A method for intelligent generation of multimodal team debriefing content based on a large language model
By acquiring and parsing multimodal collaboration data in R&D projects, identifying changes in problem status and matching task execution and version change records, generating problem handling process trajectories, and using large language models to generate team debriefing content, the problem of lack of consistency and traceability in debriefing content in existing technologies is solved, and more accurate debriefing results are achieved.
Patent Information
- Application Number
- CN202611259049.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-19
- Publication Date
- 2026-11-17
AI Technical Summary
In the research and development process of existing technologies, the scattered storage of multi-source heterogeneous collaborative data makes it difficult for the review content to accurately reflect the problem handling path. It lacks the correlation identification of problem status changes, task execution process and version modification process, and the generated review content lacks consistency and traceability.
By acquiring multimodal collaboration data from R&D projects, performing data analysis and time synchronization, identifying nodes where problem status changes occur, and matching these nodes with task execution, version changes, and test verification records, problem handling correlation records are generated, a problem handling process trajectory is constructed, and team debriefing content is generated using a large language model.
It achieves a correspondence between the review content and the actual problem-solving process, solving the problem that the review content in the existing technology lacks operational path support and is difficult to trace specific R&D records. The generated review content is more consistent and traceable.
Smart Images

Figure CN123118014A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, and in particular to a method for intelligent generation of multimodal team debriefing content based on a large language model. Background Technology
[0002] In the R&D process of software systems, smart devices, and industrial equipment, collaboration among multiple roles such as testing and development often generates heterogeneous data from multiple sources, including issue status, task assignments, instant messaging, version changes, and testing verification. Because different roles use different tools, the data is stored in a scattered manner, making it difficult to form a complete processing record. Existing debriefing methods mostly rely on manual compilation or large model text summarization, which cannot accurately identify the correspondence between issue status changes and specific version modifications, testing verifications, and other R&D operations. This makes it difficult to automatically trace the operation path, and the generated debriefing content lacks the connection between discovery, modification, and verification stages, failing to truly recreate the closed-loop processing trajectory of issues. Therefore, how to synchronize and correlate multi-source collaborative data, and reconstruct the issue processing process based on status changes to constrain the generation of accurate debriefing content from large language models, has become an urgent technical problem to be solved.
[0003] Chinese Patent Application Publication No. CN122364440A discloses a human-machine collaborative intelligent agent decision optimization method based on a large model and a review knowledge base. The method includes: constructing a real-time human-machine collaborative system based on a hierarchical language intelligent agent; the system includes: a multi-dimensional experience accumulation module, an instruction verification intermediate layer, an offline reinforcement learning policy evaluation network, a slow-brain module, and a fast-brain module; based on a game environment, the system collects multi-modal data of historical human-machine interactions in real time through a background monitoring script, and converts the multi-modal data into standardized JSON structure data through a parser; based on the standardized JSON structure data, multi-dimensional collaborative features are extracted; the multi-dimensional collaborative features are used as input corpus to a pre-trained large language model, generating structured review experiences based on preset prompts; the generated structured review experiences are input into the multi-dimensional experience accumulation module and stored in the large language model. The following describes the process in the protocol stack database: Based on the current real-time collaborative task, the SlowBrain module, using the current game map identifiers, extracts matching replay experiences from the context protocol stack database using retrieval enhancement generation technology, and injects them into the inference context to complete high-order global goal planning, generating a high-order sub-task sequence. The high-order sub-task sequence and the natural language commands issued by the user are input into the command verification intermediate layer. An offline reinforcement learning strategy is used to evaluate the natural language commands, scoring their utility probabilities. The calculated utility probabilities are compared with a preset dynamic admission threshold, and the comparison result is output. Based on the comparison result, it is determined whether the natural language command is an efficient collaborative command. If the natural language command is an efficient collaborative command, the FastBrain module converts the natural language command into low-level atomic operation commands, executes them in the game environment, and outputs the execution result.
[0004] Therefore, it is evident that the existing technology is mainly designed for experience retrieval and decision optimization processes in specific task scenarios, and cannot be applied to the closed-loop review of technical problems in R&D projects. Specifically, it suffers from the following problems: First, existing review content generation methods based on large language models typically rely on input text or pre-organized experience data, lacking the identification of the relationships between problem status changes, task execution processes, version modification processes, and testing and verification processes, resulting in the generated review content failing to accurately correspond to the actual problem handling path. Second, existing methods typically rely on keywords or simple indexing to retrieve historical records, failing to automatically match related data scattered across different R&D systems based on the state transition relationships of technical problems during the handling process, leading to information gaps between problem discovery, problem handling, and problem verification processes. Third, the existing review generation process lacks a consistency verification mechanism between the generated content and actual R&D operation records. When input data is incomplete or the relationships are inaccurate, it is difficult to locate the corresponding problem handling node and correct the content in a timely manner, affecting the authenticity and traceability of the review results. Summary of the Invention
[0005] To address this, the present invention provides a method for intelligent generation of multimodal team debriefing content based on a large language model. This method overcomes the problem of closed-loop processing in existing technologies where debriefing content is difficult to reconstruct due to the dispersed storage of multi-source heterogeneous collaborative data. This is achieved through multimodal data analysis and state association.
[0006] To achieve the above objectives, this invention provides a method for intelligently generating multimodal team debriefing content based on a large language model, comprising: Acquire multimodal collaboration data generated during the closed-loop processing of technical issues in R&D projects, and perform data parsing and time synchronization on the multimodal collaboration data to obtain a problem processing dataset. The multimodal collaboration data includes problem status records, task assignment records, instant messaging records, version change records, and test verification records. Based on the problem processing dataset, determine the status change information of the technical problem during the processing, and generate a set of problem status change nodes based on the status change information; Based on the set of problem status change nodes, the corresponding task execution records, version change records, and test verification records are matched and associated to generate problem handling association records; Based on the aforementioned problem handling association records, the problem handling process trajectory is determined according to the sequence of status changes of the technical problem sheet; Based on the problem handling process trajectory, extract the corresponding problem description information, processing operation information, and verification result information, and generate review context data; The debriefing context data is input into a preset large language model to obtain the team debriefing content of the technical problem handling process.
[0007] Furthermore, the process of determining the problem-solving process dataset includes: Extract the timestamp information from the issue status record, the task assignment record, the instant messaging record, the version change record, and the test verification record; Sort records from different data sources by time based on the timestamp information; Establish corresponding data field mapping relationships based on the data types of each data source, and convert records from different data sources into data records in a unified format. The data records, after being sorted by time and converted to the correct format, are stored to obtain the problem handling process dataset.
[0008] Furthermore, the process of identifying the status change information of the technical problem form during the processing based on the problem processing process dataset includes: Read the problem status field from the problem handling process dataset; The sequence of consecutive status records is determined based on the problem order number in the problem handling process dataset. Based on the comparison result of the single state field of two adjacent records in the continuous state record sequence, determine the state transition record in which the state has changed; The state change information of the technical problem sheet during the processing is determined based on all the state transition records.
[0009] Furthermore, the process of generating a set of problem state change nodes based on state change information includes: The state transition direction of the problem form is determined based on the previous and next problem form status fields corresponding to each state transition record. Extract the corresponding operator, operation time, and associated task number from the state transition record, and extract the corresponding processing operation record from the state transition direction; The state transition direction, the operator, the operation time, the associated task number, and the processing operation record are bound and encapsulated with the state transition record to generate a corresponding problem state change node. Based on all the problem status change nodes described in the same problem statement, a set of problem status change nodes is obtained.
[0010] Furthermore, the process of matching and associating the corresponding task execution records, version change records, and test verification records to generate issue handling association records includes: Obtain the corresponding task execution record based on the associated task number in the problem status change node; Obtain the corresponding version change record based on the task execution record; Obtain the corresponding test task number based on the version number in the version change record; The test verification record is obtained based on the test task number, and the task execution record, version change record, and test verification record are combined to generate an issue handling association record.
[0011] Furthermore, the process of obtaining the corresponding version change record includes: Retrieve commit records and extract the committer, commit time, version identifier, and modified file information from the commit records; The version identifier is compared and matched with the task number of the task execution record corresponding to the problem status change node to obtain the comparison result; Based on the comparison results, the submission record corresponding to the version identifier is determined as the version change record corresponding to the technical issue form.
[0012] Furthermore, the process of generating the problem handling process trajectory according to the state change sequence of the technical problem sheet includes: The nodes representing the changes in the problem's state are sorted according to their corresponding state transition times. Based on the state transition direction of the sorted problem state change nodes, the link relationship between adjacent nodes is established according to the comparison result of the problem single state field of two adjacent nodes, and the node links connected in the order of state change are obtained. Add the problem processing association records corresponding to each problem status change node in the node link to the corresponding node to obtain the problem processing process trajectory.
[0013] Furthermore, the process of extracting corresponding problem description information, processing operation information, and verification result information based on the problem handling process trajectory, and generating review context data, includes: Based on the node links of the problem handling process trajectory, obtain the problem proposal record of the first node in the node link where the problem status changes; Obtain the test verification record of the last node in the node link where the problem status changed; Obtain the task execution record and version change record of the problem status change node located between the first and last nodes; Based on the problem-raising records, task execution records, version change records, and test verification records, a review context data is generated.
[0014] Furthermore, the process of generating the debriefing context data includes: Based on the node order in the problem handling process trajectory, the problem proposal record, the task execution record, the version change record, and the test verification record are sequentially concatenated; Insert the corresponding issue number, task number, version number, and test task number between each record to establish explicit relationships between different records; The assembled and numbered record sequence is assembled into structured text data, generating retrospective context data that includes the process of problem generation, modification, and verification.
[0015] Furthermore, after inputting the debriefing context data into the large language model to generate team debriefing content, the process also includes: Obtain the user's modification marks on the team debriefing content; Locate the corresponding problem state change node based on the operation location information of the modified marker; Update the problem processing association record of the problem status change node according to the modification intent information of the modification mark; Based on the updated problem handling association records, the debriefing context data is regenerated, and the updated team debriefing content is generated by inputting the preset large language model.
[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: In the closed-loop processing of technical problems in R&D projects, this invention addresses the problem of scattered storage of multi-source heterogeneous data and the difficulty in automatically corresponding problem status changes with specific R&D operations during the R&D collaboration process. It collects multimodal collaborative data such as status records, task execution records, version change records, and test verification records during the R&D problem sheet processing process, and performs time synchronization and data parsing on different data sources. Furthermore, it identifies problem processing nodes based on problem sheet status changes and matches corresponding task execution, version modification, and test verification data based on these status change nodes, forming a debriefing context data reflecting the technical problem processing process. Finally, it uses the debriefing context data to constrain a large language model to generate team debriefing content, ensuring that the generated results correspond to the actual problem processing process, rather than merely summarizing based on text semantics. This solves the shortcomings of existing text generation technologies, which lack support from actual operational paths and struggle to trace specific R&D records. Attached Figure Description
[0017] Figure 1 This is a flowchart of the intelligent generation method for multimodal team debriefing content based on a large language model in this embodiment; Figure 2 This embodiment defines a flowchart for determining the dataset of the problem handling process; Figure 3 This is a flowchart illustrating the process of identifying state change information in this embodiment; Figure 4 This is a flowchart for generating a set of problem state change nodes for this embodiment. Detailed Implementation
[0018] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.
[0019] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.
[0020] Please see Figure 1 As shown, this is a flowchart of the intelligent generation method for multimodal team debriefing content based on a large language model in this embodiment. The method in this embodiment includes: Step S1: Obtain the multimodal collaboration data generated during the closed-loop processing of technical issues in the R&D project, and perform data parsing and time synchronization on the multimodal collaboration data to obtain the issue processing dataset. The multimodal collaboration data includes issue status records, task assignment records, instant messaging records, version change records, and test verification records.
[0021] In this embodiment, the technical issue form refers to an electronic document used in R&D projects to record, track, and manage technical issues in a closed loop. It is created and entered by testers or users in the R&D project after discovering a technical issue through the issue management system.
[0022] In this embodiment, the issue status record and task assignment record are obtained by directly reading from the database table of the project management platform used by the R&D project. The issue status record includes at least the issue number, status field, operator, operation time and associated task number. The status field includes at least the status fields of pending assignment, under development, pending testing, pending acceptance and closed. The task assignment record includes at least the task number, assigner, recipient, assignment time and task description.
[0023] In this embodiment, the instant messaging records are obtained by calling the application programming interface of the instant messaging tool used in the R&D project or by reading from a local log file. The instant messaging records include at least the message sender, the message receiver, the sending time, and the message content. In this embodiment, the version change record is obtained by calling the application programming interface of the version management system used by the R&D project or by reading from the commit log. The version change record includes at least the commit number, the committer, the commit time, the version identifier, and the list of modified files. In this embodiment, the test verification record is obtained by calling the application interface of the test management platform or reading the result file output by the automated testing tool. The test verification record includes at least the test task number, tester, test time, test case execution result and issue number.
[0024] Step S2: Determine the status change information of the technical problem sheet during the processing based on the problem processing process dataset, and generate a set of problem status change nodes based on the status change information.
[0025] Step S3: Based on the set of problem status change nodes, match and associate the corresponding task execution records, version change records, and test verification records to generate problem handling association records.
[0026] Step S4: Based on the problem handling association records, determine the problem handling process trajectory according to the sequence of status changes of the technical problem sheet.
[0027] Step S5: Extract the corresponding problem description information, processing operation information, and verification result information based on the problem handling process trajectory, and generate review context data.
[0028] Step S6: Input the debriefing context data into the preset large language model to obtain the team debriefing content of the technical issue handling process.
[0029] In this embodiment, the pre-trained language model based on the Transformer architecture is selected as the preset large language model. The structured input sequence is created by arranging the records of each node in the debriefing context data according to the chronological order of state changes. Each node record includes problem-posing records, task execution records, version change records, and test verification records. Each record in the structured input sequence contains a record type label, a timestamp, an association number, and a record content summary. The record type label corresponds to four types: problem-posing, task execution, version change, and test verification. The association number includes the problem number, task number, version number, and test task number. Then, the structured input sequence is processed according to preset prompt words. The template organization serves as the input text for the large language model. The preset prompt template includes role setting instructions, content organization instructions, and output format instructions. The role setting instructions require the model to describe the process as a reviewer of the development process. The content organization instructions require the model to describe the process item by item according to the record order in the structured input sequence. The output format instructions require the model to organize the output content according to the problem discovery phase, problem handling phase, and problem verification phase. Finally, the organized input text is input into the preset large language model, which then generates a team debriefing describing the entire process of a technical problem from discovery to closure, according to the record order in the structured input sequence and the format requirements of the preset prompt template.
[0030] Specifically, in the closed-loop processing of technical issues in R&D projects, this invention addresses the problem of scattered storage of multi-source heterogeneous data and the difficulty in automatically corresponding problem status changes with specific R&D operations during R&D collaboration. It collects multimodal collaborative data, including status records, task execution records, version change records, and test verification records, during the problem issue processing process, and performs time synchronization and data parsing on different data sources. Furthermore, it identifies problem processing nodes based on problem status changes and matches corresponding task execution, version modification, and test verification data based on these status change nodes, forming a debriefing context data reflecting the technical problem processing process. Finally, it uses the debriefing context data to constrain a large language model to generate team debriefing content, ensuring that the generated results correspond to the actual problem processing process, rather than merely summarizing based on textual semantics. This solves the shortcomings of existing text generation technologies, which lack support from actual operational paths and struggle to trace specific R&D records.
[0031] Please see Figure 2 As shown, this is a flowchart for determining the problem handling process dataset in this embodiment. In this embodiment, the process of determining the problem handling process dataset includes: Extract timestamp information from issue status records, task assignment records, instant messaging records, version change records, and test verification records; Records from different data sources are sorted by time based on timestamp information, so that records from different data sources are arranged on the same timeline in chronological order of occurrence. Establish corresponding data field mapping relationships based on the data types of each data source, and convert records from different data sources into data records of a unified format; The data records, after being sorted by time and converted to the correct format, are stored to obtain the problem handling process dataset.
[0032] Specifically, due to the dispersed and heterogeneous nature of multi-source collaborative data, which makes it difficult to correlate them, the timestamp information of each data record is extracted, and records from different data sources are aligned and arranged on the same timeline to eliminate temporal disorder. Furthermore, a data field mapping relationship is established, and heterogeneous records containing problem status fields and version identifiers are uniformly converted into standard format data records and stored. Through a data processing mechanism of time synchronization and structure normalization, a problem handling process dataset with coherent time sequence and uniform format is constructed, providing underlying data support for accurately matching the corresponding task execution records and test verification records based on the status changes of problem orders.
[0033] Please see Figure 3 As shown, this is a flowchart for identifying state change information in this embodiment. In this embodiment, the process of identifying the state change information of the technical problem sheet during the processing based on the problem processing process dataset includes: Read the issue status field from the issue handling process dataset; Based on the problem order number carried in the data records of the problem handling process dataset, the data records are grouped and arranged in the order of timestamp information within each group to obtain a sequence of continuous status records corresponding to the same problem order; In a continuous state record sequence, the contents of the problem single state field in two adjacent records are compared in turn. Based on the comparison result that the contents of the state fields of adjacent records are different, the next record is determined as the state transition record in which the state has changed, so as to effectively eliminate redundant records in which the state has not changed substantially. By summarizing all status transition records, we can obtain information on the status changes of the technical issue form during the processing.
[0034] Please see Figure 4 As shown, this is a flowchart of generating a set of problem state change nodes in this embodiment. In this embodiment, the process of generating a set of problem state change nodes based on state change information includes: Retrieve the status fields of the previous and next problem slips corresponding to each status transition record; The status fields of the previous and next problem orders are combined to construct the status transition direction of the problem order; Extract the corresponding operator, operation time, and associated task number from the state transition record, and extract the corresponding processing operation record from the problem handling process dataset according to the state transition direction; Bind and encapsulate the state transition direction, operator, operation time, associated task number, and processing operation record with the state transition record to generate the corresponding problem state change node; Summarize all the status change nodes of the same issue to obtain the set of issue status change nodes.
[0035] Specifically, the problem order is grouped by number and arranged chronologically. Status transition records are extracted by comparing adjacent status fields. Since simple status transition records cannot directly reflect specific R&D execution actions, the previous and subsequent statuses are combined to construct a status transition direction. This direction is then used as a retrieval criterion to extract the corresponding operators, associated task numbers, and actual processing operation records from the dataset. These are then bound to the status transition direction and encapsulated as problem status change nodes. Noise is removed through chronological comparison, and multi-source heterogeneous operation data is aggregated using the status transition direction as the anchor point. This establishes a precise mapping relationship between problem order status changes and specific task execution records.
[0036] In this embodiment, the process of matching and associating the corresponding task execution records, version change records, and test verification records to generate problem handling association records includes: Retrieve the corresponding task execution record based on the associated task number in the problem status change node; Retrieve the corresponding version change record based on the task execution record; Obtain the corresponding test task number based on the version number in the version change log; Retrieve test verification records based on test task numbers, and combine task execution records, version change records, and test verification records to generate issue handling association records.
[0037] In this embodiment, the process of obtaining the corresponding version change record includes: Read commit records from the version control system; Extract the submitter, submission time, version identifier, and modified file information from the commit history; Compare and match the version identifier with the task number of the task execution record corresponding to the issue status change node; When the comparison results match, the submission record corresponding to the version identifier will be identified as the version change record corresponding to the technical issue ticket.
[0038] Specifically, the task execution record is obtained starting from the task number in the problem status change node. Since the code submission record and business task are independent of each other, the code submission is accurately anchored to the specific task by comparing and matching the version identifier with the task number. Then, the test verification record is obtained by tracing down along the version number and combined to generate associated records. By comparing and chaining the identifiers across systems, the data silos are broken down. At the data level, the precise mapping relationship between problem status changes and specific R&D operations and verification results is established, which solves the problem that it is difficult to automatically correspond and trace status changes with specific operations.
[0039] In this embodiment, the process of generating the problem handling process trajectory according to the state change sequence of the technical problem sheet includes: Extract the state transition time corresponding to each problem state change node from the set of problem state change nodes, and sort all problem state change nodes according to the order of timestamps; Based on the state transition direction of the sorted problem state change nodes, determine whether the state field of the next problem in the previous node is consistent with the state field of the previous problem in the next node. If they are consistent, establish the link between the adjacent nodes and obtain the node link connected in the order of state change. Add the problem handling association record corresponding to each problem status change node in the node link to the corresponding node to obtain the problem handling process trajectory.
[0040] Specifically, the basic timing sequence is established by sorting nodes according to their state transition times. Since discrete nodes cannot directly reflect the causal relationship of state transitions, the link relationship is established by comparing whether the state fields of adjacent nodes are consistent. This ensures that the linked node links conform to the state machine logic and are uninterrupted. Problem processing records containing code submissions and test results are attached to the corresponding nodes. Through the progressive mechanism of timing arrangement and state field verification, a problem processing process trajectory that is temporally coherent and contains specific R&D operation details is reconstructed.
[0041] In this embodiment, the process of extracting corresponding problem description information, processing operation information, and verification result information based on the problem handling process trajectory, and generating review context data includes: Based on the node links of the problem handling process trajectory, obtain the problem submission record of the first node in the node link where the problem status changes; Obtain the test verification record of the last problematic node in the node chain; Retrieve task execution records and version change records for problem status change nodes located between the first and last nodes; Generate review context data based on issue proposing records, task execution records, version change records, and test verification records.
[0042] In this embodiment, the process of generating review context data includes: The problem-solving process is organized in the order of the nodes, including the problem-raising record, task execution record, version change record, and test verification record. Insert the corresponding issue number, task number, version number, and test task number between each record to establish explicit relationships between different records; The assembled and numbered record sequence is assembled into structured text data, generating retrospective context data that includes the process of problem generation, modification, and verification.
[0043] Specifically, corresponding records are extracted based on the beginning, end, and middle positions of the node link to ensure accurate temporal positioning of various operation information. Considering that discrete record fragments are difficult to directly reflect the causal relationship of the closed-loop processing of the problem, various records are further spliced sequentially according to the node order, and corresponding numbers are inserted between each record. By interlacing the numbers, explicit associations between different data sources are established, and the originally scattered heterogeneous records are reconstructed into review context data containing the entire process of problem generation, modification, and verification. Thus, at the data input level, a structured text sequence with strong logical constraints is provided for the large language model, overcoming the problem that the review content generated by the existing technology lacks operational associations and makes it difficult to restore the processing path.
[0044] In this embodiment, after inputting the debriefing context data into the large language model to generate team debriefing content, the method further includes: Obtain user-defined modification tags for team debriefing content; Based on the operation location information of the modified marker, the corresponding problem status change node is retrieved and located in the problem handling process trajectory; Based on the modification intent information of the modification marker, update the content of the task execution record, version change record or test verification record in the problem handling association record corresponding to the problem status change node; Based on the updated issue handling and related records, the debriefing context data is regenerated, and the updated team debriefing content is generated by inputting a preset large language model.
[0045] In this embodiment, the modification marker is a structured instruction data automatically generated by the system when a user performs selection, annotation, or content replacement operations on the team debriefing content displayed on the front end. The modification marker includes at least operation location information and modification intent information. The operation location information includes the issue number and the corresponding status node identifier, which is used to accurately map the modification operation to a specific node in the issue handling process trajectory. The modification intent information includes the deletion, replacement, or supplementation of content in the original text, so that the system can update the associated records of the corresponding node accordingly.
[0046] Specifically, by acquiring the modification markers submitted by users on the front end, which include operation location information and modification intent information, and using them as the triggering condition for data correction, the text modification actions displayed on the front end are mapped to the underlying structured data records for source tracing and correction. Based on the updated associated records, the review context data is regenerated to drive the large language model to generate updated content. This constructs a closed-loop verification mechanism from result feedback to underlying data correction and then to content regeneration, ensuring the consistency between the context data input to the model and the actual R&D operations.
[0047] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for intelligently generating multimodal team debriefing content based on a large language model, characterized in that, include: Acquire multimodal collaboration data generated during the closed-loop processing of technical issues in R&D projects, and perform data parsing and time synchronization on the multimodal collaboration data to obtain a problem processing dataset. The multimodal collaboration data includes problem status records, task assignment records, instant messaging records, version change records, and test verification records. Based on the problem processing dataset, determine the status change information of the technical problem during the processing, and generate a set of problem status change nodes based on the status change information; Based on the set of problem status change nodes, the corresponding task execution records, version change records, and test verification records are matched and associated to generate problem handling association records; Based on the aforementioned problem handling association records, the problem handling process trajectory is determined according to the sequence of status changes of the technical problem sheet; Based on the problem handling process trajectory, extract the corresponding problem description information, processing operation information, and verification result information, and generate review context data; The debriefing context data is input into a preset large language model to obtain the team debriefing content of the technical problem handling process.
2. The intelligent generation method for multimodal team debriefing content based on a large language model according to claim 1, characterized in that, The process of determining the dataset for the problem handling process includes: Extract the timestamp information from the issue status record, the task assignment record, the instant messaging record, the version change record, and the test verification record; Sort records from different data sources by time based on the timestamp information; Establish corresponding data field mapping relationships based on the data types of each data source, and convert records from different data sources into data records in a unified format. The data records, after being sorted by time and converted to the correct format, are stored to obtain the problem handling process dataset.
3. The intelligent generation method for multimodal team debriefing content based on a large language model according to claim 2, characterized in that, The process of identifying the status change information of the technical problem form during the processing based on the problem processing process dataset includes: Read the problem status field from the problem handling process dataset; The sequence of consecutive status records is determined based on the problem order number in the problem handling process dataset. Based on the comparison result of the single state field of two adjacent records in the continuous state record sequence, determine the state transition record in which the state has changed; The state change information of the technical problem sheet during the processing is determined based on all the state transition records.
4. The intelligent generation method for multimodal team debriefing content based on a large language model according to claim 3, characterized in that, The process of generating a set of problem state change nodes based on state change information includes: The state transition direction of the problem form is determined based on the previous and next problem form status fields corresponding to each state transition record. Extract the corresponding operator, operation time, and associated task number from the state transition record, and extract the corresponding processing operation record from the state transition direction; The state transition direction, the operator, the operation time, the associated task number, and the processing operation record are bound and encapsulated with the state transition record to generate a corresponding problem state change node. Based on all the problem status change nodes described in the same problem statement, a set of problem status change nodes is obtained.
5. The intelligent generation method for multimodal team debriefing content based on a large language model according to claim 4, characterized in that, The process of matching and associating the corresponding task execution records, version change records, and test verification records to generate problem handling association records includes: Obtain the corresponding task execution record based on the associated task number in the problem status change node; Obtain the corresponding version change record based on the task execution record; Obtain the corresponding test task number based on the version number in the version change record; The test verification record is obtained based on the test task number, and the task execution record, version change record, and test verification record are combined to generate an issue handling association record.
6. The intelligent generation method for multimodal team debriefing content based on a large language model according to claim 5, characterized in that, The process of obtaining the corresponding version change record includes: Retrieve commit records and extract the committer, commit time, version identifier, and modified file information from the commit records; The version identifier is compared and matched with the task number of the task execution record corresponding to the problem status change node to obtain the comparison result; Based on the comparison results, the submission record corresponding to the version identifier is determined as the version change record corresponding to the technical issue form.
7. The intelligent generation method for multimodal team debriefing content based on a large language model according to claim 6, characterized in that, The process of generating the problem handling process trajectory according to the state change sequence of the technical problem sheet includes: The nodes representing the changes in the problem's state are sorted according to their corresponding state transition times. Based on the state transition direction of the sorted problem state change nodes, the link relationship between adjacent nodes is established according to the comparison result of the problem single state field of two adjacent nodes, and the node links connected in the order of state change are obtained. Add the problem processing association records corresponding to each problem status change node in the node link to the corresponding node to obtain the problem processing process trajectory.
8. The intelligent generation method for multimodal team debriefing content based on a large language model according to claim 7, characterized in that, The process of extracting corresponding problem description information, processing operation information, and verification result information based on the problem handling process trajectory, and generating review context data includes: Based on the node links of the problem handling process trajectory, obtain the problem proposal record of the first node in the node link where the problem status changes; Obtain the test verification record of the last node in the node link where the problem status changed; Obtain the task execution record and version change record of the problem status change node located between the first and last nodes; Based on the problem-raising records, task execution records, version change records, and test verification records, a review context data is generated.
9. The intelligent generation method for multimodal team debriefing content based on a large language model according to claim 8, characterized in that, The process of generating debriefing context data includes: Based on the node order in the problem handling process trajectory, the problem proposal record, the task execution record, the version change record, and the test verification record are sequentially concatenated; Insert the corresponding issue number, task number, version number, and test task number between each record to establish explicit relationships between different records; The assembled and numbered record sequence is assembled into structured text data, generating retrospective context data that includes the process of problem generation, modification, and verification.
10. The intelligent generation method for multimodal team debriefing content based on a large language model according to claim 9, characterized in that, After inputting the debriefing context data into the large language model to generate team debriefing content, the process also includes: Obtain the user's modification marks on the team debriefing content; Locate the corresponding problem state change node based on the operation location information of the modified marker; Update the problem processing association record of the problem status change node according to the modification intent information of the modification mark; Based on the updated problem handling association records, the debriefing context data is regenerated, and the updated team debriefing content is generated by inputting the preset large language model.
Citation Information
Patent Citations
Man-machine collaborative agent decision optimization method based on large model and redisk knowledge base
CN122364440A
