139] S210, obtain the specified subtask from the task bus and execute to obtain the task execution result, the specified subtask is determined by the orchestration agent from the target execution plan, and the target execution plan is generated based on the to-be-processed query.

[0140] The work agent can subscribe to one or more specific topics according to its type. For example, a search type work agent can subscribe to the task.search topic, so that multiple work agents exist under each topic and listen to the tasks on the task bus.

[0141] When the orchestration agent publishes the specified subtask according to the task type to the corresponding topic of the task bus, all work agents listening to the task of the topic will be notified. In order to ensure that a specified subtask can only be consumed by one work agent, the work agent can perform competitive consumption, for example, the first work agent that successfully claims the specified subtask removes the message from the task bus or marks its state as processing to prevent other work agents from repeating execution.

[0142] Therefore, the work agent can obtain the specified subtask from the task bus and execute it. Specifically, the context information can be parsed from the task message, and the prompt information can be obtained from the specified subtask; the to-be-executed query is generated in combination with the context information and the prompt information; and the task execution result is obtained by executing the to-be-executed query.

[0143] Among them, the context information of the task message contains the identification of the result object considered relevant by the orchestration agent, and the prompt information in the specified subtask can contain specific content to be processed. The work agent uses a large language model to generate a series of more accurate and diversified to-be-executed queries in combination with the context information and the prompt information, and the work agent processes according to the to-be-executed query to obtain the corresponding task execution result.

[0144] S220, performing content integration processing on the task execution result to obtain result content.

[0145] At this time, the task execution result obtained contains various text segments or URLs, and the work agent can process it to generate standardized result objects. First, the content integration and cleaning of the task execution result can be performed. Specifically, the work agent can perform deduplication processing on the content in the task execution result, and clean the text content, etc., to form the result content. For example, taking a search agent as an example, it can deduplicate the search content and preliminarily clean the downloaded content to obtain the result content.

[0146] S230, performing quality evaluation and content extraction on the task execution result through a large language model to generate metadata.

[0147] The working agent can call the large language model to perform quality evaluation and content extraction processing on the content of the task execution result. The confidence of the task execution result can be obtained through the quality evaluation, and the key information can be extracted from the task execution result through the content extraction processing.

[0148] Optionally, the relevance between the result content and the prompt information can be calculated, for example, the relevance can be measured by cosine similarity. When the cosine similarity is greater than a specified similarity and the content source is mostly from authoritative websites, the confidence can be set to be relatively high. When the cosine similarity is not greater than the specified similarity or the content source is not from authoritative websites, the confidence can be set to be relatively low. The higher the confidence is. The large language model can read and analyze the result content, and summarize the result content to extract semantic tags. The confidence and the semantic tags constitute the metadata.

[0149] S240, classifying the logical relationship between the task execution result and other result objects to obtain object association data of the task execution result.

[0150] After obtaining the result content and the metadata, the working agent can continue to compare the task execution result with other result objects in the object library to identify the logical association between the other result objects in the object library and the task execution result, for example, support, contradiction, expansion, etc., and further generate the object association data.

[0151] The object association data includes the associated objects associated with the task execution result and the corresponding association types. As an implementation, the classification processing of the logical relationship between the task execution result and other result objects to obtain the object association data of the task execution result can be based on the semantic similarity between the task execution result and the result objects in the object library to determine candidate result objects. The task execution result is combined with each candidate result object to obtain a plurality of result pairs. For each result pair, the summary information of the result pair is combined with a classification template to obtain a classification prompt word, the classification prompt word includes definitions of a plurality of preset relationship types. The classification prompt word is used to guide the large language model to determine the association type of the result pair based on the definition of the preset relationship type. The candidate result object in the result pair with the specified type of the association type is determined as the associated object of the task execution result.

[0152] The object repository stores result objects corresponding to completed subtasks. Each result object has its own result content, which is denoted as the first result content for ease of description. The result content of the currently obtained task execution result is denoted as the second result content. The semantic similarity between the second result content and each first result content is calculated. The result objects corresponding to multiple second result contents are sorted according to their semantic similarity from highest to lowest. Then, a predetermined number of result objects with the highest ranking are selected as candidate result objects. Optionally, candidate result objects can also be obtained from the object repository through an approximate nearest neighbor search using the embedding vectors corresponding to the second result content.

[0153] For each candidate result object, it is combined with the task execution result to form a result pair. That is, the result pair contains a candidate result object and the task execution result, thus obtaining multiple result pairs, and the number of result pairs is the same as the number of candidate result objects.

[0154] For each result pair, the candidate result objects and task execution results within the result pair can be analyzed to clarify the logical relationship between them. Optionally, the summary information of the result pair can be combined with a classification template to obtain classification prompts. The classification template is a pre-set prompt template used to confirm the logical relationship between summaries. This prompt template can be set according to actual needs. In this embodiment of the invention, the classification template can be:

[0155] "#Role

[0156] You are a logically rigorous and experienced academic researcher. Your task is to accurately determine the logical relationship between two research abstracts.

[0157] #Relation Type Definition

[0158] Please choose the option that best describes the relationship between "Abstract B" and "Abstract A" from the following five options:

[0159] -**supports**: The conclusions, data, or arguments in Abstract B provide direct positive evidence for the core arguments in Abstract A.

[0160] -**Contradicts**: The conclusions, data, or arguments in Abstract B directly conflict with or contradict the core arguments in Abstract A.

[0161] - **expands:** Abstract B builds upon the topics discussed in Abstract A, providing more details, background, application examples, or deeper explanations, but without altering the core arguments of A. - **related:** Abstracts A and B discuss the same or similar topics, but do not form a direct supporting, contradictory, or expanding relationship.

[0162] -**unrelated**: The topics discussed by the two are basically unrelated.

[0163] #Data to be analyzed

[0164] -**Summary A (existing object)**: {{SUMMARY_OF_CANDIDATE_ARTIFA CT}}

[0165] -**Summary B (Newly Generated Object)**: {{SUMMARY_OF_NEW_ARTIFACT}}

[0166] #Output Requirements

[0167] Please output your judgment as a JSON object, without including any additional explanatory text. The JSON object must contain two fields:

[0168] 1. "relation"(string): The relationship type you selected.

[0169] 2. "justification" (string): A short sentence explaining the reasons for your judgment.

[0170] The classification template includes the definition of roles, data to be analyzed, preset relationship types, and output requirements. The data to be analyzed is the content that needs to be filled into the classification template. For example, in the classification template, summary A needs to be filled with the summary of the candidate result objects in the result pair, such as the summary information of the result content; summary B can be filled with the summary information of the result content in the task execution result.

[0171] Preset relationship types are also pre-set in the classification template, which can include several types such as support, contradiction, extension, relatedness, and irrelevantness. The definitions of each type are clearly given. After filling in the corresponding content in the data to be analyzed in the classification template, classification prompts will be obtained.

[0172] By inputting classification prompts into the large language model, the model can be guided to fully understand the definitions of each preset relationship type and analyze the data to be analyzed in the classification prompts to determine the preset relationship type to which the two belong, obtain the association type of the result pair and output it.

[0173] Each result object can obtain its corresponding association type, since the preset relationship type contains weak logical associations such as irrelevant and relevant, the result pairs with the association type being the specified type can be filtered and reserved. The specified type can be set according to actual needs, and in the embodiment of the application, the specified type is support, contradiction and extension. The candidate result object in the result pair obtained after filtering is used as the association object of the task execution result.

[0174] Optionally, the object association data can be a specific field in the metadata, or a separate field, which can be set according to actual needs. For example, the field is knowledge_links, which can contain the identification of the association object, the specific association type and the reason. The content of the field can be as follows:

[0175]

[0176] S250, merging the result content, the metadata and the object association data into a specified result object, and publishing the specified result object to the task bus.

[0177] The obtained result content, metadata and object association data are merged, so that the task execution result is standardized into a specified result object, which can be written into the object library. Moreover, the working agent can publish the specified result object to the task bus, so that the scheduling agent obtains the specified result object from the task bus and generates a query report based on the specified result object.

[0178] In order to ensure the consistency of the task state of the specified subtask, the working agent can assemble the object association data, the confidence, the semantic label and the result content when generating and publishing the specified result object, generate the specified result object, write the specified result object into the object library, update the task state of the specified subtask to completed and record the completion time, and publish the specified result object to the task bus if the specified result object is successfully written into the object library and the task state is successfully updated.

[0179] The object association data, the confidence, the semantic label and the result content can be assembled to form the specified result object. In order to avoid the situation that the task has been executed but the task state has not been modified, or the situation that the task has been executed but the corresponding specified result object has not been stored, the working agent can start a transaction at the database level, and execute two steps, one is to write the specified result object into the object library, and the other is to update the task state of the specified subtask to completed and record the completion time.

[0180] After the two steps are executed, the last construction event is executed and published, that is, the specified result object is published to the task bus. If any one step is not executed successfully, the entire transaction is rolled back. In order to improve efficiency and save network bandwidth, the published specified result object is a time message, that is, a lightweight notification, which can contain part of the content in the specified result object. For example, the structure of the message can be as follows:

[0181]

[0182] Wherein, event_type is used to explicitly indicate the nature of the event (for example, future events such as task.failed, project.completed, etc.), so as to facilitate efficient filtering by the subscriber, that is, the orchestration agent. artifact_id is the identification of the result object, which is a "pointer" or "key" to obtain the complete specified result object. confidence_score is the confidence score, which is exposed in advance by the working agent, and can allow the orchestration agent to make a preliminary, low-cost priority judgment before obtaining the complete specified result object, for example, to ignore all events with a confidence score below a certain threshold.

[0183] It should be noted that the working agent needs to publish the specified result object to a specially designed topic, for example, result.artifacts. This topic is physically isolated from the topic for publishing sub-tasks to avoid interference with each other. Thus, the orchestration agent can obtain the result object through the task bus and process the result object accordingly to generate a query report corresponding to the to-be-processed query.

[0184] The working agent in the embodiment of the application can integrate the content of the task execution result obtained by executing the specified sub-task to obtain result content, perform quality evaluation and content extraction to obtain metadata, classify according to the logical relationship with other result objects to obtain object association data, and then use these data to obtain the specified result object after publication for the orchestration agent to generate a report. Through quality evaluation and content processing, more valuable information can be obtained from the task execution result, and object association data is introduced to establish the association between different result objects, which facilitates the orchestration agent to perform deep analysis and logical reasoning, and thus the generation quality of the report can be improved.

[0185] In the embodiment, a kind of agent-based report generation method is provided, which can be applied to orchestration agent and working agent, as shown in Figure 5 The specific process of the agent-based report generation method can be as follows:

[0186] S310, the orchestration agent generates a target execution plan according to the to-be-processed query;

[0187] S320, the scheduling agent determines a specified subtask based on the dependent task and publishes the specified subtask to the task bus;

[0188] S330, the working agent obtains the specified subtask from the task bus and executes to obtain a task execution result;

[0189] S340, the working agent processes the task execution result to generate a specified result object and publishes it to the task bus;

[0190] S350, the scheduling agent obtains the specified result object from the task bus and determines a new specified subtask based on the task state of the specified subtask to publish;

[0191] S360, the scheduling agent calculates the dimension score of the specified result object in multiple dimensions for each specified result object in combination with the result content, metadata and object association data;

[0192] S370, the scheduling agent generates a report outline and the corresponding draft task using the dimension score and all result objects in the object library, and publishes the draft task to the task bus;

[0193] S380, the working agent obtains the draft task from the task bus and executes to obtain a task draft and publishes it to the task bus;

[0194] S390, the working agent integrates the task draft according to the report outline to obtain a query report corresponding to the query to be processed.

[0195] The contents of S310 to S390 above can refer to the corresponding description in the foregoing embodiments, and can also refer to Figure 6 , which shows an interaction schematic diagram of the report generation method based on an agent. Among them, the report generation system can include a scheduling agent, a plurality of working agents, a task bus and a shared state and object library.

[0196] Among them, after receiving the query to be processed, the scheduling agent can first create an entity in the shared state and object library for tracking the life cycle of the entire report generation task. Specifically, the entity can include the following contents:

[0197] “project_id: global unique identifier, used to associate all related tasks and objects

[0198] -original_query: original user query, as a basis for relevance score

[0199] -current_plan: current research plan (hierarchical task tree that can be dynamically modified)

[0200] -priority_matrix: task priority matrix, supporting dynamic adjustment

[0201] -completion_criteria: completion criteria and threshold configuration

[0202] -resource_budget: resource budget and real-time usage tracking

[0203] -global_context: global context information, including discovered key concepts and terms

[0204] The orchestration agent can then call a large language model to analyze and process the query to be processed to generate a target execution plan. The large language model can identify key research dimensions such as technical impact, market impact, and social impact, and generate corresponding sub-tasks for each dimension to obtain a preliminary execution plan. The orchestration agent then performs format and logic checks on the preliminary execution plan, and if both checks pass, the target execution plan is obtained. Each sub-task in the target execution plan can be converted into an instance or entity for subsequent processing.

[0205] The orchestration agent determines a sub-task with an empty dependency task as a specified sub-task, and publishes the specified sub-task to the task bus. The specified sub-task can be multiple, and multiple specified sub-tasks can be published to the corresponding topic of the task bus for consumption by the working agent for parallel processing.

[0206] After the orchestration agent publishes the specified sub-task, it can listen to the topic of the specified result object, such as the results.artifacts topic. The orchestration agent can also maintain the task status of all sub-tasks. After the orchestration agent publishes the specified sub-task, it can mark the internal state of the specified sub-task as published or running.

[0207] Under each topic, there are multiple working agents that competitively consume the published messages of the specified sub-task to ensure that only one working agent can consume a message. That is, once a working agent claims the message, the corresponding specified sub-task is removed from the task queue or the state is marked as processing. The working agent that successfully claims the message can send a request to the object library to update the state of the specified sub-task entity to executing, and record its number, start time, etc. to the entity.

[0208] Then the working agent can parse the fields in the message to obtain context information and obtain prompt information of the specified subtask. The two are combined to generate a more accurate and diversified query, which is then executed to obtain the task execution result. The working agent then performs standardized processing based on the task execution result to generate confidence, semantic label, object association data, result content, etc. as the specified result object and publishes it to the task bus.

[0209] After the orchestration agent listens to the specified result object, it is put into the queue, and the state of other subtasks that depend on the specified result object is updated to ready for publishing. Similarly, after publishing, it continues to listen and update the task state. Through the queue, it can also be processed in time when multiple subtasks are completed at the same time, which can avoid data loss or slow response.

[0210] The orchestration agent performs multi-dimensional evaluation on the specified result object to obtain the corresponding dimension score. The dimension score can be used to dynamically adjust the target plan or determine whether to start integrating data. If it is determined that data integration can start, the data can be used to generate a report outline and publish a draft task to the draft topic on the task bus. The working agent subscribed to the draft topic obtains the task and executes it, and then returns the task draft to the task bus.

[0211] The orchestration agent obtains the task draft from the task bus for integration to generate a query report. The generated query report can be sent to the client to display the query report to the user through the client.

[0212] The report generation method based on the agent provided in the embodiment of the application can query and decompose the orchestration agent, determine a plurality of subtasks for parallel execution to improve the report generation speed, and the working agent is not simply a task execution agent that returns a result. The working agent processes the result in multiple ways to make the information contained more rich, and then the orchestration agent can use the information to quickly generate a high-quality report.

[0213] To better implement the above method, the embodiment of the application further provides an agent-based report generation device. The agent-based report generation device is applied to an orchestration agent and can be integrated in an electronic device, which can be a terminal, a server, etc. The terminal can be a mobile phone, a tablet computer, a smart Bluetooth device, a notebook computer, a personal computer, etc. The server can be a single server or a server cluster composed of multiple servers.

[0214] For example, in this embodiment, the agent-based report generation device is integrated in a server.

[0215] For example, as Figure 7As shown, the report generation apparatus 400 of the agent applied to the orchestration agent can include a plan generation module 410, a determination module 420, a task publishing module 430, a calculation module 440, an outline generation module 450, and a report generation module 460.

[0216] The plan generation module 410 is configured to generate a target execution plan according to the to-be-processed query, the target execution plan including a plurality of sub-tasks and a dependent task corresponding to each sub-task.

[0217] The determination module 420 is configured to determine a specified sub-task based on the dependent task, and publish the specified sub-task to a task bus, so that a working agent acquires the specified sub-task and executes the same to return a specified result object, the specified result object including result content, metadata, and object association data.

[0218] The task publishing module 430 is configured to acquire the specified result object from the task bus, and determine a new specified sub-task based on a task state of the specified sub-task to publish the new specified sub-task.

[0219] The calculation module 440 is configured to, for each specified result object, calculate a dimension score of the specified result object in a plurality of dimensions in combination with the result content, the metadata, and the object association data.

[0220] The outline generation module 450 is configured to generate a report outline and a corresponding draft task by using the dimension score and all result objects in an object library, and publish the draft task to the task bus, so that the working agent returns a task draft after executing the draft task.

[0221] The report generation module 460 is configured to integrate the task draft according to the report outline to obtain a query report corresponding to the to-be-processed query.

[0222] In some embodiments, the outline generation module 450 is specifically configured to:

[0223] If the average dimension score of the result object is less than a preset value, a result graph is generated by using the result object, the result graph including nodes representing the result object, edges representing the object association data, and weights representing a connection tightness of the edges;

[0224] According to the weights of the edges in the result graph, a sub-graph is extracted from the result graph, the weights of the edges inside the sub-graph being greater than the weights of the edges between the sub-graphs, and the sub-graph including a plurality of result objects;

[0225] For each sub-graph, a sub-graph theme is generated based on all result objects in the sub-graph to obtain a report outline;

[0226] A drafting task is generated according to the report outline, and the drafting task is published to a task bus.

[0227] In some embodiments, the object association data comprises an associated object and an association type, the metadata comprises a confidence level, and the outline generation module 450 is specifically configured to:

[0228] For each result object, a node representing the result object is generated;

[0229] An edge between the result object and each of the associated objects is established;

[0230] For each edge, a weight of the edge is calculated according to the association type, a cosine similarity of object groups, and a confidence level of the object groups, the object groups being two result objects connected by the edge.

[0231] In some embodiments, the metadata comprises a confidence level and a semantic label, and the outline generation module 450 is specifically configured to:

[0232] For each subgraph, a semantic label in all result objects is extracted as a core semantic label;

[0233] According to a connection degree of the result object in the subgraph, a key object is determined from the subgraph and an abstract information of a result content of the key object is extracted to obtain a key abstract;

[0234] A theme prompt word is generated in combination with the core semantic label, the key abstract, and a theme template;

[0235] The theme prompt word is used to guide a large language model to refine a subgraph theme corresponding to the subgraph;

[0236] An association relationship between the subgraph and the subgraph theme is established to obtain a report outline.

[0237] In some embodiments, the metadata comprises a confidence level and a semantic label, and the calculation module 440 is specifically configured to:

[0238] A plurality of dimensions are obtained, the plurality of dimensions comprising a relevance dimension, a novelty dimension, a credibility dimension, and an impact dimension;

[0239] A semantic similarity between the result content and a to-be-processed query is calculated to obtain a relevance score corresponding to the relevance dimension;

[0240] The semantic label is compared with semantic labels of other result objects in an object library to obtain a novelty score corresponding to the novelty dimension;

[0241] The confidence level is taken as a credibility corresponding to the credibility dimension;

[0242] Calculate the impact of the specified result object on other sub-tasks to obtain the impact degree corresponding to the impact dimension.

[0243] In some embodiments, the object association data includes an associated object and an association type, and the agent-based report generation apparatus 400 further includes an updating module. For each specified result object, after calculating the dimension score of the specified result object in multiple dimensions based on the result content, metadata, and object association data, the updating module is specifically configured to:

[0244] If the novelty score is greater than the first score, a new sub-task is created according to the semantic label and added to the target execution plan;

[0245] If the association type includes a specified type, a generation sub-task is generated based on the specified type and added to the target execution plan;

[0246] If the impact degree is greater than the second score, a target sub-task is determined according to the scoring basis of the impact degree and removed from the target execution plan.

[0247] In some embodiments, the target execution plan further includes a task type corresponding to each sub-task, and the plan generation module 410 is specifically configured to:

[0248] Fuse the to-be-processed query with the decomposition template to generate a decomposition prompt word, the decomposition prompt word including a decomposition requirement;

[0249] Use the decomposition prompt word to guide the large language model to perform decomposition processing on the to-be-processed query according to the decomposition requirement to generate a preliminary execution plan;

[0250] If the format of the preliminary execution plan does not meet the preset format condition, a new preliminary execution plan is generated until a preset number of times is reached or the preset format condition is met;

[0251] If the format of the preliminary execution plan meets the preset format condition, the preliminary execution plan is logically verified according to the dependent tasks and task types of the sub-tasks;

[0252] If the preliminary execution plan passes the logical verification, the preliminary execution plan is determined as the target execution plan.

[0253] In implementation, each of the above modules can be implemented as an independent entity, or can be combined as the same or several entities. The specific implementation of each of the above modules can be referred to the method embodiments described above, and will not be described here.

[0254] As can be seen from the above, the agent-based report generation device for scheduling an agent in the embodiment can integrate the task execution result obtained by executing the specified subtask to obtain result content, perform quality evaluation and content extraction to obtain metadata, classify according to the logical relationship with other result objects to obtain object association data, and then use the data to obtain the specified result object after publishing to generate a report for the scheduling agent. Through quality evaluation and content processing, more valuable information can be obtained from the task execution result, and object association data is introduced to establish the association relationship between different result objects, which facilitates the scheduling agent to perform deep analysis and logical reasoning, and thus the quality of the report can be improved.

[0255] To better implement the above method, an agent-based report generation device is further provided in the embodiment of the application. The agent-based report generation device is applied to a working agent and can be integrated in an electronic device, which can be a terminal, a server or the like. The terminal can be a mobile phone, a tablet computer, a smart Bluetooth device, a notebook computer, a personal computer or the like, and the server can be a single server or a server cluster composed of multiple servers.

[0256] For example, in the embodiment, the agent-based report generation device is integrated in a server, and the method of the embodiment of the application is described in detail.

[0257] For example, as shown in Figure 8 The agent-based report generation device 500 applied to the working agent can include an execution obtaining module 510, an integration module 520, an extraction module 530, a classification module 540 and an object publishing module 550.

[0258] The execution obtaining module 510 is configured to obtain and execute the specified subtask from the task bus to obtain the task execution result, wherein the specified subtask is determined by the scheduling agent from the target execution plan, and the target execution plan is generated based on the to-be-processed query.

[0259] The integration module 520 is configured to perform content integration processing on the task execution result to obtain result content.

[0260] The extraction module 530 is configured to perform quality evaluation and content extraction processing on the task execution result by using a large language model to generate metadata.

[0261] The classification module 540 is configured to perform classification processing on the logical relationship between the task execution result and other result objects to obtain object association data of the task execution result.

[0262] The object publishing module 550 is configured to combine the result content, the metadata, and the object association data into a specified result object, and publish the specified result object on the task bus, so that the intelligent agent generates a report corresponding to the to-be-processed query based on the specified result object on the task bus.

[0263] In some embodiments, the object association data includes an association object and an association type, and the classification module 540 is specifically configured to:

[0264] determine a candidate result object based on semantic similarity between the task execution result and a result object in the object library;

[0265] combine the task execution result with each candidate result object respectively to obtain a plurality of result pairs;

[0266] for each result pair, combine the summary information of the result pair with a classification template to obtain a classification prompt word, the classification prompt word including definitions of a plurality of preset relationship types;

[0267] use the classification prompt word to guide the large language model to determine an association type of the result pair based on the definitions of the preset relationship types;

[0268] determine the candidate result object in the result pair with the specified type as an association object of the task execution result.

[0269] In some embodiments, the object publishing module 550 is specifically configured to:

[0270] assemble the object association data, the confidence, the semantic label, and the result content to generate a specified result object;

[0271] write the specified result object into the object library;

[0272] update the task state of the specified subtask to completed, and record the completion time;

[0273] if the specified result object is successfully written into the object library and the task state is successfully updated, publish the specified result object on the task bus.

[0274] In specific implementation, each of the above modules can be implemented as an independent entity, or can be combined as the same or several entities. The specific implementation of each of the above modules can be referred to the method embodiments described above, and will not be described here.

[0275] From the above, the embodiment applied to the agent-based report generation device of the working agent can integrate the task execution result obtained by executing the specified subtask to obtain the result content, perform quality evaluation and content extraction to obtain the metadata, classify according to the logical relationship with other result objects to obtain the object association data, and then use the data to obtain the specified result object publishing to generate the report for the arrangement agent. Through the quality evaluation and content processing, more valuable information can be obtained from the task execution result, and the object association data is introduced to establish the association relationship between different result objects, which is convenient for the arrangement agent to perform deep analysis and logical reasoning, and thus the quality of the report can be improved.

[0276] The embodiment of the present application also provides an electronic device, which can be a terminal, a server or the like. The terminal can be a mobile phone, a tablet computer, a smart Bluetooth device, a notebook computer, a personal computer or the like, and the server can be a single server or a server cluster composed of multiple servers.

[0277] In some embodiments, the agent-based report generation device can also be integrated in multiple electronic devices, for example, the agent-based report generation device can be integrated in multiple servers, and the multiple servers can be used to implement the agent-based report generation method of the present application.

[0278] In the embodiment, the electronic device is taken as an example to be described in detail, for example, as shown in FIG. 1, Figure 9 which shows a structural schematic diagram of the electronic device related to the embodiment of the present application, and specifically,

[0279] The electronic device can include a processor 610 with one or more processing cores, a memory 620 with one or more computer readable storage media, a power supply 630, an input module 640, a communication module 650 and the like. Those skilled in the art can understand that, Figure 9 the structure of the electronic device shown in FIG. 1 does not constitute a limitation on the electronic device, and can include more or less components than the diagram, or combine certain components, or different component arrangements. Among them:

[0280] The processor 610 is the control center of the electronic device, and connects various parts of the entire electronic device through various interfaces and lines, and performs various functions of the electronic device and processes data by running or executing software programs and / or modules stored in the memory 620 and calling data stored in the memory 620. In some embodiments, the processor 610 can include one or more processing cores; in some embodiments, the processor 610 can integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface, and application programs, etc., and the modem processor mainly processes wireless communication. It can be understood that the above-mentioned modem processor can also not be integrated into the processor 610.

[0281] The memory 620 can be used to store software programs and modules, and the processor 610 executes various functions and data processing by running the software programs and modules stored in the memory 620. The memory 620 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application program required by a function (such as a sound playing function, an image playing function, etc.), etc.; the data storage area can store data created according to the use of the electronic device, etc. In addition, the memory 620 can include a high-speed random access memory, and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state memory device. Accordingly, the memory 620 can also include a memory controller to provide access for the processor 610 to the memory 620.

[0282] The electronic device also includes a power supply 630 for supplying power to various components, and in some embodiments, the power supply 630 can be logically connected to the processor 610 through a power management system, so as to realize functions such as management of charging, discharging, and power consumption management through the power management system. The power supply 630 can also include one or more direct or alternating current power supplies, recharging systems, power failure detection circuits, power converters or inverters, power state indicators, etc.

[0283] The electronic device can also include an input module 640, which can be used to receive input digital or character information, and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0284] The electronic device can also include a communication module 650, which in some embodiments can include a wireless module, and the electronic device can perform short-range wireless transmission through the wireless module of the communication module 650, thereby providing the user with wireless broadband Internet access. For example, the communication module 650 can be used to help the user send and receive emails, browse web pages, and access streaming media, etc.

[0285] Although not shown, the electronic device can further include a display unit and the like, which will not be described herein. Specifically, in the present embodiment, the processor 610 in the electronic device will load the executable file corresponding to the process of one or more application programs into the memory 620 according to the following instructions, and run the application program stored in the memory 620 by the processor 610, thereby implementing the steps in the method of the embodiments of the present application.

[0286] The specific implementation of each operation above can refer to the previous embodiments, which will not be described herein.

[0287] Those skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructions, or by related hardware controlled by the instructions, which can be stored in a computer readable storage medium and loaded and executed by a processor.

[0288] To this end, the embodiments of the present application provide a computer readable storage medium, which stores a plurality of instructions capable of being loaded by a processor to execute the steps in any of the agent-based report generation methods provided by the embodiments of the present application.

[0289] The storage medium can include a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0290] According to an aspect of the present application, a computer program product or computer program is provided, which includes computer programs / instructions stored in a computer readable storage medium. The processor of the electronic device reads the computer programs / instructions from the computer readable storage medium, and the processor executes the computer programs / instructions, so that the electronic device executes the method provided in any of the various optional implementations of the agent aspect or the working agent aspect provided in the above embodiments.

[0291] Since the instructions stored in the storage medium can execute the steps in any of the agent-based report generation methods provided by the embodiments of the present application, the beneficial effects that can be achieved by any of the agent-based report generation methods provided by the embodiments of the present application can be achieved, which will be described in detail in the previous embodiments, and will not be described herein.

[0292] The above describes in detail the report generation method based on the agent provided by the embodiment of the present application, and the principle and implementation manner of the present application are described by applying specific examples. The above embodiment is only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, the specific implementation manner and application range will be changed according to the idea of the present application. In conclusion, the content of the specification should not be understood as the limitation of the present application.

Claims

1. An agent-based report generation method, characterized by, The method applied to an intelligent agent comprises: generating a target execution plan according to a query to be processed, the target execution plan comprising a plurality of sub-tasks and a dependent task corresponding to each sub-task; determining a specified sub-task based on the dependent task, and publishing the specified sub-task to a task bus so that a working agent obtains the specified sub-task and executes it to return a specified result object, the specified result object comprising result content, metadata, and object association data; obtaining the specified result object from the task bus, and determining a new specified sub-task based on the task state of the specified sub-task to publish the new specified sub-task; for each specified result object, calculating a dimension score of the specified result object in multiple dimensions based on the result content, metadata, and object association data; generating a report outline and a corresponding drafting task using the dimension score and all result objects in an object library, and publishing the drafting task to the task bus so that the working agent returns a task draft after executing the drafting task; integrating the task draft according to the report outline to obtain a query report corresponding to the query to be processed.

2. The method of claim 1, wherein, The method of generating a report outline and a corresponding drafting task using the dimension score and all result objects in an object library, and publishing the drafting task to the task bus comprises: if the average dimension score of the result object is less than a preset value, generating a result graph using the result object, the result graph comprising nodes representing result objects, edges representing object association data, and weights representing the connection tightness of the edges; extracting a subgraph from the result graph according to the weights of the edges in the result graph, the weights of the edges within the subgraph being greater than the weights of the edges between the subgraphs, and the subgraph comprising a plurality of result objects; for each subgraph, generating a subgraph theme based on all result objects in the subgraph to obtain a report outline; generating a drafting task according to the report outline and publishing the drafting task to the task bus.

3. The method of claim 2, wherein, The object association data comprises associated objects and association types, and the metadata comprises a confidence level, and the method of generating a result graph using the result object comprises: for each result object, generating a node representing the result object; establishing an edge between the result object and each associated object; for each edge, calculating the weight of the edge according to the association type, the cosine similarity of the object group, and the confidence level of the object group, the object group being the two result objects connected by the edge.

4. The method of claim 2, wherein, The metadata comprises a confidence level and a semantic label, and the method of generating a subgraph theme based on all result objects in the subgraph to obtain a report outline for each subgraph comprises: for each subgraph, extracting semantic labels in all result objects as core semantic labels; determining a key object from the subgraph according to the connection degree of the result object in the subgraph and extracting the abstract information of the result content of the key object to obtain a key abstract; generating a theme prompt word by combining the core semantic label, the key abstract, and a theme template; using the theme prompt word to guide a large language model to extract a subgraph theme corresponding to the subgraph; Establish an association between the subgraph and the subgraph theme to obtain a report outline.

5. The method of claim 1, wherein, The metadata includes a confidence and a semantic label, and for each specified result object, the dimension score of the specified result object in multiple dimensions is calculated in combination with the result content, metadata and object association data, including: Obtain a plurality of preset dimensions, including a relevance dimension, a novelty dimension, a credibility dimension and an impact dimension; Calculate the semantic similarity between the result content and the to-be-processed query to obtain a relevance score corresponding to the relevance dimension; Compare the semantic label with the semantic label of other result objects in the object library to obtain a novelty score corresponding to the novelty dimension; The confidence is used as the credibility corresponding to the credibility dimension; Calculate the impact of the specified result object on other subtasks to obtain an impact degree corresponding to the impact dimension.

6. The method of claim 5, wherein, The object association data includes an associated object and an association type, and after the dimension score of the specified result object in multiple dimensions is calculated in combination with the result content, metadata and object association data for each specified result object, the method further includes: If the novelty score is greater than a first score, a new subtask is created according to the semantic label and added to the target execution plan; If the association type contains a specified type, a generated subtask is generated based on the specified type and added to the target execution plan; If the impact degree is greater than a second score, a target subtask is determined according to the scoring basis of the impact degree and removed from the target execution plan.

7. The method of claim 1, wherein, The target execution plan also includes a task type corresponding to each subtask, and the target execution plan is generated according to the to-be-processed query, including: Fuse the to-be-processed query with the decomposition template to generate a decomposition prompt word, the decomposition prompt word including a decomposition requirement; Use the decomposition prompt word to guide the large language model to perform decomposition processing on the to-be-processed query according to the decomposition requirement to generate a preliminary execution plan; If the format of the preliminary execution plan does not meet the preset format condition, a new preliminary execution plan is generated until a preset number of times is reached or the preset format condition is met; If the format of the preliminary execution plan meets the preset format condition, the preliminary execution plan is logically verified according to the dependent tasks and task types of the subtasks; If the preliminary execution plan passes the logical verification, the preliminary execution plan is determined as the target execution plan.

8. An agent-based report generation method, characterized by, Applied to a work intelligent agent, the method includes: Obtain a specified subtask from a task bus and execute it to obtain a task execution result, the specified subtask being determined by an arrangement intelligent agent from a target execution plan, the target execution plan being generated based on a to-be-processed query; Perform content integration processing on the task execution result to obtain a result content; Perform quality evaluation and content extraction processing on the task execution result through a large language model to generate metadata; Classify the logical relationship between the task execution result and other result objects to obtain object association data of the task execution result; The result content, the metadata, and the object association data are combined into a specified result object, and the specified result object is published on a task bus so that an intelligent agent generates a report corresponding to the to-be-processed query based on the specified result object on the task bus.

9. The method of claim 8, wherein, The object association data includes an association object and an association type, the logical relationship between the task execution result and other result objects is classified and processed, and the object association data of the task execution result is obtained, including: Based on the semantic similarity between the task execution result and the result objects in the object library, a candidate result object is determined; The task execution result is combined with each candidate result object to obtain a plurality of result pairs; For each result pair, the summary information of the result pair is combined with a classification template to obtain a classification prompt word, and the classification prompt word includes definitions of a plurality of preset relationship types; The classification prompt word is used to guide a large language model to determine the association type of the result pair based on the definition of the preset relationship type; The candidate result object in the result pair with the specified type of association type is determined as the association object of the task execution result.

10. The method of claim 8, wherein, The result content, the metadata, and the object association data are combined into a specified result object, and the specified result object is published on a task bus so that an intelligent agent generates a report corresponding to the to-be-processed query based on the specified result object on the task bus. The object association data, the confidence, the semantic label, and the result content are assembled to generate a specified result object; The specified result object is written into the object library; The task state of the specified subtask is updated to completed, and the completion time is recorded; If the specified result object is successfully written into the object library and the task state is successfully updated, the specified result object is published on the task bus.

Images