Multi-satellite type knowledge retrieval and generation method based on multi-agent cooperation

Abstract

The application provides a multi-satellite model knowledge retrieval and generation method based on multi-agent cooperation, which can be applied to the field of artificial intelligence technology. The method comprises the following steps: in response to a text query event containing query text and at least one satellite model to be queried, generating a retrieval strategy by a knowledge task planning agent, the retrieval strategy comprising retrieval instructions and a retrieval database range; based on the retrieval instructions, performing retrieval in at least one knowledge database to obtain a candidate knowledge text set; calling a knowledge result fusion agent to filter a target knowledge text set according to the correlation degree of each candidate knowledge text with the query text and the satellite model to be queried; calling a knowledge generation verification agent to check the overall confidence of the target knowledge text set; in the case that the check fails, triggering the knowledge task planning agent to regenerate the retrieval strategy and execute the above method until the check passes or the execution round number of the method reaches a preset round number, and generating an analysis result.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and satellite data processing technology, specifically to a method for knowledge retrieval and generation of multiple satellite models based on multi-agent collaboration. Background Technology

[0002] With the development of aerospace technology, data generated by different types of satellites differ in format, semantics, and processing procedures, forming a multi-source heterogeneous data environment. Achieving efficient knowledge retrieval and question answering from massive amounts of heterogeneous satellite data is a significant challenge for aerospace data management.

[0003] Retrieval-Augmented Generation (RAG) technology improves the accuracy and credibility of responses by integrating retrieval results from external knowledge bases into the generation process of large language models. RAG technology is evolving towards Agentic RAG, whose core idea is to combine the autonomous decision-making capabilities of agents with the traditional retrieval-generation process, enabling the system to autonomously plan retrieval strategies, dynamically adjust retrieval paths, and iteratively optimize retrieval results.

[0004] However, applying intelligent agent-based retrieval enhancement generation technology to the field of multi-satellite model data processing still faces the following technical challenges: First, different satellite models may use different terms and expressions for the same physical concept. For example, "resolution" may be expressed as "spatial resolution" or "pixel resolution," and "orbit altitude" may be expressed as "orbit elevation" or "operational orbit parameters." Traditional retrieval schemes struggle to effectively handle cross-model semantic heterogeneity. Second, satellite data queries involve various scenarios such as single-model queries, cross-model comparison queries, and multi-parameter joint retrieval. Existing schemes employ fixed pipeline-style calling processes, which cannot autonomously adjust retrieval strategies based on query complexity and the number of models involved. Third, satellite models are continuously updated, and new knowledge is constantly generated. However, in existing schemes, knowledge updates rely on manual maintenance or full reconstruction, resulting in long update cycles. The update process is independent of retrieval optimization, making it difficult for the updated knowledge to be utilized by the retrieval system in a timely manner, and there is a lack of automated conflict resolution mechanisms. Summary of the Invention

[0005] In view of the above problems, this application provides a multi-agent collaborative method for knowledge retrieval and generation of multi-satellite models to improve the accuracy of knowledge retrieval and the timeliness of knowledge updates for heterogeneous data of multiple satellite models.

[0006] According to an embodiment of this application, a method for knowledge retrieval and generation of multiple satellite models based on multi-agent collaboration is provided. The method includes: responding to a query request containing query text and at least one satellite model to be queried, generating a retrieval strategy through a knowledge task planning agent, the retrieval strategy including a retrieval instruction and a retrieval database range, the retrieval instruction being used to retrieve knowledge text related to the query text and the at least one satellite model to be queried within the retrieval database range, the retrieval database range referring to at least one knowledge database corresponding to the at least one satellite model to be queried; based on the retrieval instruction, calling a knowledge document retrieval agent to perform a retrieval in the at least one knowledge database to obtain a candidate knowledge text set; calling a knowledge result fusion agent to filter a target knowledge text set according to the relevance of each candidate knowledge text in the candidate knowledge text set to the query text and the satellite model to be queried; calling a knowledge generation verification agent to verify the overall confidence of the target knowledge text set; if the verification fails, triggering the knowledge task planning agent to regenerate the retrieval strategy by generating reflection information through the knowledge generation verification agent, and re-executing the above method based on the regenerated retrieval strategy until the verification passes or the number of execution rounds of the method reaches a preset number of rounds, and generating analysis results based on the target knowledge text set.

[0007] According to an embodiment of this application, in response to a query request containing query text and at least one satellite model to be queried, generating a retrieval strategy based on a knowledge task planning agent includes: analyzing the semantic complexity of the query text and the number of the at least one satellite model to be queried through the knowledge task planning agent; when the at least one satellite model to be queried is single, generating a single-database retrieval strategy through the knowledge task planning agent, wherein the retrieval database scope is a single knowledge database corresponding to the satellite model to be queried, and the retrieval instruction is used to perform a retrieval in the single knowledge database; when the at least one satellite model to be queried is multiple, generating a multi-database parallel retrieval strategy through the knowledge task planning agent, wherein the retrieval database scope is multiple knowledge databases corresponding one-to-one with the at least one satellite model to be queried, and the retrieval instruction is used to perform a retrieval in parallel in the multiple knowledge databases; when the semantic complexity of the query text is higher than a first preset threshold or fails verification, generating a retrieval instruction for an iterative deepening retrieval strategy through the knowledge task planning agent, wherein the iterative deepening retrieval strategy includes the knowledge database corresponding to the at least one satellite model to be queried, and the retrieval instruction is used to perform multiple rounds of retrieval in conjunction with the query text.

[0008] According to an embodiment of this application, the step of invoking a knowledge document retrieval agent to perform a retrieval in at least one knowledge database based on the retrieval instruction to obtain a candidate knowledge text set includes: performing an approximate nearest neighbor retrieval of the query text in at least one knowledge database using the knowledge document retrieval agent to obtain an initial candidate text set and the semantic similarity between each initial candidate knowledge text in the initial candidate text set and the query text; matching the satellite model tags of each initial candidate knowledge text with the at least one query satellite model using the knowledge document retrieval agent to determine the relevance between each initial candidate knowledge text and the at least one query satellite model; determining a comprehensive score for each initial candidate knowledge text based on the semantic similarity and the relevance using the knowledge document retrieval agent; and selecting multiple initial candidate knowledge texts with the highest ranking from the initial candidate text set based on the comprehensive score to form the candidate knowledge text set.

[0009] According to an embodiment of this application, the step of matching the satellite model tags of each initial candidate knowledge text with the at least one satellite model to be queried through the knowledge document retrieval agent to determine the correlation between each initial candidate knowledge text and the at least one satellite model to be queried includes: calling the cross-model relationship graph through the knowledge document retrieval agent to query the association path between the satellite model tags of each initial candidate knowledge text and the at least one satellite model to be queried, wherein the cross-model relationship graph stores the association relationships between different satellite models in a graph structure; if the association path is found, the correlation between the initial candidate knowledge text and the at least one satellite model to be queried is determined to be a first value; if the association path is not found, the correlation between the initial candidate knowledge text and the at least one satellite model to be queried is determined to be a second value.

[0010] According to an embodiment of this application, the step of calling the knowledge result fusion agent to filter the target knowledge text set based on the relevance of each candidate knowledge text in the candidate knowledge text set to the query text and the satellite model to be queried includes: determining the fusion score of each candidate knowledge text based on the semantic similarity of each candidate knowledge text to the query text, the correlation between each candidate knowledge text and the at least one satellite model to be queried, and the preset credibility of the knowledge database to which each candidate knowledge text belongs; and selecting the top-ranked candidate knowledge texts as the target knowledge text set according to the descending order of the fusion scores.

[0011] According to an embodiment of this application, the method further includes: calculating the proportion of text segments whose fusion scores reach a second preset threshold in the target knowledge text set, thereby obtaining the overall confidence level. According to an embodiment of this application, the method further includes: in response to acquiring new knowledge data, retrieving multiple existing knowledge texts semantically adjacent to the new knowledge data from the knowledge database through a knowledge update agent, forming a neighborhood set; according to the relationship type between the new knowledge data and the existing knowledge texts, fusing the new knowledge data and the neighborhood set through the knowledge update agent to obtain fused new knowledge content and existing knowledge texts in the neighborhood set that are determined to be eliminated; writing the fused new knowledge content into the knowledge database corresponding to the satellite model, and deleting the existing knowledge texts that are determined to be eliminated.

[0012] According to an embodiment of this application, the method further includes: acquiring the new knowledge data uploaded by the user; and / or acquiring knowledge entry suggestion data generated by the knowledge generation and verification agent as the new knowledge data, wherein the knowledge entry suggestion data includes the query text, the at least one satellite model to be queried, and blank domain knowledge data, wherein the blank domain knowledge data is the knowledge lacking in the knowledge database determined by the knowledge generation and verification agent in the event of a failed verification, based on the difference between the knowledge content required to answer the query text and the knowledge content provided by the target knowledge text set.

[0013] According to an embodiment of this application, the knowledge database includes a general knowledge base and a model-specific knowledge base corresponding to each satellite model; the general knowledge base stores aerospace knowledge that is common across models, and the model-specific knowledge base stores domain knowledge corresponding to each satellite model.

[0014] The multi-satellite model knowledge retrieval and generation method based on multi-agent collaboration provided in this application is expected to achieve the following technical effects:

[0015] First, the knowledge document retrieval agent determines a comprehensive score based on semantic similarity and satellite model correlation during the retrieval process and selects candidate knowledge texts in order of ranking. This allows the retrieval results to incorporate cross-model correlation information on the basis of semantic matching, effectively solving the semantic omission problem caused by different expressions of the same concept among multiple satellite models, and significantly improving retrieval accuracy.

[0016] Second, the knowledge result fusion agent determines the fusion score and filters the target knowledge text set based on semantic similarity, relevance and preset credibility of the source knowledge database. It integrates multi-dimensional information to filter high-quality knowledge texts and combines the quantitative evaluation of overall confidence to provide a reliable knowledge foundation for subsequent generation and verification.

[0017] Third, the knowledge task planning agent, knowledge document retrieval agent, knowledge result fusion agent, and knowledge generation verification agent each have their own independent processing context. Each agent focuses on its own responsibility, avoiding the problem of information forgetting due to excessively long contexts for a single agent, and ensuring the processing quality of each decision-making link.

[0018] Fourth, the knowledge task planning agent autonomously decides the retrieval strategy based on the semantic complexity of the query text and the number of satellite models to be queried. It can flexibly select single-database retrieval, multi-database parallel retrieval, or iterative deepening retrieval strategy according to the query characteristics, so that simple queries can be responded to quickly with a lightweight process, and complex queries can be fully processed through the collaboration of multiple agents, and the average response time of compound queries is significantly shortened.

[0019] Fifth, the knowledge generation and verification agent performs confidence verification on the target knowledge text set. When the verification fails, it generates reflection information to trigger the knowledge task planning agent to regenerate the retrieval strategy, forming an autonomous optimization closed loop of "planning-retrieval-filtering-verification-reflection-replanning". This allows the retrieval strategy to be continuously iterated and optimized based on the verification results until the preset confidence level is reached, significantly improving the accuracy of cross-satellite model knowledge retrieval. Attached Figure Description

[0020] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0021] Figure 1 A flowchart illustrating a method for knowledge retrieval and generation of multiple satellite models based on multi-agent collaboration according to an embodiment of this application is shown in the schematic diagram.

[0022] Figure 2 This illustration schematically shows a system architecture diagram of a multi-satellite model knowledge retrieval and generation method based on multi-agent collaboration according to an embodiment of this application;

[0023] Figure 3 This illustration schematically shows a business logic diagram of a multi-satellite model knowledge retrieval and generation method based on multi-agent collaboration according to an embodiment of this application;

[0024] Figure 4 This illustration schematically shows a business logic diagram of a knowledge semantic association retrieval enhancement module according to an embodiment of this application;

[0025] Figure 5 This illustration schematically depicts the business logic of a knowledge-updating agent according to an embodiment of this application; and

[0026] Figure 6The diagram illustrates a system operation flowchart suitable for implementing a multi-satellite model knowledge retrieval and generation method based on multi-agent collaboration, according to an embodiment of this application. Detailed Implementation

[0027] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0028] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0029] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0030] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0031] Figure 1 The flowchart illustrates a method for knowledge retrieval and generation of multiple satellite models based on multi-agent collaboration according to an embodiment of this application.

[0032] like Figure 1 As shown, the multi-satellite model knowledge retrieval and generation method based on multi-agent collaboration in this application includes steps S110 to S150.

[0033] In step S110, in response to a query request containing query text and at least one satellite model to be queried, a retrieval strategy is generated by a knowledge task planning agent.

[0034] The retrieval strategy includes retrieval instructions and a retrieval database scope. The retrieval instructions are used to retrieve the query text and knowledge text related to at least one satellite model to be queried within the retrieval database scope. The retrieval database scope refers to at least one knowledge database corresponding to at least one satellite model to be queried.

[0035] In this embodiment, a knowledge task planning agent transforms the user-input query text into a structured retrieval strategy. Satellite data query scenarios are complex and diverse, ranging from simple parameter queries for a single satellite model to comparative analyses of multiple models. The knowledge task planning agent analyzes the semantic complexity of the query text and the number of satellite models involved, autonomously deciding on appropriate retrieval strategies, such as single-database retrieval, multi-database parallel retrieval, or iterative deepening retrieval strategies. This matches the optimal retrieval path for queries of varying complexity, avoiding over-processing of simple queries and under-processing of complex queries, and providing clear goals and scope guidance for subsequent retrieval execution.

[0036] In step S120, based on the retrieval instruction, the knowledge document retrieval agent is invoked to perform a retrieval in at least one knowledge database to obtain a candidate knowledge text set.

[0037] In this embodiment, based on the retrieval strategy formulated by the knowledge task planning agent, relevant knowledge text is accurately retrieved from the knowledge database corresponding to the satellite model to be queried. The knowledge database includes a general knowledge base and model-specific knowledge bases corresponding to each satellite model. The general knowledge base stores aerospace knowledge common to all satellite models, such as principles of orbital mechanics and basic theories of remote sensing. The model-specific knowledge base stores domain-specific knowledge for each satellite model, such as parameter configurations, processing procedures, and fault cases. Each model knowledge base is maintained independently, supporting refined management and independent updates of knowledge for each model. During retrieval, the knowledge document retrieval agent determines the scope of the target knowledge database based on the satellite model to be queried: if the query involves a single satellite model, the retrieval is performed in both the corresponding model-specific knowledge base and the general knowledge base; if the query involves multiple satellite models, the retrieval is performed in parallel in multiple corresponding model-specific knowledge bases and the general knowledge base.

[0038] It should be noted that in multi-satellite scenarios, different satellite models may use different terminology and expressions for the same physical concept. For example, for the concept of "resolution," technical documents for different models may use different expressions such as "spatial resolution," "pixel resolution," and "ground sampling distance." This cross-model semantic heterogeneity makes it difficult for retrieval methods that rely solely on keywords or general semantic matching to effectively recall semantically related knowledge texts from different model knowledge bases. Therefore, when performing a knowledge document retrieval agent, it is also necessary to take cross-model semantic association factors into account, incorporating the association information between knowledge text and the query model on the basis of semantic matching, thereby improving the recall rate of cross-model knowledge retrieval and enabling subsequent fusion and generation steps to obtain more comprehensive knowledge input.

[0039] In step S130, the knowledge result fusion agent is invoked to filter the target knowledge text set based on the relevance of each candidate knowledge text in the candidate knowledge text set to the query text and the satellite model to be queried.

[0040] The candidate knowledge texts returned by the knowledge document retrieval agent are comprehensively evaluated and selected. The degree of matching between each candidate knowledge text and the query requirements is measured from multiple dimensions. The knowledge text with the best evaluation results is selected as the knowledge basis for subsequent answer generation, thereby improving the quality of subsequent answer generation.

[0041] In step S140, the knowledge generation and verification agent is invoked to verify the overall confidence level of the target knowledge text set.

[0042] Before the large language model generates answers, the quality of the target knowledge text set is pre-validated. The knowledge generation and validation agent compares the overall confidence level with a preset confidence threshold. Only when the overall confidence level reaches the threshold is the target knowledge text set sent to the large language model to generate answers. This pre-validation mechanism effectively avoids the generation of low-quality answers due to insufficient retrieval quality, thereby improving the reliability and credibility of the final question-and-answer results.

[0043] In step S150, if the verification fails, the knowledge generation verification agent generates reflection information to trigger the knowledge task planning agent to regenerate the retrieval strategy, and the above method is re-executed based on the regenerated retrieval strategy (i.e., steps S120 to S140 are re-executed) until the verification passes or the number of execution rounds of the method reaches the preset number of rounds, and the analysis results are generated based on the target knowledge text set.

[0044] When the quality verification of the target knowledge text set fails, a closed-loop feedback mechanism is activated to iteratively optimize the retrieval strategy. The knowledge generation and verification agent generates reflection information for the failed verification cases and feeds it back to the knowledge task planning agent, triggering it to regenerate the retrieval strategy. Subsequently, the knowledge document retrieval agent and the knowledge result fusion agent re-execute the retrieval and filtering based on the new strategy, and perform confidence verification again. Through this autonomous optimization closed loop of "planning → retrieval → fusion → verification → reflection → replanning", the retrieval strategy can be continuously iterated and improved based on the verification results until the retrieval quality meets the preset requirements, thereby systematically improving the retrieval accuracy in complex query scenarios.

[0045] Through the above steps, the method of this application embodiment utilizes the autonomous decision-making and closed-loop collaboration capabilities of multiple agents to achieve adaptive adjustment of retrieval strategies, accurate matching of cross-model semantic associations, and iterative optimization of retrieval results, effectively improving the accuracy of knowledge retrieval and the quality of question answering in heterogeneous data scenarios of multiple satellite models.

[0046] Figure 2 The diagram illustrates a system architecture for implementing a multi-satellite model knowledge retrieval and generation method based on multi-agent collaboration according to an embodiment of this application.

[0047] like Figure 2 As shown, the system in this application embodiment adopts a three-layer architecture design, including a question-and-answer interface interaction layer, a knowledge base intelligent agent layer, and a knowledge data management layer.

[0048] The question-and-answer interface interaction layer is the system's user interaction layer. It responds to query requests containing query text and at least one satellite model to be queried, and returns the analysis results generated by the knowledge base intelligent agent layer to the user for visualization through the UI interface. The data processed by this layer comes from query requests actively entered by the user through the UI interface. The user enters query text in the graphical interface, such as "What is the procedure for adjusting the orbital parameters of satellite XX-01?" The system receives the query text, performs format standardization processing, and then passes it to the next layer.

[0049] The Q&A interface interaction layer includes a UI interface module, a query interaction module, and a results display module.

[0050] The UI module provides a graphical user interface, receives user-input queries, displays analysis results, and receives user-uploaded knowledge documents and user-initiated knowledge document deletion requests. Users can input query text through the UI's text input boxes, reuse existing queries from the historical query list, or quickly initiate queries using quick question templates. The UI displays input status and prompts in real time, guiding users to standardize their question format. Users can also submit knowledge documents for different satellite models (supporting PDF, Word, TXT, etc.) through the knowledge document upload portal. The UI automatically extracts satellite model information from the documents, selects the corresponding target model knowledge base, and passes the uploaded documents to the knowledge update agent for vectorization detection, agent inference fusion, and database processing. Furthermore, users can initiate deletion requests for knowledge documents in a specified model knowledge base. These requests are passed to the knowledge update agent, which performs corresponding deletion and knowledge maintenance operations in the knowledge base (including vector index cleanup, cross-model relationship graph updates, and version snapshot creation). During multi-turn dialogues, the UI module also manages the query history, context information, and user feedback records of the current session, maintaining session continuity.

[0051] The query interaction module standardizes the user query requests collected by the UI module, extracts structured information, and then passes it to the knowledge base agent layer. Specifically, the query interaction module receives the raw query text from the UI module, performs preprocessing operations such as removing invalid characters, standardizing punctuation, and word segmentation to generate clean, standardized query text. Then, it calls a pre-trained named entity recognition model to automatically identify the satellite model names (such as "XX-01 satellite", "XX-02 satellite", etc.) from the preprocessed query text, forming a satellite model tag list. Finally, it encapsulates the standardized query text, the identified satellite model list, the priority level automatically marked according to the urgency of the query, the request source identifier, and the timestamp into a standardized JSON request package, which is then passed to the knowledge base agent layer for multi-agent collaborative orchestration of the knowledge base.

[0052] The results display module is used to convert the analysis results output by the knowledge base intelligent agent layer into multiple formats (including structured reports, JSON data packages, chart visualizations, etc.) and display them through a UI interface according to user needs. Specifically, after receiving the analysis results output by the knowledge base intelligent agent layer, the results display module automatically selects the optimal display method based on the query type and user needs: data processing workflow queries output structured step text, parameter comparison queries output data tables, and system architecture queries output visualized charts. It then performs format conversion according to the selected display method, converting the analysis results into standard formats such as structured reports (Word / PDF format) or JSON data packages, and displays them through the UI interface module. Simultaneously, the results display module shows a confidence level indicator (high / medium / low) when displaying the analysis results to help users judge the credibility of the results, and supports exporting and downloading the results in standard format.

[0053] The technical value of the question-and-answer interface interaction layer lies in providing a unified and standardized input interface for the knowledge base intelligent agent layer. By shielding the interaction differences between different user terminals (including web browsers, mobile devices, desktop clients, etc.) through a graphical UI interface, it ensures that the knowledge base intelligent agent layer only needs to focus on business logic without having to deal with interface adaptation issues.

[0054] The knowledge base agent layer is the intelligent decision-making center of the system, carrying the knowledge base multi-agent collaborative orchestration module, the knowledge semantic association retrieval enhancement module, the knowledge update agent, and the semantic encoding module. This layer processes three types of data sources: standardized query requests transmitted from the question-and-answer interface interaction layer, including fields such as query text, satellite model list, and priority level; model-specific knowledge base retrieval results provided by the knowledge data management layer, including knowledge documents, semantic similarity scores, and document metadata; and new knowledge change notifications detected by the knowledge update agent, including knowledge document updates and deletions. The responsibilities of this layer include: multi-agent closed-loop collaboration and autonomous decision-making of retrieval strategies at the core of agent orchestration; semantic similarity retrieval and cross-model association enhancement for semantic association retrieval enhancement; vectorized detection and agent reasoning fusion and database insertion / deletion maintenance for the knowledge update agent; and knowledge read / write coordination between the knowledge base agent layer and the knowledge data management layer.

[0055] The knowledge base multi-agent collaborative orchestration module is the system's retrieval decision-making module, comprising four sub-agents: a knowledge task planning agent, a knowledge document retrieval agent, a knowledge result fusion agent, and a knowledge generation and verification agent. Each of the four sub-agents has its own independent processing context, forming a closed-loop orchestration of "planning → retrieval → fusion → verification." The core reason for adopting a multi-agent design is that each agent has its own independent context, avoiding the overly long context and potential forgetting issues that can occur when using only a single agent. The knowledge task planning agent's context, focused on retrieval strategy planning, is not diluted by the large number of intermediate results generated during retrieval. The knowledge document retrieval agent's retrieval context is not interfered with by the details of planning decisions, thus ensuring that each agent maintains high-quality decision-making within its respective area of ​​focus.

[0056] The knowledge task planning agent receives standardized query requests from the question-and-answer interface interaction layer, analyzes the query intent and generates a retrieval strategy, autonomously decides the retrieval strategy through the reasoning ability of the large language model, and passes the complete retrieval plan to the knowledge document retrieval agent.

[0057] The knowledge document retrieval agent is used to perform multi-path knowledge retrieval, parse the retrieval instructions passed by the knowledge task planning agent, execute the corresponding retrieval strategy according to the retrieval mode, call the knowledge semantic association retrieval enhancement module to perform retrieval in the dedicated knowledge base and general knowledge base of each model, and output the retrieval result set.

[0058] The knowledge result fusion agent is used to perform weighted fusion and confidence evaluation of multi-way retrieval results. After receiving the retrieval result set of the knowledge document retrieval agent, it calculates a comprehensive score for each document (considering semantic similarity, model knowledge base credibility and satellite model relevance). The top-N documents are selected as the fused candidate text slice set according to the comprehensive score in descending order, and the overall confidence is calculated.

[0059] The knowledge generation and verification agent is used to generate answers and perform quality verification. It receives a candidate text slice set, an overall confidence score, and the original query text from the knowledge result fusion agent. When the overall confidence score is greater than a confidence threshold, the agent inputs the candidate text slice set and the original query text into a large language model for inference, outputting the final analysis result to the result display module of the question-and-answer interface for UI display. When the overall confidence score is less than a preset confidence threshold, the agent identifies missing domain knowledge data in the knowledge database based on the difference between the knowledge content required to answer the query text and the knowledge content provided by the target knowledge text set. This missing domain knowledge data, along with the query text and at least one satellite model to be queried, is packaged into a knowledge database entry suggestion and passed to the knowledge update agent. Simultaneously, improvement feedback is generated and sent to the knowledge task planning agent. This feedback includes the overall confidence score, the score distribution of each candidate knowledge text, an analysis of the reasons for failed verification, and suggestions for adjusting the retrieval strategy, triggering a re-decision by the knowledge task planning agent.

[0060] The knowledge semantic association retrieval enhancement module, as a callable tool module of the knowledge document retrieval agent, is independent of the knowledge base multi-agent collaborative orchestration module. Upon receiving retrieval requests from the knowledge document retrieval agent, the module first performs semantic vector encoding on the query text using the semantic encoding module. It then performs semantic similarity retrieval in the vector indexes of each model-specific knowledge base and the general knowledge base to obtain an initial candidate set. Next, it enhances the association of the candidate set using a cross-model relationship graph. Finally, it performs a final ranking of the candidates using a comprehensive score weighted fusion method, outputting the retrieval results after semantic association enhancement and comprehensive score ranking (each result includes the original text of the text slice, semantic similarity score, comprehensive ranking score, document source model knowledge base identifier, satellite model tag, and other metadata), which is returned to the knowledge document retrieval agent.

[0061] The knowledge update agent is the system's dynamic knowledge base maintenance module. It receives knowledge update trigger signals from the question-and-answer interface layer and the knowledge base agent layer. These signals include user-uploaded satellite model knowledge documents and manually deleted knowledge documents via the UI, as well as knowledge entry suggestions generated by the knowledge generation and verification agent during the current query. The knowledge update agent performs a complete three-step update process: vectorized detection of new knowledge, agent reasoning fusion, and entry into the knowledge base along with deletion of old knowledge. After the update is complete, the knowledge update agent returns the operation results (including update status, a list of changed documents, and the new version number) to the UI module of the question-and-answer interface layer for display.

[0062] The semantic encoding module, as a shared foundational component, employs a single embedding model (Qwen3-Embedding) to provide unified semantic encoding services for the knowledge semantic association retrieval enhancement module and the knowledge update agent. All text vector transformations (including user query text, knowledge text slices, query context, etc.) are completed through this embedding model, converting natural language text into dense semantic vectors of fixed dimensions (default 1024 dimensions), ensuring that all text vectors reside in the same semantic space and supporting accurate semantic similarity calculation. The single encoding latency does not exceed 100 milliseconds.

[0063] The knowledge data management layer is the system's knowledge storage and retrieval infrastructure, responsible for the construction, retrieval, and continuous maintenance of the retrieval-enhanced knowledge base stored by satellite model domain. This layer handles two types of data sources: knowledge read / write requests from the knowledge base agent layer, vector retrieval requests from the knowledge semantic association retrieval enhancement module, and incremental index update requests from the knowledge update agent. The responsibilities of this layer include the storage and management of the satellite model-domain-based knowledge base, efficient construction of vector indexes and approximate nearest neighbor retrieval, maintenance and querying of cross-model relationship graphs, and the creation and rollback management of knowledge base versions.

[0064] The knowledge data management layer includes a general knowledge base and satellite-specific knowledge bases (such as the XX-01 satellite knowledge base, XX-02 satellite knowledge base, XX-03 satellite knowledge base, etc.), as well as a vector indexing engine, a cross-model relationship graph, and a version manager. The knowledge data management layer adopts a three-dimensional knowledge organization architecture of "model-specific domain storage + unified index + relationship graph," storing and managing general aerospace knowledge and domain knowledge for each satellite model separately, while maintaining the relationships between knowledge from different models through a cross-model relationship graph.

[0065] A general knowledge base stores cross-model, universal aerospace fundamentals (such as orbital mechanics principles and remote sensing basic theories), while each model's dedicated knowledge base stores domain-specific knowledge (e.g., the XX-01 satellite knowledge base stores parameter configurations, processing procedures, and fault cases for XX-01 satellite, while the XX-02 satellite knowledge base stores corresponding knowledge for XX-02 satellite). This model-specific domain storage makes the size of each model's knowledge base controllable, improves retrieval efficiency, and supports independent updates and maintenance for each model's knowledge base.

[0066] The process for building the enhanced knowledge base is as follows: After the original knowledge documents (including technical documents, parameter manuals, operation and maintenance experience, etc.) are input, the system segments the documents according to preset parameters (default block size 512 tokens, default overlap length 50 tokens), splitting long documents into multiple slices. Each slice is accompanied by metadata information (including satellite model tags, document source, creation time, and the slice's position offset in the original text). The sliced ​​text is converted into semantic vectors by the semantic encoding module and stored along with the metadata in the corresponding model's vector index, forming searchable knowledge entries.

[0067] The vector indexing engine is used for efficient indexing and near-nearest neighbor retrieval of document vectors in various knowledge bases. All knowledge bases share the same vector indexing engine (implemented based on FAISS or Milvus), while the indexes of different knowledge bases are isolated through a namespace mechanism, supporting both single-model retrieval and cross-model unified retrieval modes. The vector indexing engine uses the IVF-PQ algorithm to build high-performance vector indexes, responding to vector retrieval requests (including query vectors, retrieval quantity, and target knowledge base identifier) ​​submitted by the knowledge semantic association retrieval enhancement module or the knowledge document retrieval agent. It performs near-nearest neighbor search within the namespace of the specified knowledge base, with a single retrieval latency of no more than 50 milliseconds. Simultaneously, the vector indexing engine responds to incremental update requests from the knowledge update agent, adding or deleting a small number of vector records without rebuilding the entire index.

[0068] A cross-satellite relationship graph is used to maintain semantic associations and complementary relationships between knowledge from different satellite models, providing relationship enhancement capabilities for cross-model retrieval in the knowledge semantic association retrieval enhancement module. The cross-satellite relationship graph stores the associations between knowledge from different models in a graph structure (e.g., "comparison between parameters of XX-01 satellite and XX-02 satellite", "complementary relationship between data from one sensor and data from another sensor"), establishing knowledge nodes and associated edges, implemented through a graph database (e.g., Neo4j). The cross-satellite relationship graph responds to association query requests from the knowledge semantic association retrieval enhancement module, querying association paths between knowledge based on conditions such as relationship type and association strength, and returning a list of association paths for comprehensive scoring and re-ranking to weight cross-satellite association degrees. Simultaneously, the cross-satellite relationship graph responds to association change notifications from the knowledge update agent, dynamically updating the relationship nodes and edges in the graph.

[0069] The version manager automatically creates version snapshots after each knowledge base update, recording the version number, update time, and change summary. It supports knowledge base version rollback and change auditing. Each time a knowledge update agent completes an update for a specific knowledge base model, the version manager automatically creates a snapshot of the current knowledge base state. When an error is detected during a knowledge update or an administrator initiates a rollback request, the version manager restores the corresponding knowledge base model to the specified historical version snapshot state. The version manager also records a complete knowledge change log, supporting querying of change history by time, model, change type, and other dimensions.

[0070] The system's data flow is as follows: Users input query requests through the UI of the question-and-answer interface layer. After format standardization and satellite model entity recognition, the standardized query request is passed to the knowledge base intelligent agent layer for multi-agent collaborative orchestration. This involves a closed-loop processing flow: knowledge task planning agent → knowledge document retrieval agent (calling the knowledge semantic association retrieval enhancement module) → knowledge result fusion agent → knowledge generation and verification agent. During processing, the knowledge base intelligent agent layer acquires domain knowledge from the dedicated knowledge bases of each satellite model through the knowledge data management layer. The knowledge update agent monitors new knowledge entry, performs vectorization detection, agent reasoning fusion, and deletion of both new and old knowledge, returning the results to the UI module of the question-and-answer interface layer for display. Finally, the generated analysis results are returned to the user through the question-and-answer interface layer, and the results display module performs multi-format conversion and display according to user needs.

[0071] The three-layer architecture described above achieves separation of concerns: the question-and-answer interface layer focuses on graphical user interaction and question input / output display; the knowledge base agent layer focuses on agent orchestration decision-making and semantic association retrieval algorithms; and the knowledge data management layer focuses on knowledge storage and retrieval categorized by satellite model. Each layer communicates through clearly defined data interfaces, ensuring that changes within any layer will not affect the normal operation of other layers, thus guaranteeing the system's maintainability and scalability.

[0072] Figure 3 The illustration shows a schematic diagram of the business logic of a multi-satellite model knowledge retrieval and generation method based on multi-agent collaboration according to an embodiment of this application.

[0073] like Figure 3 As shown, the multi-agent collaborative orchestration of the knowledge base includes four sub-agents: a knowledge task planning agent, a knowledge document retrieval agent, a knowledge result fusion agent, and a knowledge generation and verification agent. Each of the four sub-agents has its own independent processing context, forming a closed-loop orchestration of "planning → retrieval → fusion → verification → reflection → replanning".

[0074] In some embodiments, the knowledge task planning agent receives standardized query requests, analyzes the query intent through a large language model, and generates a retrieval strategy. Corresponding to step S110, in response to a query request containing query text and at least one satellite model to be queried, a retrieval strategy is generated based on the knowledge task planning agent. S110 includes steps S111 to S114.

[0075] In step S111, the semantic complexity of the query text and the number of at least one satellite model to be queried are analyzed by the knowledge task planning agent.

[0076] The knowledge task planning agent receives standardized query requests containing query text, at least one satellite model to be queried, query context, and priority level. The agent concatenates the query text, at least one satellite model to be queried, and query context into structured input, invokes a large language model for reasoning, and generates a retrieval strategy decision. Specifically, the agent analyzes the semantic complexity of the query text using the large language model (comprehensively evaluating dimensions such as query length, number of satellite models involved, and whether it contains complex semantic patterns), and, combined with the number of at least one satellite model to be queried, autonomously decides whether to adopt a single-database retrieval strategy, a multi-database parallel retrieval strategy, or an iterative deepening retrieval strategy. The retrieval instructions include the retrieval mode, a list of target knowledge bases, and retrieval parameters (including an upper limit on the number of returned results and a semantic matching quality threshold).

[0077] In step S112, when at least one satellite model to be queried is a single model, a single-database retrieval strategy is generated through a knowledge task planning agent. The single-database retrieval strategy searches a single knowledge database corresponding to the satellite model to be queried, and the retrieval command is used to perform the retrieval within that single knowledge database. Under this strategy, only the knowledge document retrieval agent and the knowledge generation and verification agent are activated, enabling a lightweight path for quickly responding to simple queries for a single model and avoiding unnecessary processing overhead caused by multi-agent collaboration.

[0078] In step S113, when there are multiple satellite models to be queried, a multi-database parallel retrieval strategy is generated through a knowledge task planning agent. The multi-database parallel retrieval strategy involves searching multiple knowledge databases that correspond one-to-one with at least one satellite model to be queried, and retrieval instructions are used to execute retrieval in parallel across multiple knowledge databases. Under this strategy, three agents—a knowledge document retrieval agent, a knowledge result fusion agent, and a knowledge generation and verification agent—are activated. These agents execute retrieval in parallel across multiple model knowledge bases according to the retrieval instructions to handle complex query scenarios such as cross-model comparative analysis.

[0079] In step S114, if the semantic complexity of the query text exceeds a first preset threshold or fails verification, a retrieval instruction for an iterative deepening retrieval strategy is generated by the knowledge task planning agent. The iterative deepening retrieval strategy involves searching a knowledge database that includes at least one satellite model to be queried, and the retrieval instruction is used to execute multiple rounds of retrieval in conjunction with the query text. Under this strategy, multiple rounds of retrieval are executed by rewriting the query and adjusting retrieval parameters to gradually approach the answer, addressing complex query scenarios with semantic ambiguity or insufficient initial retrieval quality.

[0080] Therefore, the knowledge task planning agent can autonomously decide on the retrieval strategy based on the semantic complexity of the query and the number of satellite models involved, enabling simple queries to be responded to quickly with a lightweight process, and complex queries to be fully processed through the collaboration of multiple agents, thus achieving adaptive matching between the retrieval strategy and the query features.

[0081] In some embodiments, the knowledge document retrieval agent is used to perform single-database or multi-database knowledge retrieval according to retrieval instructions. The knowledge document retrieval agent performs retrieval in the target model knowledge base and the general knowledge base by invoking the knowledge semantic association retrieval enhancement module (the specific processing flow of the knowledge semantic association retrieval enhancement will be discussed later). Figure 4 (Detailed description in the text), outputting a search result set. Each record in the search result set contains metadata fields such as slice identifier, original text of the slice, semantic similarity score, source model knowledge base identifier, and satellite model tag.

[0082] The knowledge document retrieval agent also calculates the retrieval hit rate H during the retrieval process. hit This refers to the proportion of text segments in the search results whose semantic similarity reaches a preset threshold. When H... hit When the similarity threshold is lower than the preset threshold (default 0.5), the knowledge document retrieval agent triggers a retrieval scope expansion strategy (such as lowering the similarity threshold or increasing the Top-K value) to ensure that key knowledge is not missed.

[0083] The knowledge document retrieval agent maintains its processing context during the retrieval process, including a list of retrieved knowledge bases and the retrieved results, which are used for deduplication and search scope expansion in subsequent searches. After obtaining the search results, the knowledge document retrieval agent performs summary deduplication, retains the highest-scoring text slices that are repeated across knowledge bases, and outputs the retrieval result set and hit rate report after sorting them in descending order of comprehensive score.

[0084] In some embodiments, the knowledge result fusion agent receives a candidate knowledge text set from the knowledge document retrieval agent. Each candidate knowledge text in the candidate knowledge text set carries semantic similarity to the query text, relevance to at least one satellite model to be queried, and a comprehensive score. Corresponding to step S130, the knowledge result fusion agent filters the target knowledge text set based on the relevance of each candidate knowledge text in the candidate knowledge text set to the query text and the satellite model to be queried, and calculates the overall confidence level. Step S130 includes steps S131 to S132.

[0085] In step S131, the knowledge result fusion agent determines the fusion score of each candidate knowledge text based on the semantic similarity of each candidate knowledge text to the query text, the correlation between each candidate knowledge text and at least one satellite model to be queried, and the preset credibility of the knowledge database from which each candidate knowledge text originates. The fusion score is used to comprehensively measure the overall performance of each candidate knowledge text in three dimensions: semantic relevance, model matching degree, and source reliability.

[0086] In step S132, the knowledge result fusion agent sorts the candidate knowledge texts in descending order according to the fusion score, and selects the top-ranked candidate knowledge texts as the target knowledge text set. During the sorting and selection process, the knowledge result fusion agent automatically performs cross-knowledge base deduplication, retaining the text slice with the highest fusion score for duplicate text slices. Thus, the knowledge texts with the best overall quality are selected from the candidate knowledge texts, providing refined knowledge input for subsequent answer generation.

[0087] The knowledge outcome fusion agent is also used to calculate the overall confidence score of the target knowledge text set. Specifically, the agent calculates the percentage of text segments in the target knowledge text set whose fusion score reaches a second preset threshold, thus obtaining the overall confidence score. In one example, the second preset threshold can be set to 0.4, meaning that target knowledge texts with a fusion score greater than or equal to 0.4 are considered high-quality texts, and the overall confidence score reflects the proportion of high-quality texts in the target knowledge text set. This overall confidence score is used to quantitatively evaluate the overall quality of the target knowledge text set, providing a quantifiable basis for subsequent verification steps.

[0088] In some embodiments, the knowledge generation and verification agent is used to validate the target knowledge text set based on the overall confidence level. Specifically, the knowledge generation and verification agent determines whether the overall confidence level is less than a preset confidence threshold. In one example, the preset confidence threshold can be set to 0.6. If the overall confidence level is greater than or equal to the preset confidence threshold, the validation passes, and the knowledge generation and verification agent inputs the target knowledge text set and query text into the large language model to generate and output the final analysis result. If the overall confidence level is less than the preset confidence threshold, the validation fails, and the knowledge generation and verification agent generates reflective information to trigger the knowledge task planning agent to regenerate the retrieval strategy. Thus, the quality of the target knowledge text set is pre-validated before the large language model generates the answer, ensuring that the answer generation stage only begins when the knowledge quality meets the preset conditions. This avoids low-quality answer generation due to insufficient retrieval quality, improving the reliability and credibility of the final question-and-answer results.

[0089] In some embodiments, after the knowledge generation and verification agent completes the verification, it performs corresponding follow-up processing based on the verification result.

[0090] If the verification passes, the knowledge generation and verification agent inputs the target knowledge text set and the query text into the large language model to generate the final analysis result. This final analysis result is passed to the result display module of the question-and-answer interface interaction layer, where it performs multi-format conversion and visualization rendering according to the query type and user needs. Simultaneously, the knowledge generation and verification agent determines blank domain knowledge data based on the difference between the knowledge content required to answer the query text and the knowledge content provided by the target knowledge text set. This blank domain knowledge data, along with the query text and at least one satellite model to be queried, is packaged into knowledge entry suggestion data and passed to the knowledge update agent. Thus, the knowledge generation and verification agent, while performing generation and verification functions, also plays a role in providing retrieval quality feedback. Blank domain knowledge data refers to the knowledge lacking in the knowledge database, determined by the knowledge generation and verification agent based on the difference between the knowledge content required to answer the query text and the knowledge content provided by the target knowledge text set. Feeding the knowledge entry suggestion data back to the knowledge update agent triggers the incremental knowledge update process, ensuring that new knowledge is promptly entered into the database and reflected in subsequent searches, achieving a two-way collaborative closed loop between retrieval and knowledge updating.

[0091] For example, if a user queries "What are the design parameters of the heat dissipation system for satellite XX-03?", the large language model, in the process of generating the final analysis result, first deduces the knowledge subdomains required to answer the query based on the semantic connotation of the query text. This includes the type of heat dissipation system (active or passive heat dissipation), heat dissipation power parameters, heat dissipation loop design, heat dissipation material specifications, etc., forming a demand list. Subsequently, the large language model checks which knowledge is actually matched in the candidate knowledge text set (i.e., the actual knowledge content provided in the retrieval): for example, the search results may contain "basic orbital parameters of satellite XX-03" and "payload configuration of satellite XX-03", but no content related to the heat dissipation system. After comparing the demand list with the supply list item by item, the large language model identifies the knowledge domain "design parameters of the heat dissipation system for satellite XX-03" as existing in the demand list but completely missing in the supply list, thus determining this knowledge domain as blank domain knowledge data. The knowledge generation and verification agent packages the blank domain knowledge data with the query text "Design parameters of the heat dissipation system of XX-03 satellite" and the satellite model to be queried "XX-03 satellite" into knowledge storage suggestion data, and passes it to the knowledge update agent to trigger the subsequent knowledge incremental update process.

[0092] If the verification fails, the knowledge generation verification agent does not proceed to the answer generation stage. Instead, it generates reflection information and passes it to the knowledge task planning agent, triggering the agent to regenerate the retrieval strategy. The knowledge task planning agent then passes the regenerated retrieval strategy to the knowledge document retrieval agent, which performs the retrieval again based on the new strategy (returning to step S120) until the verification passes or the preset number of rounds is reached. The reflection information may include the overall confidence level of the target knowledge text set, the distribution of the fusion scores of each knowledge text, and the reasons for the verification failure. The knowledge task planning agent adjusts its retrieval strategy based on the reflection information (e.g., expanding the scope of the retrieval database, adjusting retrieval parameters, rewriting the query expression, etc.), and then the knowledge document retrieval agent performs the retrieval again based on the regenerated retrieval strategy. This reflection iteration process continues until the verification passes or the preset number of rounds is reached, thereby continuously optimizing the retrieval quality through a closed-loop feedback mechanism.

[0093] In some embodiments of this application, the iterative retry mechanism of the knowledge task planning agent also serves to perform secondary verification of knowledge gaps. After receiving the reflection information, the knowledge task planning agent re-formulates the retrieval strategy, including adjusting the scope of the retrieval database (such as expanding from a single-model knowledge base to multiple model knowledge bases or a general knowledge base), adjusting retrieval parameters (such as lowering the similarity threshold, increasing the number of returned candidate knowledge texts), or rewriting the query statement, and then re-initiating the retrieval.

[0094] This iterative process can be used to verify the authenticity of the blank domain knowledge data initially identified by the knowledge generation verification agent. Specifically, if after multiple iterations and using at least two different retrieval strategies, a certain knowledge domain consistently fails to return relevant knowledge text, then the knowledge domain is confirmed to be a genuine knowledge blank, meaning that the knowledge database itself does indeed lack relevant knowledge. Conversely, if adjusting the retrieval strategy (e.g., lowering the similarity threshold) successfully returns previously undetected knowledge text, then the knowledge domain is confirmed not to be a knowledge blank, but rather a missed detection caused by an inappropriate previous retrieval strategy.

[0095] Through the aforementioned iterative verification mechanism, the knowledge task planning agent assists the knowledge generation verification agent in distinguishing whether the missed detection is due to a lack of knowledge in the knowledge base itself or an inappropriate retrieval strategy. This makes the finally confirmed blank domain knowledge data more reliable and avoids misjudging actual knowledge in the knowledge base as a knowledge gap due to retrieval strategy issues.

[0096] In existing technologies, knowledge base maintenance relies on manual discovery of knowledge gaps, which is inefficient and prone to delays. This application addresses this by having a knowledge generation and verification agent automatically identify knowledge data in blank domains, a knowledge task planning agent iteratively verify its authenticity through multiple rounds, and a knowledge update agent automatically perform incremental updates. This forms an automated "identification → verification → update" chain that requires no manual intervention, making retrieval failures the starting point for knowledge base improvement and significantly reducing the human cost of knowledge base maintenance.

[0097] Figure 4 The illustration shows a schematic diagram of the business logic of the knowledge semantic association retrieval enhancement module according to an embodiment of this application.

[0098] like Figure 4 As shown, the knowledge semantic association retrieval enhancement module, as a callable tool module of the knowledge document retrieval intelligent agent, responds to the call of the knowledge document retrieval intelligent agent, receives the retrieval request transmitted by the knowledge document retrieval intelligent agent, the retrieval request includes query text, target satellite model list and retrieval parameters, enhances and optimizes the retrieval results through a three-level processing flow of "semantic similarity retrieval → cross-model association enhancement → comprehensive score ranking", and returns the processed candidate knowledge text set to the knowledge document retrieval intelligent agent.

[0099] The knowledge semantic association retrieval enhancement module first encodes the query text into a semantic vector through the semantic encoding module, and performs an approximate nearest neighbor retrieval of the query text in at least one knowledge database to obtain an initial candidate text set and the semantic similarity between each initial candidate knowledge text in the initial candidate text set and the query text. This step corresponds to step S120, which includes steps S121 to S124.

[0100] In step S121, the semantic encoding module uses a single embedding model to convert the query text into a dense semantic vector of fixed dimensions. Then, it performs an approximate nearest neighbor retrieval in the vector indexes of the target model knowledge base and the general knowledge base, and calculates the cosine similarity between the query vector and each text segment vector.

[0101]

[0102] in, For query text With text fragments The cosine similarity between them takes values ​​in the range [0,1]. For query text The semantic vector encoded by the semantic encoding module; For text fragments The semantic vector encoded by the semantic encoding module; and These are the L2 norms of the corresponding vectors. This step obtains the initial set of candidate text fragments and the semantic similarity score between each candidate text fragment and the query text.

[0103] The knowledge semantic association retrieval enhancement module utilizes a cross-model relationship graph to enhance the cross-model association of the initial ranking results. The cross-model relationship graph stores the associations between different satellite models in a graph structure, such as "the comparison relationship between the parameters of XX-01 satellite and the parameters of XX-02 satellite" and "the complementary relationship between data from one sensor and data from another sensor", which is implemented through a graph database (such as Neo4j).

[0104] In this embodiment, when each knowledge text is stored in the knowledge database, the system adds a corresponding satellite model tag to the knowledge text. This tag, as structured metadata, is stored in a vector index in association with the text content and is used to identify the specific satellite model involved or to which the knowledge text belongs. The tag value is consistent with the node identifier in the cross-model relationship graph. The cross-model relationship graph stores the associations between different satellite models in a graph structure. The nodes in the graph are uniquely identified by the satellite model, and the edges represent the associations between different models, such as technology inheritance, parameter comparison, common components, or functional similarity.

[0105] In step S122, the knowledge document retrieval agent invokes the cross-model relationship graph to match the satellite model tags of each initial candidate knowledge text with at least one satellite model to be queried, and queries the association path between the satellite model tags of each initial candidate knowledge text and at least one satellite model to be queried, so as to determine the degree of association between each initial candidate knowledge text and at least one satellite model to be queried.

[0106] After extracting the satellite model tags carried by the initial candidate knowledge text, the knowledge document retrieval agent uses the tag value as the query key to locate the corresponding model node in the cross-model relationship graph and searches for the association path between this node and at least one satellite model node specified in the query request. If the path exists, it is determined that the initial candidate knowledge text is related to the satellite model to be queried, and the association degree is the first value; if it does not exist, the association degree is the second value.

[0107] correlation A binary decision method is used: if a related path is found, the satellite model tag of the initial candidate knowledge text is determined. With at least one list of satellite models to be queried If there is a direct intersection, then the correlation is... For example, the first value. If no related path is found, the relevance is determined to be the second value, for example... .

[0108] This mechanism addresses the semantic omission problem caused by "different expressions of the same concept" in multi-satellite scenarios. Different satellite models may use different terms and expressions for the same physical concept. For example, for the concept of "resolution," technical documents for different models may use different expressions such as "spatial resolution," "pixel resolution," and "ground sampling distance." Through cross-model relationship graph enhancement, even if a text fragment has low semantic similarity, if its model has a related path to the query model, the enhanced association can improve the ranking of the text fragment, thereby effectively recalling semantically related knowledge texts from different model knowledge bases.

[0109] In step S123, the knowledge semantic association retrieval enhancement module determines the comprehensive score of each initial candidate knowledge text based on semantic similarity and relevance. The ranking score is determined by semantic similarity. and cross-model correlation Calculated using a linear combination of preset weights, taking into account both the semantic relevance and model association of the document:

[0110]

[0111] in, For text fragments The final ranking score; To query text q and text fragments Semantic similarity between them; For relevance; Weights for semantic similarity scores (default 0.7, range [0.5, 0.9]); The weight of the correlation (default 0.3) satisfies .

[0112] In step S124, the knowledge semantic association retrieval enhancement module selects the top-K initial candidate knowledge texts from the initial candidate text set based on the comprehensive score, forming a candidate knowledge text set, and returns the candidate knowledge text set to the knowledge document retrieval agent. Each candidate knowledge text in the candidate knowledge text set carries semantic similarity to the query text, association with at least one satellite model to be queried, and a comprehensive score, which are used by the subsequent knowledge result fusion agent for fusion and filtering.

[0113] Through the above processing flow, the knowledge semantic association retrieval enhancement module of this application embodiment incorporates cross-model association information on the basis of semantic matching, which effectively improves the recall and accuracy of cross-model knowledge retrieval in the scenario of heterogeneous data of multiple satellite models, so that the subsequent fusion and generation steps can obtain more comprehensive knowledge input.

[0114] Figure 5 The illustration schematically shows a business logic diagram of a knowledge updating agent according to an embodiment of this application.

[0115] like Figure 5 As shown, the knowledge update agent is the system's knowledge base dynamic maintenance module. In response to the acquisition of new knowledge data, it executes a three-step closed-loop knowledge incremental update process: "vectorized spatial retrieval and change detection → relation type-based fusion processing → incremental entry and deletion of old knowledge", thereby realizing the automated incremental maintenance of the knowledge base.

[0116] The sources of new knowledge data include user-uploaded knowledge documents and / or knowledge entry suggestion data generated by the knowledge generation and verification agent. Users can upload satellite model knowledge documents (supporting PDF, Word, TXT, and other formats) through the UI interface. The UI interface automatically extracts the satellite model information from the document, selects the corresponding target model knowledge base, and passes the uploaded document as new knowledge data to the knowledge update agent. Knowledge entry suggestion data includes the query text, at least one satellite model to be queried, and blank domain knowledge data. The knowledge generation and verification agent passes the knowledge entry suggestion data to the knowledge update agent after completing the question-and-answer process.

[0117] In response to acquiring new knowledge data, the knowledge updating agent first encodes the new knowledge data into a semantic vector using a semantic encoding module. Then, it retrieves multiple existing knowledge texts semantically adjacent to the new knowledge data from the knowledge database, forming a neighborhood set. Specifically, the knowledge updating agent encodes the new knowledge into a semantic vector using the semantic encoding module. ,in, For new knowledge content, This is the encoding function for the semantic encoding module. Subsequently, the knowledge updating agent performs an approximate nearest neighbor search in the vector space of the existing knowledge base to find the nearest neighbor... The top-K semantically nearest existing knowledge text slices (default K=10) form a neighborhood set N. Neighborhood set N will serve as the input candidate set for fusion reasoning in step two, used to determine the relationship type between new knowledge and existing knowledge.

[0118] The knowledge updating agent, based on the relationship type between new knowledge data and existing knowledge text, fuses the new knowledge data with the neighborhood set to obtain the fused new knowledge content and the existing knowledge text in the neighborhood set that is determined to be discarded. Specifically, the knowledge updating agent inputs the new knowledge data and existing knowledge text fragments from the neighborhood set N into a large language model, which analyzes the relationship type between the new knowledge and each existing knowledge fragment in the neighborhood set. Relationship types include four categories: addition, supplementation, update, and conflict. When the relationship type is addition, the new knowledge is entirely new and unrelated to existing knowledge; the new knowledge data is directly added to the database as the fused new knowledge content. When the relationship type is supplementation, the new knowledge is an extension of existing knowledge; the new knowledge data is fused with the existing knowledge text to obtain the fused new knowledge content, and this existing knowledge text is determined to be discarded. When the relationship type is update, the new knowledge is a modification or replacement of existing knowledge; the new knowledge data is fused with the existing knowledge text to obtain the fused new knowledge content, and this existing knowledge text is determined to be discarded. When the relationship type is conflicting, the new knowledge contradicts the existing knowledge. Based on preset rules, the system will either retain the new knowledge data or the existing knowledge text as the new knowledge content after fusion, or prompt the user for confirmation.

[0119] The entire fusion reasoning process is driven entirely by the semantic understanding and reasoning capabilities of the large language model. No additional threshold calculations are required. It can accurately identify various complex relationships between new and existing knowledge and adapt to diverse knowledge association scenarios in satellite data.

[0120] The knowledge update agent performs knowledge base update operations based on the fusion processing results. Specifically, it writes the fused new knowledge content into the knowledge database of the corresponding satellite model and deletes existing knowledge texts that are deemed to need to be phased out. For the fused new knowledge content, it performs segmentation and vectorization processing and writes it into the vector index of the corresponding model's knowledge base; for existing knowledge texts that the large language model determines need to be phased out, it removes the corresponding vector records and metadata information from the vector index. If the received request is to delete a specific knowledge document, it directly deletes the text fragments related to the document in the knowledge base and updates the relationships in the cross-model relationship graph.

[0121] After the update is complete, the knowledge update agent creates a version snapshot of the current model's knowledge base, records the version number, update time, and change summary, writes it to the version manager, and returns the operation result information (including update status, number of added slices, number of deleted slices, and new version number) to the UI module of the question-and-answer interface interaction layer for display. Model-based domain storage allows each model's knowledge base to be updated and maintained independently; an update to one model's knowledge base will not affect the normal operation of other models' knowledge bases, improving the system's maintainability.

[0122] Through the aforementioned three-step closed-loop knowledge incremental update process, the knowledge update intelligent agent in this embodiment of the application achieves automated incremental maintenance of knowledge from generation to storage. The latency from knowledge generation to storage is shortened from several days in traditional manual updates to minutes, significantly improving the timeliness of the knowledge base. Simultaneously, by feeding back knowledge storage suggestion data to the knowledge update intelligent agent, new knowledge can be promptly stored and reflected in subsequent searches, achieving a collaborative closed loop between retrieval and knowledge updating.

[0123] Figure 6 The diagram illustrates a system operation flowchart suitable for implementing a multi-satellite model knowledge retrieval and generation method based on multi-agent collaboration, according to an embodiment of this application.

[0124] like Figure 6 As shown, the system operation flow of this application embodiment includes seven stages, which are sequentially processed by the question-and-answer interface interaction layer, the knowledge base intelligent agent layer, and the knowledge data management layer, forming an end-to-end closed-loop process.

[0125] Phase 1, completed at the question-and-answer interface interaction layer, aims to convert user-input query requests into a standardized format that the system can process. After the user inputs a query request through the UI interface, the UI interface module passes the raw query text and session metadata to the query interaction module. The query interaction module first performs text preprocessing (removing invalid characters, standardizing punctuation, and word segmentation), and then calls a pre-trained named entity recognition model to automatically extract the relevant satellite model names (such as "XX-01 satellite", "XX-02 satellite", etc.) from the query text, forming a satellite model tag list. The extracted satellite model list, along with the standardized query text, request source identifier, priority level, and other information, is packaged into a standardized JSON request package and passed to the knowledge base agent layer for multi-agent collaborative orchestration of the knowledge base.

[0126] Phase 2, completed within the knowledge task planning agent at the knowledge base agent layer, aims to analyze query intent and generate the optimal retrieval strategy. Upon receiving a standardized query request, the knowledge task planning agent concatenates the query text, satellite model list, and query context into structured input. It then uses a large language model to infer and generate a structured text for retrieval strategy decisions. Based on this structured text, the knowledge task planning agent determines whether to employ a single-database retrieval (query involving only a single satellite model), a multi-database parallel retrieval (query involving comparison or joint analysis of multiple models), or an iteratively refined retrieval strategy. Finally, it transmits the complete retrieval plan (including retrieval mode, target model knowledge base list, and retrieval parameters) to the knowledge document retrieval agent.

[0127] Phase 3 is completed in the knowledge document retrieval agent at the knowledge base agent layer, simultaneously invoking the knowledge semantic association retrieval enhancement module and the vector indexing engine and cross-model relationship graph of the knowledge data management layer. The knowledge document retrieval agent parses the retrieval instructions passed by the knowledge task planning agent, executes the corresponding retrieval strategy according to the retrieval mode, and invokes the knowledge semantic association retrieval enhancement module to perform the retrieval. The knowledge semantic association retrieval enhancement module first performs semantic vector encoding on the query through the semantic encoding module, performs approximate nearest neighbor retrieval in each model knowledge base for initial semantic similarity ranking, then uses the cross-model relationship graph to enhance the relevance of candidate documents, and finally refines the ranking based on the comprehensive score to output the final retrieval results. The output retrieval result set (each record containing metadata fields such as slice identifier, original text slice, semantic similarity score, source model knowledge base identifier, and satellite model tag) is passed to the knowledge result fusion agent.

[0128] Phase 4 is completed in the knowledge result fusion agent at the knowledge base agent layer. Its goal is to perform weighted fusion of multi-path retrieval results and evaluate the overall confidence level. After receiving the retrieval result set, the knowledge result fusion agent determines the fusion score of each candidate knowledge text based on the semantic similarity of each candidate knowledge text to the query text, the correlation between each candidate knowledge text and at least one satellite model to be queried, and the preset credibility of the knowledge database from which each candidate knowledge text originates. The top-ranked candidate knowledge texts are then selected as the target knowledge text set, arranged in descending order of fusion score. Simultaneously, the knowledge result fusion agent calculates the percentage of text segments in the target knowledge text set whose fusion score reaches a preset threshold, obtaining the overall confidence level. The target knowledge text set and the overall confidence level are then passed to the knowledge generation and verification agent.

[0129] Phase 5 is completed in the knowledge generation and verification agent at the knowledge base agent layer. The goal is to generate high-quality answers based on the target knowledge text set and perform quality verification. The knowledge generation and verification agent determines whether the overall confidence level is less than a preset confidence threshold. When the overall confidence level is greater than or equal to the preset confidence threshold, the verification passes. The knowledge generation and verification agent then inputs the target knowledge text set and the original query text into the large language model for reasoning, outputs the final analysis result, and passes it to the result display module of the question-and-answer interface interaction layer for UI display. Simultaneously, the knowledge generation and verification agent identifies blank domain knowledge data based on the difference between the knowledge content required to answer the query text and the knowledge content provided by the target knowledge text set. It packages the query text involved in this query, at least one satellite model to be queried, and blank domain knowledge data into knowledge entry suggestion data and passes it to the knowledge update agent. When the overall confidence level is lower than the preset confidence threshold, the verification fails. The knowledge generation and verification agent generates reflection information and passes it to the knowledge task planning agent, triggering a new round of retrieval strategy planning, forming an autonomous optimization closed loop of "planning → retrieval → fusion → verification → reflection → replanning".

[0130] Phase 6 is completed in the Q&A interface interaction layer. Its goal is to convert the analysis results into multiple formats and display them according to user needs. After receiving the final analysis results, the results display module automatically selects the optimal display method based on the query type and user requirements: data processing workflow queries output structured step text, parameter comparison queries output data tables, and system architecture queries output visual charts. Format conversion is then performed according to the selected display method, converting the analysis results into standard formats such as structured reports (Word / PDF format) or JSON data packets, and then rendering and displaying them visually in the UI interface. Users can download the results as standard format files using the export function in the UI interface.

[0131] Phase 7 is completed in the knowledge update agent at the knowledge base agent layer. It simultaneously invokes the semantic encoding module and the model-specific knowledge base group, vector indexing engine, cross-model relationship graph, and version manager at the knowledge data management layer. The goal is to write new knowledge data into the corresponding model knowledge base and perform knowledge document deletion operations. After receiving satellite model knowledge documents uploaded by users through the UI interface or knowledge entry suggestions submitted by the knowledge generation and verification agent, the knowledge update agent first encodes the new knowledge data into semantic vectors using the semantic encoding module. It then performs an approximate nearest neighbor search in the vector space of the existing knowledge base to find the Top-K existing knowledge text slices that are semantically closest to the new knowledge, forming a neighborhood set. Subsequently, the knowledge update agent inputs the new knowledge data and the neighborhood set into a large language model. The large language model infers and determines the relationship type between the new knowledge and existing knowledge (addition, supplementation, update, or conflict), outputting the fused new knowledge content and the existing knowledge texts in the neighborhood set that are determined to be discarded. Finally, the knowledge update agent will perform chunking and vectorization processing on the merged new knowledge and write it into the vector index of the corresponding model knowledge base. At the same time, it will delete the existing knowledge text and its vector index that are determined to be eliminated, create a version snapshot, and return the operation result information to the UI interface module of the question-and-answer interface interaction layer for display.

[0132] Through the seven stages described above, this embodiment of the application achieves end-to-end closed-loop processing from user query request access to final answer display. Specifically, the four-agent closed-loop processing, orchestrated collaboratively by multiple agents in the knowledge base, is responsible for query understanding, retrieval execution, result fusion, and quality verification. The knowledge data management layer, with its knowledge base stored in domains according to satellite model, provides knowledge retrieval and storage services. The knowledge update agent independently performs incremental updates and returns the operation results for display. Throughout the process, retrieval and updating form a two-way collaborative closed loop: the knowledge generation and verification agent feeds back knowledge entry suggestions to the knowledge update agent, triggering incremental updates; after completing the update, the knowledge update agent returns the operation results for display, ensuring that new knowledge is promptly entered into the database and reflected in subsequent searches, achieving a dual-drive of retrieval and knowledge updating.

[0133] The multi-agent collaborative knowledge retrieval and generation method for multiple satellite models provided in this application has independent processing contexts for the knowledge task planning agent, knowledge document retrieval agent, knowledge result fusion agent, and knowledge generation verification agent. Each agent focuses on its own responsibility, avoiding information forgetting due to excessively long contexts and ensuring the processing quality of each decision-making stage. The knowledge task planning agent autonomously decides the retrieval strategy based on the semantic complexity of the query text and the number of satellite models to be queried, matching the optimal retrieval path for queries of different complexities. This allows simple queries to be responded to quickly with a lightweight process, and complex queries to be fully processed through multi-agent collaboration, shortening the average response time of compound queries. The knowledge generation verification agent performs confidence verification on the target knowledge text set, and triggers the knowledge task planning agent to regenerate the retrieval strategy when the verification fails, forming an autonomous optimization closed loop of "planning → retrieval → fusion → verification → reflection → replanning", improving the accuracy of cross-satellite model knowledge retrieval. The semantic similarity of the knowledge semantic association retrieval module is enhanced. The three-level processing of retrieval, cross-model association enhancement, and comprehensive score ranking and reordering enables the retrieval results to incorporate cross-model association information on the basis of semantic matching, effectively solving the semantic omission problem caused by different expressions of the same concept among multiple satellite models, and significantly improving the recall and accuracy of cross-model knowledge retrieval. By integrating knowledge results with intelligent agents to filter high-quality knowledge texts based on multi-dimensional information such as comprehensive semantic similarity, model association, and source knowledge base credibility, and combining the quantitative evaluation of overall confidence, a reliable knowledge foundation is provided for subsequent answer generation, improving the credibility of the final question-and-answer results. Through the automated incremental update of the knowledge update agent and the intelligent fusion reasoning mechanism based on a large language model, the average latency from knowledge generation to database entry is shortened from several days in traditional manual updates to several minutes. Each model knowledge base can be updated and maintained independently, significantly improving the timeliness of the knowledge base and the maintainability of the system. Through the "retrieval → verification → update" two-way collaborative closed loop formed by the knowledge base multi-agent collaborative orchestration and the knowledge update agent, new knowledge can be promptly entered into the database and reflected in subsequent retrievals, realizing a dual-drive of retrieval and knowledge update.

[0134] This application also provides a multi-satellite model knowledge retrieval and generation system based on multi-agent collaboration, which can be deployed on one or more processors.

[0135] In one possible implementation, the above method can be executed on a single processor. This processor can be a central processing unit (CPU), graphics processing unit (GPU), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other form of programmable logic device. The single processor runs a knowledge task planning agent, a knowledge document retrieval agent, a knowledge result fusion agent, and a knowledge generation and verification agent. These agents share the processor's computing resources through time-division multiplexing or pipelined parallelism. Data interaction between the agents occurs through inter-process communication or function calls.

[0136] In another possible implementation, the above method can be deployed across multiple processors. These processors can be integrated into the same electronic device or distributed across multiple electronic devices, communicating via wired or wireless network connections. For example, a knowledge task planning agent can run on a first processor, a knowledge document retrieval agent on a second processor, a knowledge result fusion agent on a third processor, and a knowledge generation and verification agent on a fourth processor. The processors communicate and collaborate through standardized data interfaces. Each processor independently executes the computational tasks of its respective agent, achieving parallel processing among the agents.

[0137] In another possible implementation, the knowledge updating agent and the semantic encoding module can be deployed on separate processors. The knowledge updating agent runs on the fifth processor, continuously monitoring new knowledge data and executing incremental update processes; the semantic encoding module runs on the sixth processor, providing unified semantic encoding services for all modules in the system that require text vector conversion. After the knowledge updating agent completes the knowledge base update, it returns the operation result information to the question-and-answer interface interaction layer for display through a standardized interface.

[0138] The deployment of multiple processors enhances the system's parallel processing capabilities and fault tolerance. When one processor fails, others can continue operating, ensuring the availability of core system functions. Simultaneously, each agent possesses an independent processing context. With multiple processors deployed, the context environments of each agent are completely isolated and do not interfere with each other, further avoiding information forgetting due to excessively long contexts and ensuring that each agent maintains high-quality decision-making within its respective domain of focus. Those skilled in the art can flexibly choose single-processor or multi-processor deployment methods based on the computing resources, response time requirements, and cost constraints of the actual application scenario; all such choices fall within the scope of this application.

[0139] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

[0140] The embodiments of this application have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of this application. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Without departing from the scope of this application, those skilled in the art can make various substitutions and modifications, all of which should fall within the scope of this application.

Claims

1. A method for knowledge retrieval and generation of multiple satellite models based on multi-agent collaboration, characterized in that, The method includes: In response to a query request containing query text and at least one satellite model to be queried, a retrieval strategy is generated by a knowledge task planning agent. The retrieval strategy includes a retrieval instruction and a retrieval database range. The retrieval instruction is used to retrieve knowledge text related to the query text and the at least one satellite model to be queried within the retrieval database range. The retrieval database range refers to at least one knowledge database corresponding to the at least one satellite model to be queried. Based on the retrieval instruction, a knowledge document retrieval agent is invoked to perform a retrieval in at least one knowledge database to obtain a candidate knowledge text set. The knowledge result fusion agent is invoked to filter the target knowledge text set based on the relevance of each candidate knowledge text in the candidate knowledge text set to the query text and the satellite model to be queried; The knowledge generation and verification agent is invoked to verify the overall confidence level of the target knowledge text set; If the verification fails, the knowledge generation verification agent generates reflection information to trigger the knowledge task planning agent to regenerate the retrieval strategy, and the above method is re-executed based on the regenerated retrieval strategy until the verification passes or the number of execution rounds of the method reaches the preset number of rounds, and the analysis result is generated based on the target knowledge text set.

2. The method according to claim 1, characterized in that, In response to a query request containing query text and at least one satellite model to be queried, the knowledge task planning agent generates a retrieval strategy including: The knowledge task planning agent analyzes the semantic complexity of the query text and the number of at least one satellite model to be queried. When the at least one satellite model to be queried is a single one, a single-database retrieval strategy is generated by the knowledge task planning agent. The single-database retrieval strategy is that the retrieval database is a single knowledge database corresponding to the satellite model to be queried. The retrieval instruction is used to perform a retrieval in the single knowledge database. When there are multiple types of at least one satellite to be queried, a multi-database parallel retrieval strategy is generated by the knowledge task planning agent. The multi-database parallel retrieval strategy is that the retrieval database range is multiple knowledge databases that correspond one-to-one with the at least one satellite to be queried. The retrieval instruction is used to perform retrieval in parallel in the multiple knowledge databases. If the semantic complexity of the query text exceeds a first preset threshold or fails verification, the knowledge task planning agent generates a retrieval instruction for an iterative deepening retrieval strategy. The iterative deepening retrieval strategy includes a knowledge database corresponding to at least one satellite model to be queried. The retrieval instruction is used to perform multiple rounds of retrieval in conjunction with the query text.

3. The method according to claim 1, characterized in that, Based on the retrieval instruction, the knowledge document retrieval agent is invoked to perform a retrieval in at least one knowledge database to obtain a candidate knowledge text set, including: The knowledge document retrieval agent performs an approximate nearest neighbor retrieval of the query text in at least one knowledge database to obtain an initial candidate text set and the semantic similarity between each initial candidate knowledge text in the initial candidate text set and the query text. The knowledge document retrieval agent matches the satellite model tags of each initial candidate knowledge text with the at least one satellite model to be queried, thereby determining the correlation between each initial candidate knowledge text and the at least one satellite model to be queried. The knowledge document retrieval agent determines the comprehensive score of each initial candidate knowledge text based on the semantic similarity and the relevance. Based on the comprehensive score, a number of initial candidate knowledge texts with the highest ranking are selected from the initial candidate text set to form the candidate knowledge text set.

4. The method according to claim 3, characterized in that, The step of matching the satellite model tags of each initial candidate knowledge text with the at least one satellite model to be queried through the knowledge document retrieval agent to determine the correlation between each initial candidate knowledge text and the at least one satellite model to be queried includes: The knowledge document retrieval agent invokes the cross-model relationship graph to query the association path between the satellite model tags of each of the initial candidate knowledge texts and the at least one satellite model to be queried. The cross-model relationship graph stores the association relationship between different satellite models in a graph structure. If the associated path is found, the correlation between the initial candidate knowledge text and the at least one satellite model to be queried is determined to be a first value; If the associated path is not found, the correlation between the initial candidate knowledge text and the at least one satellite model to be queried is determined to be a second value.

5. The method according to claim 1, characterized in that, The knowledge result fusion agent, based on the relevance of each candidate knowledge text in the candidate knowledge text set to the query text and the satellite model to be queried, filters the target knowledge text set, including: The knowledge result fusion agent determines the fusion score of each candidate knowledge text based on the semantic similarity of each candidate knowledge text to the query text, the correlation between each candidate knowledge text and the at least one satellite model to be queried, and the preset credibility of the knowledge database to which each candidate knowledge text belongs. Based on the descending order of the fusion scores, the top-ranked candidate knowledge texts are selected as the target knowledge text set.

6. The method according to claim 5, characterized in that, The method further includes: The overall confidence level is obtained by calculating the proportion of text segments whose fusion scores reach a second preset threshold in the target knowledge text set.

7. The method according to claim 1, characterized in that, The method further includes: In response to acquiring new knowledge data, the knowledge updating agent retrieves multiple existing knowledge texts that are semantically adjacent to the new knowledge data from the knowledge database to form a neighborhood set. Based on the relationship type between the new knowledge data and the existing knowledge text, the knowledge update agent fuses the new knowledge data with the neighborhood set to obtain the fused new knowledge content and the existing knowledge text in the neighborhood set that is determined to be eliminated. The fused new knowledge content is written into the knowledge database of the corresponding satellite model, and the existing knowledge text that is determined to be obsolete is deleted.

8. The method according to claim 7, characterized in that, The method further includes: Obtain the new knowledge data uploaded by the user; and / or The knowledge entry suggestion data generated by the knowledge generation and verification agent is obtained as the new knowledge data. The knowledge entry suggestion data includes the query text, the at least one satellite model to be queried, and blank domain knowledge data. The blank domain knowledge data is the knowledge that the knowledge database lacks, determined by the knowledge generation and verification agent when the verification fails, based on the difference between the knowledge content required to answer the query text and the knowledge content provided by the target knowledge text set.

9. The method according to claim 8, characterized in that, The method further includes: During the iteration of the regenerated retrieval strategy, if the knowledge content required for the query text fails to return relevant knowledge text after using at least two different retrieval strategies, then the knowledge content is confirmed to be blank domain knowledge data.

10. The method according to claim 1, characterized in that, The knowledge database includes a general knowledge base and a model-specific knowledge base corresponding to each satellite model; the general knowledge base stores aerospace knowledge that is common to all satellite models, and the model-specific knowledge base stores domain knowledge corresponding to each satellite model.

Images