Intelligent Big Data Retrieval Methods and Systems Integrating Multimodal Large Models

By integrating intelligent big data retrieval methods with multimodal large models, a set of retrieval configurations specific to R&D scenarios is generated, which solves the problem of insufficient adaptability of traditional data retrieval methods in complex R&D scenarios, and realizes efficient and accurate data acquisition and utilization, thereby improving R&D efficiency.

CN123115382APending Publication Date: 2026-11-17GUANGZHOU ESLITE NETWORK TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611308332.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-27
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Traditional data retrieval methods cannot adapt to the complex and ever-changing needs of R&D scenarios, resulting in inaccurate and incomplete retrieval results, lack of flexibility and adaptability, and difficulty in meeting the requirements for accurate data acquisition and efficient utilization during the R&D process.

Method used

An intelligent big data retrieval method integrating multimodal large models is adopted. By acquiring dynamic scene elements from data acquisition terminals, a set of retrieval configurations specific to the R&D scenario is generated. Combined with a multi-path collaborative retrieval method, real-time adjustment and optimization are achieved to generate the final retrieval results.

Benefits of technology

It improves the accuracy, comprehensiveness, and real-time performance of data retrieval, enhances R&D efficiency and quality, and ensures a high degree of consistency between retrieval strategies and R&D scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115382A_ABST
    Figure CN123115382A_ABST
Patent Text Reader

Abstract

This application provides an intelligent big data retrieval method and system integrating a multimodal large model, relating to the field of big data retrieval technology. First, dynamic scene elements from a customized development and research scenario are acquired by a data acquisition terminal and input into a multimodal large model to generate a retrieval configuration set specific to the development scenario. Next, real-time state features are input through the multimodal large model's interactive interface to update the configuration set. Then, a multi-path collaborative retrieval is initiated based on the updated configuration set to obtain an initial data set. Finally, the initial data set is input into the multimodal large model for scenario-based filtering and optimization to generate a final retrieval result set. This invention can adjust the retrieval strategy in real time according to the dynamic changes of the development scenario, improving the accuracy and comprehensiveness of data retrieval.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of big data retrieval technology, and more specifically, to an intelligent big data retrieval method and system that integrates multimodal big models. Background Technology

[0002] In customized development scenarios for integrated data acquisition terminal applications, data retrieval is a crucial step in obtaining key information and supporting R&D decisions. However, existing data retrieval methods often employ a single mode, making it difficult to adapt to the complex and ever-changing needs of R&D scenarios. Different development stages have different data requirements, variations in terminal hardware characteristics affect data acquisition and processing methods, and application adaptation and data association requirements also differ. Traditional retrieval methods cannot comprehensively consider these dynamic scenario elements, resulting in inaccurate and incomplete retrieval results, failing to meet the needs for precise data acquisition and efficient utilization during the R&D process. Furthermore, traditional methods lack flexibility and adaptability when facing real-time changes in the R&D scenario, failing to adjust retrieval strategies in a timely manner, leading to low retrieval efficiency and hindering the rapid advancement of R&D work. Summary of the Invention

[0003] In view of this, the purpose of this application is to provide an intelligent big data retrieval method and system that integrates multimodal big models.

[0004] According to a first aspect of this application, an intelligent big data retrieval method integrating multimodal large models is provided, the method comprising: The dynamic scene elements in the customized development and research scenario of the integrated application of data acquisition terminal are obtained. The dynamic scene elements include development stage characteristics, terminal hardware characteristics, application adaptation requirements and data association requirements. The dynamic scene elements are input into a deep-trained multimodal large model to generate a search configuration set specifically for the R&D scene. The search configuration set specifically for the R&D scene includes the definition of search dimensions, the method of element association, and the interaction interface of the multimodal large model. Through the multimodal large model interaction interface, the real-time state features of the R&D scenario are input into the multimodal large model. Based on the adjustment instructions output by the multimodal large model, the search configuration set dedicated to the R&D scenario is updated to obtain the adapted and adjusted search configuration set dedicated to the R&D scenario. Based on the adapted and adjusted R&D scenario-specific retrieval configuration set, a multi-path collaborative retrieval method is initiated to perform targeted data mining operations on the R&D-related big data resource library, thereby obtaining the initial data set corresponding to the targeted data mining operation results. The initial dataset is input into the multimodal large model, and the multimodal large model performs scenario-based filtering and optimization on the initial dataset to generate the final retrieval result set.

[0005] According to a second aspect of this application, an intelligent big data retrieval system integrating multimodal big models is provided. The intelligent big data retrieval system integrating multimodal big models includes a processor and a readable storage medium. The readable storage medium stores a program that, when executed by the processor, implements the aforementioned intelligent big data retrieval method integrating multimodal big models.

[0006] Based on any of the above aspects, by acquiring dynamic scene elements from the customized development and research scenario of the data acquisition terminal and integrating them with the application, and inputting them into a deeply trained multimodal large model, a dedicated search configuration set highly adapted to the development scenario can be generated. This dedicated search configuration set comprehensively considers multiple factors such as development stage characteristics, terminal hardware characteristics, application adaptation requirements, and data association needs, making the definition of search dimensions, element association methods, and the setting of multimodal large model interaction interfaces more accurate and reasonable. By inputting real-time state characteristics of the development scenario, the multimodal large model can output adjustment instructions and update the search configuration set in a timely manner, ensuring that the search strategy always remains highly consistent with the actual situation of the development scenario. Activating the multi-path collaborative search method can fully utilize the big data resource library to perform targeted data mining operations and obtain an initial data set. Finally, by performing scenario-based filtering and optimization processing on the initial data set through the multimodal large model, a final search result set that meets the development needs can be generated, greatly improving the accuracy, comprehensiveness, and real-time performance of data retrieval, and effectively improving development efficiency and quality. Attached Figure Description

[0007] Figure 1 A flowchart illustrating the intelligent big data retrieval method integrating multimodal large models provided in this application embodiment is shown. Figure 2 This paper illustrates a schematic diagram of the component structure of an intelligent big data retrieval system that integrates multimodal big models, as provided in an embodiment of this application. Detailed Implementation

[0008] Figure 1 A flowchart illustrating the intelligent big data retrieval method integrating multimodal large models provided in this application is shown, and the detailed steps are described below.

[0009] Step S110: Obtain dynamic scene elements in the customized development and research scenario of the data acquisition terminal fusion application. The dynamic scene elements include development stage characteristics, terminal hardware characteristics, application adaptation requirements, and data association requirements.

[0010] In this embodiment, dynamic scene elements are acquired through a multi-source data acquisition component deployed in the R&D environment. Development phase characteristics are extracted through the API interface of the R&D project management platform, covering structured information such as the current development phase identifier, phase start time, planned completion time, percentage of completed tasks, and a list of unresolved issues. Terminal hardware characteristics are obtained through real-time scanning of the data acquisition terminal using hardware testing tools, including parameters such as hardware model, core processor architecture, memory capacity, storage media type, sensor configuration list, and communication module specifications. Application adaptation requirements are extracted from requirement entries in the requirement management system. Each requirement includes the name of the target application, application version range, functional module adaptation requirements, performance index requirements, and compatibility constraints. Data association requirements are parsed from the metadata information of the R&D database. The metadata records the relationship definition between different data entities, involving information such as data entity names, attribute fields, association types, and association fields. During the data acquisition process, sensitive data involving R&D personnel operation behavior and device login credentials are processed using data anonymization technology. Sensitive fields are replaced with specific identifiers and transmitted to the processing node through an encrypted transmission channel to ensure data security.

[0011] Step S120: Input the dynamic scene elements into a deep-trained multimodal large model to generate a search configuration set specifically for the R&D scene. The search configuration set specifically for the R&D scene includes the search dimension definition, element association method and multimodal large model interaction interface.

[0012] In this embodiment, dynamic scene elements need to be preprocessed before being input into the multimodal large model. First, development stage features, terminal hardware characteristics, application adaptation requirements, and data association needs are converted into key-value pair structured data, supplementing the necessary fields of the semi-structured data. Next, the structured data is feature-standardized, unifying the time format to year, month, day, hour, minute, and second, and using international standard units for hardware parameters. The preprocessed data is input through the multimodal large model's input layer. The text feature extraction submodule of the input layer uses an attention-based text encoder to process text-type elements, while the structured data processing submodule converts the structured data into fixed-dimensional feature vectors through an embedding layer. The intermediate layer of the multimodal large model consists of a cross-attention module and a feedforward neural network layer. The cross-attention module captures the relationships between elements of different modalities, while the feedforward neural network layer deeply processes and fuses features. The output layer contains three parallel branches, generating the retrieval dimension definition, element association method, and multimodal large model interaction interface, respectively. Each branch outputs a probability distribution through a softmax activation function, which is then decoded to obtain specific configuration parameters.

[0013] Step S121: Extract the development stage features from the dynamic scene elements, break them down to obtain development process node information, and each development process node information corresponds to a specific implementation link of the customized development of the data acquisition terminal fusion application.

[0014] In this embodiment, feature extraction during the development phase is performed using a development phase feature extractor. First, the development phase feature data is segmented into individual word units. Then, based on a predefined keyword library for development process nodes, the segmentation results are matched to identify the names of the development process nodes. Next, the nodes are ordered according to their logical sequence to determine their position within the development process. For each node, implementation description information is extracted, including task objectives, involved development roles, required resource support, and output requirements.

[0015] Step S122: Extract the terminal hardware characteristic information from the dynamic scene elements, and decompose it to obtain hardware structure association information. Each piece of hardware structure association information corresponds to a hardware component and functional implementation path of the data acquisition terminal.

[0016] In this embodiment, the terminal hardware characteristic information is processed by a hardware characteristic analysis engine. The engine first categorizes the hardware characteristic information according to the hardware system hierarchy (core processing layer, storage layer, sensing layer, communication layer, power management layer, etc.). Within each layer, it identifies specific hardware components, such as the processor model, number of cores, and clock speed range in the core processing layer, and various sensors in the sensing layer. For each hardware component, it analyzes its functional implementation path, covering its working principle description, connection methods with other hardware (bus type, interface specifications), data transmission protocol, control instruction set, etc.

[0017] Step S123: Extract the application adaptation requirement information from the dynamic scene elements, and decompose it to obtain application compatibility association information. Each application compatibility association information corresponds to the adaptation direction and association logic between the data acquisition terminal and the target application.

[0018] In this embodiment, application adaptation request information is extracted by the adaptation request analysis module. The module uses semantic role labeling methods from natural language processing to understand the request information, identifying the subject (data acquisition terminal), object (target application), action (adaptation), and constraints (operating system version, API interface version, data format, etc.). Based on semantic understanding, adaptation directions are determined, including functional modules, data interfaces, UI interfaces, performance metrics, etc. For each adaptation direction, association logic is constructed to describe the mapping relationship between specific attributes of the data acquisition terminal and specific requirements of the target application, such as data format field correspondence, data type conversion rules, data length limits, etc.

[0019] Step S124: Extract the data association requirement information from the dynamic scene elements, and break it down to obtain data interaction association information. Each data interaction association information corresponds to the flow and collaboration logic of different types of data in the R&D process.

[0020] In this embodiment, data association requirement information is extracted using a data requirement analysis tool. The tool first identifies data entities in the R&D process, such as design document data, test case data, code file data, hardware parameter data, and user feedback data. It then extracts the attribute information of each data entity, including data entity ID, name, type, size, creation time, update time, and the corresponding R&D stage. The tool analyzes the flow relationships between different data entities, covering the source of generation, flow path, recipient, processing method, and storage location. Based on these flow relationships, a collaborative logic is constructed to describe the collaborative working methods of data entities in R&D. For example, after the design document data review is passed, code development data is generated; after the code development data is completed, it is associated with test case data for testing; and the test results are fed back to the design document data for iterative optimization.

[0021] Step S125: Input the development process node information, the hardware structure association information, the application compatibility association information and the data interaction association information into the deep-trained multimodal large model, perform multi-dimensional association modeling operation of R&D scenario, and obtain the result of multi-dimensional association modeling operation of R&D scenario.

[0022] In this embodiment, the four types of information input into the multimodal large model need to be vectorized. Development process node information is converted into feature vectors through a process node embedding layer, considering node order relationships and attribute features during embedding. Hardware structure association information is converted through a hardware structure embedding layer, considering the hierarchical structure and connection relationships of hardware components. Application compatibility association information is converted through an application compatibility embedding layer, considering adaptation direction and related logical semantic features. Data interaction association information is converted through a data interaction embedding layer, considering data entity relationships and flow path features. The intramodal attention module of the multimodal large model's intermediate processing layer captures the feature dependencies within each modality, while the cross-modal attention module uses one modality feature as the query vector and another as the key-value vector, capturing intermodal relationships through attention calculation. The fused multimodal feature vectors are input into a graph neural network layer, mapped to graph nodes. A research and development scenario association graph is constructed based on association weights, and node features are updated through graph convolution operations. Finally, the node features and association graph structure are output as the modeling result.

[0023] Step S126: Based on the results of the multi-dimensional association modeling operation of the R&D scenario, the core search dimensions specific to the R&D scenario are mapped, and each core search dimension corresponds to a key R&D requirement direction.

[0024] In this embodiment, the process starts with the association graph structure and graph node feature vectors from the association modeling results. A density-based clustering algorithm is used to cluster the graph node feature vectors, dividing similar nodes into different clusters, and calculating the central feature vector of each cluster. The central feature vectors are then input into a pre-trained text generation model to generate semantic description text, extracting a set of core keywords. Based on domain knowledge, it is determined whether a cluster constitutes an independent core retrieval dimension. The boundary range of the dimension is clearly defined, specifying the included sub-topics, data types, and associated R&D stages.

[0025] Step S1261: Extract the development process node association information from the multi-dimensional association modeling operation results of the R&D scenario, identify the key data requirement direction corresponding to each development process node, and each key data requirement direction corresponds to the data type that the development process node must obtain during the R&D process.

[0026] In this embodiment, the association information of development process nodes is extracted from the association graph structure of the association modeling results. The association graph is traversed to identify the development process node graph nodes and their associated edges, and data entity type graph nodes are filtered. Based on the data entity attribute characteristics and their role in the R&D process, the key data requirement directions for each development process node are determined, and the data types that the node must acquire are described.

[0027] Step S1262: Extract the hardware structure association information from the multi-dimensional association modeling operation results of the R&D scenario, identify the technical parameter data requirements corresponding to each hardware component, and identify the data content required for hardware debugging and optimization for each technical parameter data requirement.

[0028] In this embodiment, hardware structure association information is extracted from the association graph structure of the association modeling results. The association graph is traversed to locate hardware component graph nodes, their attribute characteristics and association relationships are analyzed, and associated technical parameter data entity graph nodes are extracted. Based on the technical parameter data entity graph node information, the technical parameter data requirements of the hardware components are identified, and the parameter types, ranges, and accuracy requirements for hardware debugging and optimization are described.

[0029] Step S1263: Extract application compatibility association information from the multi-dimensional association modeling operation results of the R&D scenario, identify the adaptation data requirements in the process of data acquisition terminal and target application docking, and each adaptation data requirement corresponds to the data support content to ensure the normal operation of the application.

[0030] In this embodiment, application compatibility association information is extracted from the association graph structure of the association modeling results. Edges representing the connection relationship between the data acquisition terminal and the target application are filtered, and attribute information such as the connection interface type, data transmission protocol, and compatibility constraints of the edges are analyzed to locate the adaptive data entity graph nodes participating in the connection. Based on the adaptive data entity graph node information, adaptive data requirements are identified, including connection interface definition data, compatibility test data, and adaptive driver data.

[0031] Step S1264: Extract the data interaction association information from the multi-dimensional association modeling operation results of the R&D scenario, identify the collaborative data requirements in different types of data flow processes, and ensure that each collaborative data requirement corresponds to data association content that ensures effective data interaction.

[0032] In this embodiment, data interaction association information is extracted from the association graph structure of the association modeling results. The edges representing data flow relationships are traversed, and attribute information such as flow direction, conditions, and frequency is analyzed. Based on the edge attribute information and the characteristics of the connected data entity graph nodes, collaborative data requirements are identified, covering data flow rule definitions, format conversion rules, consistency verification rules, and conflict resolution strategies.

[0033] Step S1265: Input the identified key data demand directions, technical parameter data demands, adaptation data demands, and collaborative data demands into the deep-trained multimodal large model, and perform a similar merging operation on the demand dimensions to obtain the merged demand dimensions.

[0034] In this embodiment, the four types of requirements are converted into a unified text description format containing fields such as requirement name, description, data type involved, and related R&D stages, and then input into the multimodal large model requirement merging module. The requirement merging module extracts the textual features of the requirements, calculates the cosine similarity of the requirement feature vectors, merges requirements with similarity higher than a threshold, integrates the requirement descriptions, retains common and unique features, and names the merged requirements with a unified name. After merging, a consistency check is performed, and requirement inconsistencies are resolved by referring to the results of the association modeling.

[0035] Step S1266: Based on the overall development goals of the R&D scenario, perform an importance ranking operation on the merged requirement dimensions to obtain the ranked requirement dimensions, and determine the priority of each requirement dimension in the retrieval process.

[0036] In this embodiment, the overall development goals are extracted from the R&D project objective description document, including core functions, performance indicators, development cycle, and cost control objectives. An evaluation indicator system is established, including indicators such as the correlation between requirement dimensions and core functional objectives, the weight of their impact on performance indicators, their proportion in the critical path of the development cycle, and their proportion of implementation cost. Experts are organized to score each requirement dimension on each indicator according to high, medium, and low levels. A weighted summation method is used to calculate the comprehensive score, and the requirement dimensions are sorted from high to low scores to determine the retrieval priority.

[0037] Step S1267: Define a defined coverage area for each of the sorted demand dimensions, and define the data types and data characteristics included in each demand dimension.

[0038] In this embodiment, based on the requirement description and related R&D stages, and referring to the R&D database metadata classification standard, the data types included in the coverage are defined, such as document type, code type, data type, multimedia type and their subclasses. Data characteristics are clearly defined, including format characteristics (text, binary, table format), content characteristics (keywords, core concepts, technical terms), source characteristics (generating department, time, tool), and association characteristics (related project stage, hardware module, application function), etc.

[0039] Step S1268: Mark the potential correlation points between each demand dimension and other demand dimensions in the demand dimensions with a defined coverage, and obtain the demand dimensions with marked potential correlation points.

[0040] In this embodiment, based on the association graph structure derived from the association modeling results, the connection edges between clusters corresponding to each requirement dimension are analyzed. Requirement dimensions corresponding to edges with association strength exceeding a threshold are marked as potentially related points. The association type of these points (data dependency, process sequence, functional complementarity), the data entities involved, and triggering conditions are described.

[0041] Step S1269: Based on the coverage, priority, and potential related points of the demand dimensions marked with potential related points, generate the retrieval execution standard for each demand dimension, and define the retrieval method and judgment basis for the data of that dimension.

[0042] In this embodiment, the retrieval method is determined based on the characteristics of the data types covered by the demand dimension. For document data, a combination of keyword and semantic retrieval is used; for code data, a combination of code features and semantic understanding is used; and for hardware parameter data, structured query is employed. The judgment criteria are based on priority and potential correlation points. High-priority dimension data is returned first, and correlation dimension retrieval is triggered when retrieving dimension data with potential correlation points. The criteria for determining whether data belongs to a given dimension include whether it contains core keywords, whether the data type is within the coverage scope, and whether the source and correlation features match.

[0043] Step S12610: Define the requirement dimension with retrieval execution standard as the core retrieval dimension exclusive to the R&D scenario. Each core retrieval dimension exclusive to the R&D scenario can accurately correspond to a core data requirement of the R&D scenario.

[0044] In this embodiment, the completeness and clarity of the retrieval execution standards for each requirement dimension are checked. After confirming that there are no errors, they are formally defined as the core retrieval dimensions, assigned unique identifiers, and associated with the cluster identifiers in the association modeling results to ensure that they can accurately correspond to the core R&D data requirements.

[0045] Step S127: Based on the results of the multi-dimensional association modeling operation of the R&D scenario, generate the interaction association logic between the core search dimensions and define the collaborative search rules for data of different dimensions.

[0046] In this embodiment, edges between clusters corresponding to the core retrieval dimensions are extracted from the association graph structure of the association modeling results. The association type (data dependency, process collaboration, functional complementarity), strength, and direction are analyzed. The collaborative retrieval order is determined based on the association type and direction, and the association strength affects the collaborative retrieval trigger threshold. Data sharing conditions are defined, clarifying the timing, format requirements, and scope limitations for data sharing. Association judgment criteria are set, determining the basis for data association identification, strength description, and result processing methods.

[0047] Step S128: Based on the development interface standard of the data acquisition terminal, generate the multimodal large model interaction interface between the multimodal large model and the retrieval execution link, and define the data transmission format and linkage triggering conditions.

[0048] In this embodiment, the development interface standards are obtained by consulting the hardware development manual and interface specification documents, including interface types, communication protocols, data encoding methods, call frequency limits, authentication and authorization methods, etc. The data transmission format is defined: text data uses JSON format containing fields such as request ID, core search dimension identifier, and search parameters; binary data is Base64 encoded and embedded in JSON. Linkage triggering conditions are set, such as dynamic scene element updates triggering a search configuration update request, search completion triggering a result return request, and insufficient or ambiguous data triggering a model-assisted decision-making request. An exception handling mechanism is also defined.

[0049] Step S129: Integrate the core search dimensions, the interaction logic between the core search dimensions, and the multimodal large model interaction interface to form the basic structure of a search configuration set specific to the R&D scenario.

[0050] In this embodiment, a hierarchical JSON object format is used to organize the basic structure. The top layer contains fields for the core retrieval dimension set, the interaction and association logic set, and the multimodal large model interaction interface definition. The core retrieval dimension set array elements include subfields such as dimension identifier, name, coverage, priority, and retrieval execution criteria; the interaction and association logic set array elements include subfields such as source / target dimension identifier, association type, strength, collaborative retrieval order, data sharing conditions, and association judgment criteria; the interaction interface definition includes fields such as interface type, communication protocol, data transmission format template, linkage trigger condition list, and authentication / authorization method. Consistency in referencing is ensured during integration, and syntax validation is performed.

[0051] Step S1210: Based on historical search case data of the R&D scenario, perform scenario adaptability enhancement operation on the basic structure of the R&D scenario-specific search configuration set to obtain the R&D scenario-specific search configuration set. Each component of the R&D scenario-specific search configuration set corresponds to the actual needs of the R&D scenario.

[0052] In this embodiment, historical search case data, including search request parameters, result lists, user reviews, problems, and solutions, is extracted from the search log database. Historical data is analyzed to identify infrastructure adaptation issues, such as low relevance of results in core search dimensions and conflicts in collaborative search rules. Configuration adjustments are made to address these issues, optimizing the search execution standards for core search dimensions, the collaborative rules for interactive logic, and the data transmission format and triggering conditions of the interactive interface. After adjustments, the effectiveness is verified through simulated search tests, resulting in a final search configuration set specific to the R&D scenario.

[0053] Step S130: Input the real-time state features of the R&D scenario into the multimodal large model through the multimodal large model interaction interface, and update the R&D scenario-specific retrieval configuration set based on the adjustment instructions output by the multimodal large model to obtain the adapted and adjusted R&D scenario-specific retrieval configuration set.

[0054] In this embodiment, the real-time status characteristics of the R&D scenario are collected by the R&D environment real-time monitoring component every five minutes. These characteristics are then encapsulated according to the interaction interface format and sent to the multimodal large model via a specified communication protocol. The multimodal large model extracts and analyzes the real-time status characteristics, compares them with the currently stored search configuration, identifies the configuration items that need adjustment, and generates an adjustment instruction containing fields such as the target configuration item identifier, adjustment type, and content. The search configuration management module parses the instruction, executes the adjustment operation, verifies the completeness and consistency of the updated configuration, saves it as an adapted search configuration set, updates the version number, and records the adjustment log.

[0055] Step S131: Through the R&D status monitoring stage of the data acquisition terminal, collect real-time development progress information of the R&D scenario. The real-time development progress information includes the currently executed development process node and the completion status of the development process node.

[0056] In this embodiment, the development process monitoring component of the R&D status monitoring stage is synchronized with the project management platform in real time. It periodically calls the progress query API to obtain the current development process node information, including node identifier, name, and planned / actual start time. The percentage of completed subtasks under a node is used as the completion percentage. Supplementary information on the completion status, such as the number of issue reports, the number of resolved issues, and the number of blocked tasks, is collected and encapsulated into a JSON object containing fields such as node identifier, name, completion percentage, number of issues, and number of blocked tasks.

[0057] Step S132: Through the R&D status monitoring stage of the data acquisition terminal, collect real-time hardware adaptation information of the R&D scenario. The real-time hardware adaptation information includes the current hardware debugging results and hardware parameter adjustment status of the data acquisition terminal.

[0058] In this embodiment, the hardware monitoring submodule of the R&D status monitoring stage is connected to the hardware debugging interface of the data acquisition terminal. It sends parameter reading commands to obtain hardware debugging results, including the working status of hardware components, real-time sensor data, communication module signal strength, power module voltage and current, etc. It monitors hardware configuration file change records and debugging tool operation logs to obtain parameter adjustment information, and integrates them into structured data containing fields such as hardware component identifiers, working status, current values ​​of key parameters, recent adjustment records, and debugging result summaries.

[0059] Step S133: Through the R&D status monitoring stage of the data acquisition terminal, collect real-time application docking information of the R&D scenario. The real-time application docking information includes the temporary docking effect between the data acquisition terminal and the target application and feedback on adaptation issues.

[0060] In this embodiment, the application integration monitoring component in the R&D status monitoring stage simulates user operations or executes the integration process in a test environment through automated test scripts to evaluate the temporary integration effect, including indicators such as integration success rate, data transmission latency, data integrity, and function call return codes. Error logs, exception prompts, and warning messages during the integration process are captured as adaptation issue feedback. Each feedback includes the occurrence time, operation steps, error type, description, and stack trace information. The integrated performance indicators and issue feedback are structured data containing fields such as target application identifier, integration operation type, success rate, average latency, data integrity verification result, and issue feedback list.

[0061] Step S134: Input the real-time development progress information, the real-time hardware adaptation information, and the real-time application docking information into the deep-trained multimodal large model through the multimodal large model interaction interface. Based on the deep-trained multimodal large model, perform feature extraction operations on the real-time development progress information, the real-time hardware adaptation information, and the real-time application docking information to obtain the real-time status features of the R&D scenario. The real-time status features of the R&D scenario include the current R&D stage identifier, hardware configuration parameters, and application interface status.

[0062] In this embodiment, the three types of real-time information are encapsulated according to the interaction interface format and sent to the multimodal large model input layer via HTTP / HTTPS protocol. The three modal processing submodules of the input layer process the following: real-time development progress information is converted into feature vectors via an embedding layer; structured parameter embedding processing of real-time hardware adaptation information is performed, where the original sensor data is extracted for temporal features via a one-dimensional convolutional layer and then concatenated into a hardware adaptation feature vector; and the problem feedback text of real-time application integration information is encoded by a BERT text encoder, and after evaluation metric embedding processing, it is concatenated into an application integration feature vector. The multimodal fusion layer uses a cross-attention mechanism to fuse the three feature vectors, inputs them to the feature classification layer, and outputs real-time status features containing the current development stage identifier, hardware configuration parameters, and application interface status via a fully connected neural network.

[0063] Step S135: Perform correlation analysis on the real-time status features of the R&D scenario and the existing components of the R&D scenario-specific retrieval configuration set to identify the core retrieval dimensions and interaction logic in the R&D scenario-specific retrieval configuration set that do not match the real-time status features of the R&D scenario.

[0064] In this embodiment, a mapping relationship is established between real-time status features and the components of the retrieval configuration. The priority of the core retrieval dimensions is identified based on the current development stage, the coverage of the core retrieval dimensions is associated with hardware configuration parameters, and the interaction logic coordination rules are associated with application interface status. Mismatches are identified by comparing the mapping relationships, such as low priority of the "test case design" dimension during the testing phase, failure to update the corresponding dimension coverage after adding a new sensor, or failure to adjust the relevant dimension coordination rules when the application interface is abnormal. The identifiers and specific manifestations of the mismatched dimensions and associated logic are recorded.

[0065] Step S136: Based on the current R&D stage identifier and hardware configuration parameters in the real-time status features of the R&D scenario, update the coverage parameters and priority values ​​of the search core dimensions that differ in the R&D scenario-specific search configuration set to obtain the updated search core dimensions.

[0066] In this embodiment, the reasonable priority of each core search dimension at the current stage is determined based on the current R&D stage identifier and R&D process specifications, and the priority values ​​of dimensions with differences are adjusted. According to changes in hardware configuration parameters, the coverage of the corresponding core search dimensions is updated; for example, if a new sensor is added, the sensor data type and other information are added to the relevant dimension coverage. The "priority" and "coverage description" fields of the corresponding dimension in the configuration set are updated, and the reference relationship between the updated dimension and other configuration items is checked.

[0067] Step S137: Based on the application interface status in the real-time status characteristics of the R&D scenario, adjust the collaborative retrieval rules of the interaction association logic that differ in the R&D scenario-specific retrieval configuration set to obtain the adjusted interaction association logic.

[0068] In this embodiment, the application interface status includes information such as connection status, data transmission status, and error code statistics. The collaborative retrieval rules of the interaction association logic are adjusted according to the interface status. For example, when the interface frequently encounters errors, the retrieval priority of relevant debugging data dimensions is increased, and data sharing conditions and association judgment criteria are adjusted to adapt to changes in the interface status.

[0069] Step S1371: Extract the adaptation requirement change information from the real-time status features of the R&D scenario, and identify the core retrieval dimensions and related logic involved in the adaptation requirement change information.

[0070] In this embodiment, the current application interface status is compared with that of the previous cycle to identify status change characteristics and analyze the changes in adaptation requirements reflected. For example, frequent occurrences of communication error codes reflect an increased demand for communication protocol adaptation and error handling data retrieval. Based on these changes in requirements, the relevant core retrieval dimensions and associated logic are determined.

[0071] Step S1372: Analyze the interaction and association logic that differs in the R&D scenario-specific search configuration set, identify the specific content in the original collaborative search rules that conflicts with the information on the change in adaptation requirements, analyze the reasons for the conflict, and distinguish between issues related to search order settings, data sharing condition definitions, or association judgment criteria settings.

[0072] In this embodiment, differing interactive association logic is analyzed, and collaborative retrieval order, data sharing conditions, and association judgment criteria are extracted. These are compared with information on changes in adaptation requirements to identify conflicting content. The causes of the conflicts are analyzed to distinguish whether the issues lie in the retrieval order, data sharing conditions, or association judgment criteria.

[0073] Step S1373: Based on the specific reasons for the conflict and the specific requirements in the adaptation requirement change information, redesign the collaborative retrieval order between the relevant core retrieval dimensions to obtain the redesigned collaborative retrieval order, which is synchronized with the current R&D progress.

[0074] In this embodiment, the collaborative retrieval order is redesigned based on the priority requirements of the adaptation needs change information and the development progress determined by the current development stage identifier. For example, during the interface debugging stage, error handling and connection recovery strategy data are retrieved first and represented in the form of an ordered list.

[0075] Step S1374: Redefine the data sharing conditions between different core search dimensions, determine the triggering time, data format requirements and sharing scope restrictions for data sharing, and obtain the redefined data sharing conditions, which meet the current adaptation requirements.

[0076] In this embodiment, the data sharing trigger timing is set according to the interface status change events in the adaptation requirement change information, such as triggering connection recovery strategy data sharing when the interface is disconnected. The data format requirements are determined based on the data transmission format of the interaction interface and the coverage of the receiver dimension. The sharing scope limitation is set according to data sensitivity and receiver needs, and expressed in the form of a combination of conditional expressions and descriptive text.

[0077] Step S1375: Reset the association judgment criteria between different core search dimensions, determine the identification basis of data association, the description of association strength and the processing method of association results, and obtain the reset association judgment criteria. The reset association judgment criteria improve the adaptability of association logic.

[0078] In this embodiment, new data types or fields are added to update the data association identification criteria based on changes in adaptation requirements. For example, if a new error code type is added, a matching rule for that field is added. The association strength is expressed using a hierarchical system or a numerical range, and the processing method is determined based on the association strength. For example, strongly associated data is automatically merged, moderately associated data prompts the user for confirmation, and weakly associated data only records the relationship.

[0079] Step S1376: Input the redesigned collaborative retrieval order, the redefined data sharing conditions, and the redefined association judgment criteria into the deep-trained multimodal large model, and perform a logical validity verification operation in combination with the historical interaction data of the R&D scenario to obtain the logical validity verification operation result.

[0080] In this embodiment, the multimodal large model compares the newly designed collaborative retrieval order, data sharing conditions, association judgment criteria, and successful and failed cases of historical interaction data to evaluate logical conflicts, whether the requirements are met, and whether they are consistent with historical experience, and generates a verification result containing verification conclusions, a list of logical problems, and improvement suggestions.

[0081] Step S1377: Based on the results of the logical validity verification operation, perform detailed adjustments to the redesigned collaborative retrieval order, the redefined data sharing conditions, and the redefined association judgment criteria to eliminate logical loopholes and potential conflicts, and obtain the adjusted collaborative retrieval order, the adjusted data sharing conditions, and the adjusted association judgment criteria.

[0082] In this embodiment, the relevant content is adjusted to address the logical problems pointed out by the verification results, such as supplementing the data sharing conditions for the "connecting" state of the interface, breaking the circular dependency of the collaborative retrieval order, etc., and then verifying again until it passes.

[0083] Step S1378: Integrate the adjusted collaborative retrieval order, the adjusted data sharing conditions, and the adjusted association judgment criteria to form the adjusted interactive association logic.

[0084] In this embodiment, the adjusted content is filled into the corresponding fields of the interactive association logic to ensure that it is correctly associated with the source / target dimension identifier and there are no reference errors.

[0085] Step S1379: The adjusted interaction association logic forms an effective linkage with other components in the adapted and adjusted R&D scenario-specific search configuration set, so that the adapted and adjusted R&D scenario-specific search configuration set can respond collaboratively to changes in the adaptation requirements of the R&D scenario.

[0086] In this embodiment, the reference relationship between the adjusted interactive association logic and the updated core search dimensions is checked to ensure that the dimension identifiers of the collaborative search order list exist in the core search dimension set, and that the data types involved in the data sharing conditions are within the coverage of the core search dimensions, thereby achieving effective linkage.

[0087] Step S138: Based on the changes in data transmission requirements in the real-time status characteristics of the R&D scenario, adjust the data transmission format and linkage triggering conditions of the multimodal large model interaction interface in the R&D scenario-specific retrieval configuration set to obtain the adjusted multimodal large model interaction interface.

[0088] In this embodiment, hardware configuration parameters and application interface information based on real-time status characteristics are analyzed to determine changes in data transmission requirements. For example, if the communication module bandwidth decreases, the data volume needs to be reduced; if the data transmission frequency increases, the interface call frequency needs to be adjusted. The data transmission format is adjusted by using compression, efficient encoding, and fragmented transmission. The linkage triggering conditions are adjusted by modifying thresholds, adding types, adjusting frequency limits, etc., and relevant fields of the interaction interface are updated.

[0089] Step S139: Integrate the updated core search dimensions, the adjusted interactive association logic, and the adjusted multimodal large model interaction interface to obtain a set of search configurations adapted to the R&D scenario.

[0090] In this embodiment, the corresponding old part in the search configuration set is replaced, and integrity, consistency and syntax checks are performed. The result is saved as the adapted search configuration set, the version number is updated and the adjustment time and content summary are recorded.

[0091] Step S140: Based on the adapted and adjusted R&D scenario-specific retrieval configuration set, initiate a multi-path collaborative retrieval method to perform targeted data mining operations on the R&D-related big data resource library, and obtain the initial data set corresponding to the targeted data mining operation results.

[0092] In this embodiment, the adapted and adjusted search configuration set specific to the R&D scenario is first loaded into the search management module. The search management module parses the configuration set and extracts the core search dimensions, interaction logic, and multimodal large model interaction interface information. Based on the core search dimensions, multiple independent search paths are created, each corresponding to a core search dimension and responsible for the search task related to that dimension.

[0093] Multi-path collaborative retrieval is implemented through a collaborative scheduler, which coordinates the execution of each retrieval path based on the collaborative retrieval order and data sharing conditions in the interaction logic. Targeted data mining operations are performed in a research and development-related big data resource repository, which includes various data sources such as document databases, code repositories, test data storage, and hardware parameter databases. Each retrieval path connects to the big data resource repository through a corresponding data source interface and performs retrieval operations according to the retrieval execution standards of the core retrieval dimensions, such as sending query requests, filtering data, and extracting relevant data entries.

[0094] During the retrieval process, each retrieval path shares data and collaborates according to the instructions of the coordinator. When a retrieval path meets the data sharing conditions, it sends its retrieval results to other related retrieval paths; after a retrieval path completes its own retrieval task, it triggers the execution of subsequent retrieval paths according to the collaborative retrieval order. The results of the targeted data mining operation include data entries obtained by all retrieval paths and data shared between paths. This data is initially integrated to form an initial dataset.

[0095] Step S141: Analyze the core search dimensions in the adapted and adjusted R&D scenario-specific search configuration set, and determine the target data type and data association attributes corresponding to each core search dimension.

[0096] In this embodiment, the core search dimension fields in the adapted and adjusted R&D scenario-specific search configuration set are parsed. Each core search dimension includes a coverage description field, which details the data types covered by that dimension. Target data types are extracted based on the coverage description. Referring to the metadata classification standards of R&D-related big data resource libraries, data types are divided into major categories such as document types (e.g., requirement documents, design documents, test reports), code types (e.g., source code files, script files, configuration files), data types (e.g., structured data tables, unstructured log data, raw sensor data), and multimedia types (e.g., design drawings, flowcharts, demonstration videos). Each major category is further subdivided into specific data type names.

[0097] Data association attributes are extracted from the association information and interactive association logic of the core search dimensions. These include association fields between data in this dimension and data in other dimensions (such as document ID, project number, hardware model), association types (such as inclusion relationship, reference relationship, causal relationship), and association strength (such as strong association, weak association). For example, the target data types corresponding to the core search dimension "sensor data acquisition and processing" include "sensor raw data file," "filter algorithm source code," and "sensor calibration parameter table." The data association attributes include the association fields "sensor model" and "project ID," and the association types "inclusion relationship" (raw data is included in the test report) and "reference relationship" (source code references the calibration parameter table).

[0098] Step S142: Analyze the interactive association logic in the adapted and adjusted R&D scenario-specific search configuration set to determine the collaborative search order and data interaction method between different core search dimensions.

[0099] In this embodiment, the set of interactive association logic is traversed. For each interactive association logic, the source dimension identifier, target dimension identifier, collaborative retrieval order list, and data sharing condition set are extracted. The collaborative retrieval order list clarifies the execution order of different core retrieval dimensions in the collaborative retrieval process. For example, the list [Dimension A identifier, Dimension B identifier, Dimension C identifier] indicates that the retrieval of Dimension A is executed first, then the retrieval of Dimension B is executed, and finally the retrieval of Dimension C is executed.

[0100] The data interaction method is determined based on a set of data sharing conditions, including the direction of data sharing (from the source dimension to the target dimension or bidirectionally), the format requirements of the shared data (such as JSON, CSV), the sharing channel (such as memory sharing, message queue), and the sharing frequency (such as real-time sharing, scheduled batch sharing). For example, if the data sharing conditions set of a certain interaction logic stipulates that "when new data is retrieved in the source dimension, the data is immediately shared to the target dimension in JSON format via a message queue," then the data interaction method is unidirectional (source to target), JSON format, message queue channel, and real-time sharing.

[0101] Step S143: Based on the target data type and data association attributes corresponding to each core search dimension, generate multiple independent search paths. Each search path contains a set of search statements for a core search dimension.

[0102] In this embodiment, each retrieval path consists of components such as a path identifier, a corresponding core retrieval dimension identifier, a target data source list, a set of retrieval statements, and a result storage buffer. The target data source list is determined based on the target data type, and each target data type corresponds to one or more data sources in the big data resource library (such as a document database or code repository).

[0103] The set of search statements is generated based on the search execution criteria and data association attributes of the core search dimensions. For document-type target data types, the search statements use a structured query language (such as SQL) combined with full-text search syntax, including keywords, data type filtering conditions, and related field matching conditions. For code-type target data types, the search statements use a code-specific query language, including conditions such as function names, variable names, code snippet characteristics, and comment keywords. For data-type target data types, the search statements use a database query language, including conditions such as field value ranges, data timestamp ranges, and related attribute matching. Each search statement corresponds to a specific search request, and the set of search statements is organized in an ordered list, arranged according to execution priority.

[0104] Step S144: Based on the interactive association logic in the adapted and adjusted R&D scenario-specific search configuration set, configure collaborative triggering conditions for each search path, and define the data sharing timing and linkage execution rules between different search paths.

[0105] In this embodiment, the collaborative triggering conditions are extracted from the data sharing conditions and collaborative retrieval order in the interaction association logic. For each retrieval path, corresponding collaborative triggering conditions are configured according to the role (source dimension or target dimension) of its corresponding core retrieval dimension in the interaction association logic.

[0106] As the source dimension retrieval path, the collaborative triggering condition defines when its retrieval results are shared with the retrieval path corresponding to the target dimension. For example, when the number of data entries in the result buffer of retrieval path A reaches a set threshold, data sharing to retrieval path B is triggered; when retrieval path A retrieves a specific type of data (such as error logs), data sharing to retrieval path C is immediately triggered.

[0107] As the target dimension search path, the collaborative triggering condition defines when to receive shared data from the source dimension search path and how to adjust its own search behavior (linked execution rules) based on the shared data. For example, after receiving shared data from search path A, search path B adds keywords from the shared data to its own search statement and performs supplementary search; after receiving shared data, search path C pauses its current search, prioritizes processing related information in the shared data, and then continues the search. The collaborative triggering condition is stored in the search path configuration information in the form of conditional expressions that can be parsed by the rule engine.

[0108] Step S145: Input the retrieval parameters of each retrieval path and the collaborative triggering conditions into the retrieval execution stage to start the parallel retrieval operation of multiple retrieval paths.

[0109] In this embodiment, the retrieval execution phase includes multiple retrieval executors, each responsible for executing the retrieval task for one retrieval path. Retrieval parameters include the target data source list for the retrieval path, the set of retrieval statements, and the address of the result storage buffer. Coordination triggering conditions are loaded into the coordination scheduling module of the retrieval execution phase. The coordination scheduling module is responsible for monitoring the execution status of each retrieval path and coordinating data sharing and coordinated execution when the triggering conditions are met.

[0110] When a parallel retrieval operation is initiated, the retrieval execution phase assigns an independent retrieval executor to each retrieval path. The executor connects to the corresponding target data source based on the retrieval parameters and executes the retrieval requests in the set of retrieval statements sequentially. During execution, the executor stores the retrieved data entries in a result storage buffer and reports the execution status (e.g., executing, completed, error encountered) to the collaborative scheduling module in real time. The collaborative scheduling module determines whether to trigger data sharing or coordinated execution operations based on the execution status of each retrieval path and the collaborative triggering conditions.

[0111] Step S146: During the parallel retrieval process, the retrieval progress and preliminary data fragments of each retrieval path are obtained in real time through the multimodal large model interaction interface in the adapted and adjusted R&D scenario-specific retrieval configuration set.

[0112] In this embodiment, the multimodal large model interaction interface in the adapted and adjusted R&D scenario-specific search configuration set defines the reporting format and frequency of search progress and data fragments. The search executor of each search path generates a search progress report according to the frequency defined by the interface (e.g., once every ten seconds). The report includes information such as path identifier, number of executed search statements, total number of search statements, number of retrieved data entries, and current execution status (e.g., normal, blocked, error).

[0113] Preliminary data fragments refer to a portion of data entries acquired in real time by the retrieval executor during the retrieval process, which are encapsulated according to the data transmission format defined by the interface. The reporting timing of preliminary data fragments can be either timed (e.g., once every thirty seconds) or event-triggered (e.g., immediately upon retrieval of important data entries).

[0114] The retrieval executor sends a retrieval progress report and preliminary data fragments to the multimodal large model via a communication protocol (such as HTTP) specified by the multimodal large model interaction interface. After receiving this information, the multimodal large model uses it to analyze the retrieval process in real time and generate adjustment instructions in subsequent steps (as described in step S130).

[0115] Step S147: Based on the collaborative triggering condition, perform cross-association processing on the preliminary data fragments of different retrieval paths, extract the association information between the data of different paths, and obtain a cross-association data set. The cross-association data set supplements the shortcomings of single-path retrieval.

[0116] In this embodiment, the cross-association processing operation is performed by the association processing unit in the collaborative scheduling module. The association processing unit monitors the preliminary data fragments of each retrieval path in real time, and initiates cross-association processing when the collaborative triggering conditions are met.

[0117] Step S1471: Monitor the retrieval progress of each retrieval path. When any retrieval path meets the initial data fragment generation threshold set in the collaborative triggering conditions, trigger the cross-association processing flow.

[0118] In this embodiment, the initial data fragment generation threshold is extracted from the collaborative triggering conditions, including a data entry quantity threshold (e.g., the result buffer of the retrieval path contains more than a set number of data entries), a data type threshold (e.g., at least one data entry of a specific type is retrieved), and a time threshold (e.g., the retrieval path has been executed for more than a set time). The association processing unit periodically checks the progress status and result buffer of each retrieval path. When it finds that a retrieval path meets any of the threshold conditions, it immediately triggers the cross-association processing flow and records the triggered retrieval path identifier and triggering condition type.

[0119] Step S1472: Obtain the preliminary data fragments generated by the retrieval path that meet the preliminary data fragment generation threshold, and parse the core data identifier and data attribute information in each preliminary data fragment to obtain the first core data identifier and the first data attribute information.

[0120] In this embodiment, the result buffer of the retrieval path that meets the threshold is accessed, and preliminary data fragments are read from it. Each preliminary data fragment is stored in a structured format, including data entry ID, data content, data type, generation time, source data source, core data identifier field, and a list of attribute fields.

[0121] Core data identifiers are key fields used to associate different data fragments, such as project number, document ID, hardware model, and error code. The first core data identifier is extracted by parsing the core data identifier field of a data fragment. Data attribute information includes the data format, size, associated R&D stage, and involved functional modules. The first data attribute information is extracted by parsing the attribute field list of a data fragment. The first core data identifier and the first data attribute information are stored in a temporary association data table as the baseline information for association.

[0122] Step S1473: Obtain the preliminary data fragments currently generated by other parallel retrieval paths, parse the core data identifiers and data attribute information in them, and obtain the second core data identifiers and second data attribute information.

[0123] In this embodiment, other parallel retrieval paths refer to all retrieval paths currently being executed, excluding the retrieval path that triggers the cross-association processing flow. The association processing unit queries the result buffer status of these retrieval paths through the interface of the retrieval execution stage to obtain the generated preliminary data fragments.

[0124] Similar to step S1472, the acquired preliminary data fragments are parsed to extract the second core data identifier (the core data identifier of each data fragment) and the second data attribute information (the data attributes of each data fragment), and stored in a temporary associated data table, which is distinguished from the first core data identifier and the first data attribute information.

[0125] Step S1474: Input the first core data identifier, the first data attribute information, the second core data identifier, and the second data attribute information into the association analysis step to construct a data association index table.

[0126] In this embodiment, the association analysis process includes a core identifier matching module and an attribute information comparison module. The core identifier matching module performs precise matching and fuzzy matching between the first core data identifier and the second core data identifier. Precise matching means that the field values ​​of the core identifiers are exactly the same (e.g., the project numbers are all the same string); fuzzy matching means that the field values ​​of the core identifiers are similar (e.g., the document titles contain the same keyword combination) or there is a hierarchical relationship (e.g., the relationship between the parent document ID and the child document ID).

[0127] The attribute information comparison module compares the first data attribute information with the second data attribute information and calculates the attribute similarity. The attribute similarity is calculated according to the type of the attribute field. For numeric attributes (such as data size), the ratio of the absolute difference to the threshold is calculated. For textual attributes (such as functional module descriptions), semantic similarity is calculated. For enumerated attributes (such as data types), it is calculated whether they belong to the same enumerated value.

[0128] The data association index table records all successfully matched core data identifier pairs and their corresponding attribute similarities. Each record in the table contains information such as the source data fragment ID (from the trigger path or other paths), the target data fragment ID (from other paths or the trigger path), the core identifier matching type (exact or fuzzy), the matched core identifier field, the attribute similarity score, and the association type (based on core identifier or attribute similarity).

[0129] Step S1475: Based on the data association index table, find data entries with consistent core data identifiers or complementary data attribute information in data segments of different retrieval paths to obtain relevant data entries.

[0130] In this embodiment, data entries with consistent core data identifiers are obtained by querying records in the data association index table where the core identifier matching type is exact match. These data entries have the same core identifier field value and belong to the same research and development object or are related to the same event. Data entries with complementary data attribute information are obtained by querying records where the attribute similarity score is higher than a set threshold and the attribute fields have a complementary relationship. Attribute complementarity means that the attribute field of one data entry contains key information missing from another data entry. For example, if data entry A contains the hardware model and error phenomenon, and data entry B contains the same hardware model and error solution, then the two attributes are complementary.

[0131] Related data entries are stored in the form of data entry ID pairs. Each ID pair contains two related data entry IDs from different retrieval paths and is marked with the association criteria (consistent core identifiers or complementary attributes).

[0132] Step S1476: Based on the attribute matching degree between the relevant data items and the association requirements in the R&D scenario, perform association strength analysis on the relevant data items to determine the tightness of the data association between the relevant data items.

[0133] In this embodiment, the attribute matching degree is obtained from the attribute similarity score in the data association index table. The association requirements in the R&D scenario are determined by the interactive association logic in the R&D scenario-specific retrieval configuration set after adaptation and adjustment. Different association types (such as data dependency and functional collaboration) correspond to different association strength requirements.

[0134] Association strength analysis combines attribute matching degree with association requirements to calculate the association strength value. For example, for data dependency type association requirements, attribute matching degree has a higher weight, and the association strength value of related data entries with consistent core identifiers is the attribute similarity score multiplied by a set weight coefficient. For functional collaboration type association requirements, in addition to attribute matching degree, the generation time difference of data entries is also considered; the smaller the time difference, the higher the association strength value. Association strength values ​​are usually divided into multiple levels (such as very strong, strong, medium, and weak), each level corresponding to a numerical range. The degree of closeness between related data entries is determined based on the calculated association strength value.

[0135] Step S1477: Extract the association information between relevant data entries whose data association is close enough to meet the requirements. The association information includes the data source path, association attribute type, and association logical relationship.

[0136] In this embodiment, the tightness of data association is considered to meet the requirements when the association strength level reaches a set threshold (e.g., reaching "strong" or "extremely strong" level). For these related data entries, their association information is extracted. The data source path refers to the retrieval path identifier to which the related data entry belongs; the association attribute type refers to the core data identifier field name or complementary attribute field name that leads to the association; the association logical relationship refers to the relationship type determined according to the association requirements of the R&D scenario, such as data dependency, causal relationship, inclusion relationship, reference relationship, etc. The association information is appended to the record of the related data entry ID pair to form an association information record.

[0137] Step S1478: Bind the extracted association information with the corresponding related data entries to form association data groups. Each association data group contains related data entries from different search paths and their corresponding association information.

[0138] In this embodiment, the binding operation is implemented by creating an associated data group data structure. Each associated data group consists of three parts: a group identifier, a list of data entries, and a list of associated information. The list of data entries contains the complete content of all related data entries (such as data content and data attributes); the list of associated information contains the associated information records corresponding to the data entry pairs. The association between data entries is achieved by referencing the index position of the data entry in the list in the associated information record. For example, the first and second entries in the data entry list are associated through the first record in the associated information list, which contains their source paths, associated attribute types, and logical association relationships.

[0139] Step S1479: Based on the logical relationship in the associated information, obtain the missing associated data from the relevant retrieval path, perform a completeness supplementation operation on each associated data group, and obtain the supplemented associated data group.

[0140] In this embodiment, the logical relationships in the association information indicate the data types that may be missing in the associated data group. For example, if the logical relationship is "data A is a prerequisite for data B", then the associated data group should contain both data A and data B. If the current data entry list only contains data A and data B is missing, then data B needs to be obtained from the retrieval path to which data B belongs.

[0141] The integrity supplementation operation first analyzes the association information of each associated data group to determine the types and quantities of data entries that should be included. Then, it checks whether the current data entry list meets the requirements. If any entries are missing, a supplementary retrieval request is generated and sent to the relevant retrieval path (the retrieval path to which data B belongs). Upon receiving the supplementary request, the relevant retrieval path performs a targeted search based on the core data identifier and attribute conditions in the request, returning the retrieved missing data entries to the association processing unit. The association processing unit adds the missing data entries to the data entry list of the associated data group and updates the association information list, resulting in the supplemented associated data group.

[0142] Step S14710: Summarize all the supplemented related data groups to obtain the cross-related data set.

[0143] In this embodiment, the aggregation operation stores all supplemented related data groups into a unified collection structure, where each element is a related data group object. Simultaneously, an index is created for the cross-related data collection. The index key includes the group identifier, core data identifier, and logical relationship type of the related data group, facilitating subsequent data integration and query operations. The cross-related data collection contains relevant data entries and their associated information from different search paths, supplementing related data that might be missed in a single-path search.

[0144] Step S148: According to the interaction association logic in the adapted and adjusted R&D scenario-specific retrieval configuration set, perform data integration operation on the cross-related data set to form a structured intermediate data set.

[0145] In this embodiment, the data integration operation is performed based on the association judgment criteria and collaborative retrieval rules in the interactive association logic. First, each associated data group in the cross-association data set is parsed to extract data entries and association information. According to the type of association logic relationship, the data entries are organized into a hierarchical structure or a network structure. For example, for associated data groups with an inclusion relationship, the parent data entry is used as the upper-level node and the child data entry as the lower-level node, forming a hierarchical structure; for associated data groups with a many-to-many relationship, the data entries are used as nodes and the association information as edges, forming a network structure.

[0146] Then, the integrated data structure is standardized by unifying data formats (such as date formats and numerical units), adding missing attribute fields (such as uniformly adding a "association strength" attribute), and removing duplicate data entries (based on data entry ID). The standardized structured data is stored as an intermediate dataset, which is stored in the form of a graph database or relational database to support complex relational queries and subsequent processing.

[0147] Step S149: Perform a path redundancy data removal operation on the structured intermediate data set, and use the intermediate data set after removing redundancy as the initial data set obtained by the targeted data mining operation. The initial data set covers all core search dimensions and reflects the collaborative correlation value of data in different dimensions.

[0148] In this embodiment, path redundancy data refers to data entries that appear repeatedly in multiple search paths, or data entries that are not associated with any of the core search dimensions, or data entries whose association strength is lower than a set threshold. The path redundancy data removal operation includes three steps: First, by comparing data entry IDs, the latest version or the version with the richest association information among the duplicate data entries is identified and retained, while the remaining duplicate versions are deleted; second, it is checked whether each data entry is associated with at least one core search dimension (determined by the association dimension identifier in the data attributes), and unrelated data entries are deleted; finally, data entries and their association information with association strength lower than the set threshold are filtered out.

[0149] After the elimination process is completed, the remaining structured data is determined as the initial dataset. A coverage check is performed on the initial dataset to ensure that all core search dimensions have corresponding data entries; a correlation check is also performed to ensure that the collaborative relationships between data entries are complete. The initial dataset is ultimately stored in the form of structured files (such as JSON or XML) or database tables, ready to be input into the multimodal large-scale model for subsequent processing.

[0150] For example, the method may also include: step S1410: during the parallel retrieval process of multiple retrieval paths, the real-time retrieval status of each retrieval path is fed back to the deep-trained multimodal large model through the multimodal large model interaction interface in the adapted and adjusted R&D scenario-specific retrieval configuration set.

[0151] In this embodiment, the retrieval executor of each retrieval path periodically (e.g., every five seconds) collects its own real-time retrieval status, including the currently executed retrieval statement index, the amount of data retrieved, the remaining retrieval task volume, the average retrieval rate, the number of errors, and the reason for the most recent error. The real-time retrieval status is encapsulated according to the status report format defined by the multimodal large model interaction interface and sent to the deeply trained multimodal large model through the interface's communication channel. The multimodal large model receives and stores the real-time retrieval status of all retrieval paths for subsequent retrieval path optimization decisions.

[0152] Step S1411: Analyze the real-time retrieval status of each retrieval path based on the deep-trained multimodal large model. If the data retrieval volume per unit time of the retrieval path is found to be lower than the first preset threshold, or the similarity between the retrieval result and the target data pattern is lower than the second preset threshold, then mark the retrieval path as a retrieval path to be optimized.

[0153] In this embodiment, the multimodal large model internally loads a retrieval performance evaluation model, which receives the real-time retrieval status of each retrieval path as input. First, it calculates the data retrieval volume per unit time, i.e., the amount of retrieved data divided by the retrieval execution time, and compares it with a first preset threshold (the minimum retrieval rate set based on historical retrieval performance data). If it is lower than the threshold, it indicates low retrieval efficiency.

[0154] Simultaneously, the multimodal large model retrieves the most recent search result fragments from the real-time search status and calculates their similarity with the target data pattern (ideal data features extracted from the coverage of the core search dimensions and the search execution criteria). The similarity calculation employs a semantic similarity model and a feature matching algorithm to comprehensively evaluate the degree of matching in terms of data content, data attributes, and relationships. If the similarity is lower than a second preset threshold (set according to the search accuracy requirements of the R&D scenario), it indicates that the search result quality is poor.

[0155] The multimodal large model performs the above two evaluations on each retrieval path. If either evaluation fails, the identifier of the retrieval path is added to the list of retrieval paths to be optimized, and the type of evaluation that failed (efficiency or quality) is recorded.

[0156] Step S1412: For retrieval paths marked as to be optimized because the amount of data retrieved per unit time is lower than the first preset threshold, according to the storage structure of the big data resource library, the number of concurrent retrieval threads of the retrieval path is increased or its data scanning range is expanded to obtain the retrieval path with adjusted parameters.

[0157] In this embodiment, the storage structure information of the big data resource repository is obtained through the resource repository metadata interface, including data sharding method, index distribution, server node load, etc. For the retrieval path to be optimized with low data retrieval volume per unit time, the reasons for its low retrieval efficiency are first analyzed. It may be due to insufficient number of concurrent threads (unable to fully utilize the parallel processing capability of the resource repository) or too narrow data scanning range (missing some data source shards).

[0158] If the repository metadata indicates that the server node connected to the current search path has a low load and the search statement supports parallel execution, then the number of concurrent search threads for that search path is increased, for example, from the original set number to a new number. Each thread is responsible for searching one data shard of the repository. If the repository metadata indicates that the current data scan range does not include all relevant data source shards, then the data scan range is expanded, and the missing shards are added to the target data source list. The configuration information of the search path after parameter adjustment (number of concurrent threads, data source list) is updated, and the search executor continues to execute the search task according to the new configuration.

[0159] Step S1413: For the search paths to be optimized that are marked due to similarity being lower than the second preset threshold, optimize the search keywords of the search path or adjust its search ranking algorithm based on the acquired preliminary data fragments to obtain the search path after direction correction.

[0160] In this embodiment, the acquired preliminary data fragments contain some data entries with high similarity to the target data pattern. By analyzing the characteristics of these highly similar data entries (such as keywords, data structure, and semantic topics), optimized search keywords are extracted. For example, specific terms or phrases that frequently appear in highly similar data entries are added to the search keyword list, while irrelevant keywords contained in low-similarity data entries are removed from the list.

[0161] Adjustments to the search ranking algorithm include modifying algorithm parameters (such as adjusting the weights of different features, increasing semantic similarity weights, and decreasing keyword matching weights) or switching the ranking algorithm type (such as switching from keyword frequency-based ranking to machine learning model-based ranking). The revised search path uses optimized search keywords and the adjusted search ranking algorithm to regenerate the search statement set, replacing the original search statements, and continues to execute the search task.

[0162] Step S1414: Monitor the status changes of the retrieval path after parameter adjustment and the retrieval path after direction correction in real time, record the changes in the data retrieval volume and retrieval result similarity per unit time, and form adjustment effect feedback information.

[0163] In this embodiment, the multimodal large model continuously receives the real-time retrieval status of the retrieval path after parameter adjustment and direction correction through the multimodal large model interaction interface. The focus is on monitoring the changing trends of data retrieval volume per unit time (for the path after parameter adjustment) and retrieval result similarity (for the path after direction correction). Every monitoring cycle (e.g., ten seconds), the current value is recorded and compared with the value before adjustment, calculating the rate of change (e.g., percentage increase in retrieval volume, percentage increase in similarity). The adjustment effect feedback information includes the retrieval path identifier, adjustment type (parameter or direction), indicator values ​​before and after adjustment, rate of change, and current status (still needs optimization or has met the target). This feedback information is sent back to the multimodal large model in real time.

[0164] Step S1415: Based on the feedback information of the adjustment effect, dynamically adjust the number of concurrent threads or the scanning range of the search path after parameter adjustment, and dynamically adjust the search keywords or ranking algorithm weights of the search path after direction correction to obtain the continuously optimized search path.

[0165] In this embodiment, the multimodal large model determines the effectiveness of the adjustment measures based on the feedback information of the adjustment effect. For the retrieval path after parameter adjustment, if the data retrieval volume per unit time is still lower than the first preset threshold and the rate of change is positive (there is an improvement but it does not meet the standard), then the number of concurrent threads will be increased (within the range allowed by the server node load) or the scanning range will be further expanded; if the rate of change is negative (the situation is worse after adjustment), then it will revert to the previous parameter configuration.

[0166] For the retrieval path after direction correction, if the similarity of the retrieval results is still lower than the second preset threshold and the rate of change is positive, then continue to optimize the retrieval keywords (add more keywords extracted from the new high similarity results) or adjust the ranking algorithm weights (further increase the weight of effective features); if the rate of change is negative, then revert to the original retrieval keywords and ranking algorithm. This dynamic adjustment process is repeated until the retrieval path's indicators meet the standards (the number of data retrieved per unit time is higher than the first threshold and the similarity is higher than the second threshold), resulting in a continuously optimized retrieval path.

[0167] Step S1416: During the parallel retrieval process of multiple continuously optimized retrieval paths, retrieval resources are dynamically allocated according to the real-time retrieval needs of each retrieval path, prioritizing the resource supply of retrieval paths corresponding to the core retrieval dimensions in the adapted and adjusted R&D scenario-specific retrieval configuration set.

[0168] In this embodiment, retrieval resources include CPU processing time, memory space, network bandwidth, and the number of data source connections. The real-time retrieval requirements of each continuously optimized retrieval path are measured by indicators such as the complexity of its retrieval statement, the amount of target data, and its priority. The priority values ​​of the core retrieval dimensions in the adapted and adjusted R&D scenario-specific retrieval configuration set determine the resource allocation priority of its corresponding retrieval path.

[0169] Dynamic resource allocation is implemented by a resource scheduler, which monitors the resource usage and real-time retrieval needs of each retrieval path in real time, as well as the system's total resource supply capacity. Based on the priority of core retrieval dimensions, more resource shares are allocated to high-priority retrieval paths. For example, when system CPU resources are scarce, more CPU time slices are allocated to retrieval paths corresponding to core retrieval dimensions; when network bandwidth is insufficient, priority is given to ensuring data source connection bandwidth for core retrieval paths. The resource allocation strategy is implemented through a resource scheduling algorithm that comprehensively considers factors such as priority, real-time needs, and resource utilization to ensure that core retrieval paths can complete their retrieval tasks with priority.

[0170] Step S1417: When a continuously optimized search path completes the preset search target, the search results of the search path are shared in real time with other related continuously optimized search paths.

[0171] In this embodiment, the preset search objectives are set according to the search execution standards of the core search dimensions, including the number of retrieved data entries reaching the expected value, all search statements being executed, and the average similarity of the search results reaching a set threshold. When the continuously optimized search path meets any of the above objective conditions, it is determined that the search task has been completed.

[0172] This retrieval path sends all its search results (stored in a results buffer) to other related, continuously optimized retrieval paths via an internal data bus, according to the data sharing format and scope defined in the interaction and association logic. Related retrieval paths refer to paths that have a collaborative retrieval relationship with this path in the interaction and association logic (such as target dimension paths). The shared data provides other paths with data references (such as core data identifiers that can be associated) and guidance on search direction (such as the data types or attributes that should be prioritized for retrieval).

[0173] Step S1418: Based on the shared search results, other related continuously optimized search paths adjust their own search strategies to avoid duplicate search behavior, improve overall search efficiency, and obtain the search paths after strategy adjustment.

[0174] In this embodiment, after receiving the shared search results, the relevant search path first parses the results to extract core data identifiers and data attribute information. It then compares this information with the data entries it has already retrieved to identify the data range covered by the shared results, thereby avoiding duplicate searches of the same data range (canceling the corresponding search statements).

[0175] Simultaneously, based on the data characteristics in the shared results (such as frequently occurring keywords and new data types), the system adjusts its search strategy. This includes updating the keyword list (adding new keywords from the shared results), adjusting the priority of search statements (prioritizing search statements for data types highly relevant to the shared results), and optimizing filtering conditions (narrowing the search scope based on the attribute range of the shared results). Through these adjustments, relevant search paths can more accurately locate the required data, reduce invalid searches, and improve overall search efficiency, thus becoming the search paths after strategy adjustments.

[0176] Step S1419: After all the search operations of the search paths after the strategy adjustment are completed, summarize the search results of all the search paths after the strategy adjustment and the collaborative data between the paths to form an initial data set.

[0177] In this embodiment, the completion of the retrieval task for all strategy-adjusted retrieval paths is indicated by their satisfaction of the preset retrieval objectives. The retrieval management module waits for all paths to complete before initiating a summary operation. The summary includes retrieval result data entries for each path, shared collaborative data between paths (such as cross-referenced data sets), and retrieval process logs (such as optimization records and shared records). The summary operation integrates the above content into a unified data structure, removes duplicate data entries, retains all collaborative association information, and forms the final initial data set. This initial data set is stored in a temporary database, ready to be input into a multimodal large model for scenario-based filtering and optimization.

[0178] Step S150: Input the initial data set into the multimodal large model, and perform scenario-based filtering and optimization processing on the initial data set through the multimodal large model to generate the final retrieval result set.

[0179] In this embodiment, the initial dataset is encapsulated according to the input format defined by the multimodal large model interaction interface and sent to the deeply trained multimodal large model through the interface channel. The multimodal large model internally loads a scenario-based processing module, which includes a data filtering submodule and a data optimization submodule. The data filtering submodule scores and filters data entries in the initial dataset based on the real-time status characteristics of the R&D scenario and the priority of the core retrieval dimensions. The data optimization submodule performs correlation integration, hierarchical division, and structured organization on the filtered data, ultimately generating a final retrieval result set that meets the needs of the R&D scenario.

[0180] Step S151: Input the initial data set into the deep-trained multimodal large model through the multimodal large model interaction interface in the adapted and adjusted R&D scenario-specific retrieval configuration set, and start the scenario-based screening process of the multimodal large model.

[0181] In this embodiment, the multimodal large model interaction interface in the adapted and adjusted R&D scenario-specific retrieval configuration set defines the input protocol and format requirements for the initial data set. The initial data set is converted into a binary stream or structured text format required by the interface, and data header information (such as data size, number of data entries, and timestamp) is added. It is then sent to the input layer of the multimodal large model through an encrypted transmission channel.

[0182] After receiving the initial dataset, the multimodal large model depackages and validates the data to ensure its integrity. Once validation is successful, it activates the scenario-based filtering process, loading the filtering model parameters and rule base corresponding to the current R&D scenario, and prepares to process the initial dataset.

[0183] Step S152: Extract the development process node association information corresponding to each data entry in the initial dataset based on the deep-trained multimodal large model, and analyze the correlation between each data entry and the current R&D stage.

[0184] In this embodiment, the development process association analysis layer of the multimodal large model is responsible for handling this task. Each data entry is input into this layer, where a text encoder encodes the content of the data entry to obtain a content feature vector. Simultaneously, the development process node identifier (such as the R&D stage corresponding to the data generation time) is extracted from the attribute fields of the data entry. The current R&D stage is obtained from the real-time status features of the R&D scenario, and its corresponding stage feature vector (pre-stored in the model) is used to calculate similarity with the content feature vector of the data entry.

[0185] The correlation analysis combines the matching degree of development process node identifiers with the similarity of content feature vectors. If the development process node identifier of a data entry is consistent with the current R&D stage, the basic correlation is high; if they are inconsistent, but the content feature vector is highly similar to the stage feature vector (the data content is relevant to the current stage), the correlation is increased accordingly. The final correlation is represented by a value between 0 and 1, with a higher value indicating a higher correlation.

[0186] Step S153: Extract the hardware parameter association information corresponding to each data entry in the initial data set based on the deep-trained multimodal large model, and analyze the matching degree between each data entry and the current hardware state of the data acquisition terminal.

[0187] In this embodiment, the hardware parameter matching layer of the multimodal large model handles this task. Hardware parameter association information in the data entries is extracted using regular expression matching and named entity recognition techniques, including hardware model, core parameter values ​​(such as processor frequency, memory capacity, and sensor accuracy), and interface type. The current hardware status of the data acquisition terminal is obtained from the hardware configuration parameters in the real-time status characteristics of the R&D scenario, including the currently used hardware model and parameter values.

[0188] The matching degree analysis compares the hardware model of the data entry with the current hardware model. If they match, it further compares the compatibility range of core parameter values ​​(e.g., whether the sensor accuracy in the data entry is within the accuracy range of the current hardware sensor). If the hardware models do not match, it analyzes whether the hardware parameters in the data entry are suitable for the current hardware upgrade or replacement scenario (e.g., whether they are parameters of a compatible upgrade model of the current hardware). The matching degree is also represented by a value between 0 and 1, combining the model matching and parameter compatibility.

[0189] Step S154: Extract the application docking association information corresponding to each data entry in the initial dataset based on the deep-trained multimodal large model, and analyze the matching degree between each data entry and the target application adaptation requirements.

[0190] In this embodiment, the application adaptation analysis layer of the multimodal large model handles this task. Application integration information for data entries is obtained through text classification and keyword extraction techniques, including the adapted application name, application version, interface type, data format requirements, and compatibility identifiers. Target application adaptation requirements are extracted from application adaptation request information in the real-time status characteristics of the development scenario, including the target application name, target version, required interface list, and data format standards.

[0191] Matching analysis first checks whether the application name and version of the data entries match the target application. Then, it verifies whether the interface type and data format meet the target requirements. Finally, it assesses the impact of compatibility indicators (such as "compatible," "incompatible," or "requires modification") on adaptation. Matching score is represented by a value between 0 and 1, combining the results of name and version matching, interface format conformity, and compatibility evaluation.

[0192] Step S155: Based on the relevance of each data entry to the current R&D stage, the matching degree with the current hardware status of the data acquisition terminal, and the matching degree with the adaptation requirements of the target application, perform a comprehensive scenario adaptation evaluation operation on each data entry to obtain a comprehensive score result for each data entry.

[0193] In this embodiment, the comprehensive scene adaptation evaluation operation is performed in the comprehensive evaluation layer of the multimodal large model. This comprehensive evaluation layer integrates the three indicators of relevance, matching degree and matching degree into a comprehensive score.

[0194] Step S1551: Set evaluation dimensions for the relevance of data entries to the current R&D stage. Each evaluation dimension corresponds to the specific data requirements of a key link in the R&D stage, thus obtaining the first set of evaluation dimensions.

[0195] In this embodiment, the key steps in the current R&D phase are defined through the R&D process specification document. For example, key steps in the requirements analysis phase include user requirements research, requirements specification definition, and requirements review; key steps in the development phase include architecture design, module development, and unit testing. The specific data requirements for each key step constitute an evaluation dimension. For example, the data evaluation dimensions for the requirements research phase include "data includes user interview records," "data covers all target user groups," and "data records user pain point descriptions." The first set of evaluation dimensions is a list of these evaluation dimensions, each with a unique identifier and descriptive name.

[0196] Step S1552: Set the evaluation dimensions for the matching degree between data entries and the current hardware status of the data acquisition terminal. Each evaluation dimension corresponds to the technical parameter requirements of a component in the hardware structure for the data, thus obtaining the second set of evaluation dimensions.

[0197] In this embodiment, the hardware structure of the data acquisition terminal is analyzed from the hardware configuration parameters, including components such as a core processor, memory, storage, sensors, communication modules, and power modules. The technical parameters of each component constitute an evaluation dimension. For example, the evaluation dimensions for the core processor include "the processor model in the data is consistent with the current model," "the clock frequency parameters in the data are within the current hardware support range," and "the instruction set in the data is compatible with the current processor." The evaluation dimensions for sensors include "the sensor type in the data is consistent with the current configuration" and "the sampling frequency in the data does not exceed the current sensor's maximum frequency," etc. The second set of evaluation dimensions is a list of these evaluation dimensions.

[0198] Step S1553: Set the evaluation dimensions for the matching degree between data entries and the adaptation requirements of the target application. Each evaluation dimension corresponds to the data compatibility requirements of an adaptation link in the application docking process, thus obtaining the third set of evaluation dimensions.

[0199] In this embodiment, the adaptation process during application integration is determined based on application adaptation requirements, including interface protocol adaptation, data format adaptation, functional module adaptation, performance indicator adaptation, and security policy adaptation. The compatibility requirements for each adaptation step constitute an evaluation dimension. For example, the evaluation dimensions for interface protocol adaptation include "the interface type in the data matches the target interface," "the communication protocol version in the data meets the target requirements," and "the data contains example code for interface calls." The evaluation dimensions for data format adaptation include "the data structure in the data matches the target format definition," "the field types in the data meet the target requirements," and "the data contains format conversion examples," etc. The third set of evaluation dimensions is a list of these evaluation dimensions.

[0200] Step S1554: Based on the overall requirements of the R&D scenario and historical evaluation data, set specific evaluation execution standards for each evaluation dimension in the first evaluation dimension set, the second evaluation dimension set, and the third evaluation dimension set, to obtain an evaluation execution standard set.

[0201] In this embodiment, the overall requirements of the R&D scenario include functional completeness, performance reliability, development efficiency, and cost control. Historical evaluation data includes successful and unsuccessful cases of data evaluation in similar past R&D scenarios. Combining the overall requirements and historical data, evaluation execution standards are set for each evaluation dimension. The standards include judgment criteria (e.g., "there are clear user interview transcripts in the data"), compliance levels (e.g., "fully compliant", "partially compliant", "not compliant"), and level scoring rules (e.g., a set score for fully compliant, a set score for partially compliant, and a set score for not compliant). The set of evaluation execution standards is a summary of the execution standards for all evaluation dimensions, indexed by the dimension identifier.

[0202] Step S1555: Compare the relevant information corresponding to each data entry with each evaluation execution standard in the set of evaluation execution standards one by one, record the compliance status of each data entry in each evaluation dimension, and obtain the dimension compliance record of each data entry.

[0203] In this embodiment, the relevant information for each data entry includes data content, data attributes, associated R&D stages, and involved hardware modules. The comparison operation is performed one by one according to the order of the evaluation dimension set. For each evaluation dimension, the relevant information of the data entry is checked to see if it conforms to the evaluation execution standard. For example, for the dimension "data contains user interview records" in the first evaluation dimension set, the data content is checked to see if there are text paragraphs marked as user interviews; for the dimension "sensor sampling frequency does not exceed the maximum frequency" in the second evaluation dimension set, the sampling frequency value in the data is checked to see if it is less than or equal to the current sensor's maximum frequency parameter. The comparison results are recorded as a compliance level and corresponding score. The comparison results of all dimensions constitute the dimensional compliance record for this data entry, and the record format is a key-value pair list of dimension identifier-compliance level-score.

[0204] Step S1556: Based on the importance of each evaluation dimension to the current R&D scenario, assign corresponding weight coefficients to each evaluation dimension in the first set of evaluation dimensions, the second set of evaluation dimensions, and the third set of evaluation dimensions to obtain a set of weight coefficients.

[0205] In this embodiment, the importance of the evaluation dimensions is determined using an expert scoring method. Experts in the R&D field are invited to rate the importance of each evaluation dimension (e.g., 1-5 points), based on the dimension's impact on the current R&D scenario objectives. The expert scores are averaged to obtain the importance score for each evaluation dimension. The weight coefficient is obtained by normalizing the importance scores; that is, the weight coefficient for each dimension equals its importance score divided by the sum of the importance scores of all dimensions. The set of weight coefficients is stored in dictionary form, with the key being the evaluation dimension identifier and the value being the corresponding weight coefficient.

[0206] Step S1557: Based on the dimensional compliance record of each data entry and the corresponding weight coefficient in the set of weight coefficients, perform a comprehensive calculation operation on the multi-dimensional evaluation results to obtain the comprehensive evaluation score of each data entry.

[0207] In this embodiment, the comprehensive calculation operation performs a weighted summation of the scores for all evaluation dimensions of each data entry. For each dimension in the dimension compliance record, its score is multiplied by the corresponding weight coefficient in the weight coefficient set to obtain a weighted score. The weighted scores of all dimensions are summed to obtain the comprehensive evaluation score of the data entry. For example, if a data entry has a score of 80 in dimension A (weight 0.3), a score of 90 in dimension B (weight 0.2), and a score of 70 in dimension C (weight 0.5), then the comprehensive evaluation score is 77.

[0208] Step S1558: Based on the distribution range of the comprehensive evaluation scores of all data items, divide all data items into different adaptation evaluation levels, with each adaptation evaluation level corresponding to a type of scenario adaptation degree.

[0209] In this embodiment, the distribution range of the comprehensive evaluation score is determined by statistically analyzing the scores of all data items, including the minimum, maximum, average, and standard deviation. The score range is divided into multiple intervals using equidistant grouping or cluster analysis, with each interval corresponding to a suitability evaluation level. For example, a score between 90 and 100 is considered "Excellent Suitability," 75-89 is "Good Suitability," 60-74 is "Basic Suitability," and below 60 is "Inadequate Suitability." Each level has a name, a score range, and a description of the degree of suitability (e.g., "Excellent Suitability" indicates that the data fully meets the requirements of the current scenario and can be used directly).

[0210] Step S1559: Perform feature analysis on the data entries in each adaptation evaluation level, extract the common features of the data entries in the adaptation evaluation level in terms of relevance, matching degree and matching degree, and obtain the common features of the level.

[0211] In this embodiment, the feature analysis operation summarizes and analyzes all data entries in each adaptation evaluation level. For relevance (with the R&D stage), the conformity rate of data entries in this level across each dimension of the first evaluation dimension set is calculated (number of entries conforming to that dimension / total number of entries), and the top few dimensions with the highest conformity rates are identified as common features of relevance. For matching (with hardware status), the conformity rate of each dimension of the second evaluation dimension set is calculated, and common features are identified. For matching (with application requirements), the conformity rate of each dimension of the third evaluation dimension set is calculated, and common features are identified. The common features of the level are the set of these high-conformity dimensions, describing the typical characteristics of data entries in this level.

[0212] Step S15510: Based on the common characteristics of the levels, perform a secondary calibration operation on the comprehensive evaluation score of each data item to obtain the comprehensive score result of each data item.

[0213] In this embodiment, the secondary calibration operation fine-tunes the comprehensive evaluation score based on the data item's suitability assessment level and its common characteristics. If a data item belongs to a certain level but lacks a key common characteristic of that level (e.g., the common characteristic of the "Excellent Suitability" level includes "data contains requirement review records," while an item scores high in other dimensions but lacks this characteristic), its comprehensive evaluation score is reduced. If a data item belongs to a lower level but possesses common characteristics of a higher level (e.g., the "Basic Suitability" item possesses some common characteristics of "Excellent Suitability"), its comprehensive evaluation score is increased. The calibration range is determined based on the importance weight of the common characteristics, and the final comprehensive score result serves as the final quantitative indicator of the data item's scenario suitability.

[0214] Step S156: Based on the comprehensive score of each data item, select data items with scores higher than the preset threshold to form the first stage of the selected data set.

[0215] In this embodiment, the preset threshold is set according to the accuracy requirements of the search results in the R&D scenario, referencing the lowest comprehensive score of data entries marked as "useful" by R&D personnel in historical evaluation data. All data entries in the initial dataset are traversed, and their comprehensive scores are compared with the preset threshold. Data entries with scores higher than the threshold are copied to a new dataset, forming the first-stage filtered dataset. Simultaneously, the scores and reasons for the filtered data entries (e.g., "score below the threshold") are recorded for subsequent analysis of the effectiveness of the search configuration.

[0216] Step S157: Feed back each data entry in the first stage of the filtered data set to the corresponding retrieval path, triggering a secondary supplementary retrieval operation of the retrieval path to obtain extended data associated with the core data entry.

[0217] In this embodiment, each data entry in the first-stage filtered dataset includes its source retrieval path identifier. This identifier is used to feed back the data entry information (core data identifier, data attributes, and comprehensive score) to the corresponding retrieval path. Upon receiving the feedback, the retrieval path initiates a secondary supplementary retrieval operation. Based on the core data identifier and attribute information of the data entry, it retrieves extended data associated with that core data entry from the big data resource library. Extended data includes other data referencing the core data, source data referenced by the core data, related data belonging to the same project or module, and detailed data supplementing the core data. The secondary supplementary retrieval uses more precise search criteria, limiting it to the scope related to the core data identifier, to obtain targeted extended data.

[0218] Step S158: Merge the extended data with the first stage filtered data set to generate a merged data set. Calculate and label the association weights between different data items in the merged data set using the multimodal large model to form the second stage optimized data set.

[0219] In this embodiment, the merging operation adds extended data to the first-stage filtered data set, removes duplicate data entries (determined by data entry ID), and retains all unique data entries to generate a merged data set. The multimodal large model performs association analysis on all data entries in the merged data set, calculating the association weight between any two data entries. The association weight is calculated based on factors such as the matching degree of core data identifiers, attribute similarity, content semantic similarity, and the strength of the logical relationship between the data entries, and is represented by a value between 0 and 1, with larger values ​​indicating a stronger association.

[0220] After the association weights are calculated, an association list is added for each data entry. Each element in the list contains the associated data entry ID and the corresponding association weight. Simultaneously, entries with association weights higher than a set threshold are marked as "strongly associated," and the rest are marked as "weakly associated." The set containing the data entries, the association list, and the association strength labels constitutes the second-stage optimization dataset.

[0221] Step S159: Based on the supporting role of data in R&D needs, perform data hierarchy division on the second-stage optimized data set to divide it into core supporting data hierarchy, auxiliary supporting data hierarchy, and reference data hierarchy.

[0222] In this embodiment, the supporting role of data in R&D needs is determined by analyzing the comprehensive score results and correlation weights of data items. The core supporting data level includes data items with the highest comprehensive score (e.g., top 20%) and the highest total correlation weight with other data items. These data directly address key issues in the current R&D stage and serve as the primary basis for R&D decisions. The auxiliary supporting data level includes data items with moderate comprehensive score results (e.g., 20%-60%) and strong correlations with the core supporting data. These data supplement the core data, providing more details or supporting information. The reference data level includes data items with lower comprehensive score results (e.g., below 60%) but potentially useful for future R&D stages, or data items with weak correlations with the core / auxiliary data. These data serve as potential reference resources. The data hierarchy partitioning operation allocates data items in the second-stage optimized data set to the three levels, each with its own independent storage area and index.

[0223] Step S1510: Extract all data entries of the core support data level and the auxiliary support data level, and perform final structured organization according to the interaction association logic in the adapted and adjusted R&D scenario-specific retrieval configuration set to generate a final retrieval result set that meets the customized development requirements of the data acquisition terminal fusion application.

[0224] In this embodiment, the final structured processing operation first merges all data entries from the core support data level and the auxiliary support data level. Then, based on the interaction association logic in the adapted and adjusted R&D scenario-specific retrieval configuration set, the merged entries are organized. Data entries are grouped into different dimension categories according to the core retrieval dimensions and their priority defined in the interaction association logic; a hierarchical or network structure is constructed between data entries (e.g., core data entries as root nodes, auxiliary data entries as child nodes) based on the association type and strength in the interaction association logic. Structured processing also includes unifying data formats (e.g., converting all document-type data to PDF format, adding syntax highlighting to code-type data), adding standardized metadata (e.g., data source, creation time, association dimensions, and description of supporting functions), and generating a data catalog (a list of entries organized by dimension category and hierarchical structure). The final generated retrieval result set is provided in the form of a compressed package or database export file, containing structured data files, a data catalog, and association relationship documentation, which can be directly used for customized development of data acquisition terminal fusion applications.

[0225] Furthermore, Figure 2 A schematic diagram of the hardware structure of an intelligent big data retrieval system 100 that integrates multimodal large models for implementing the methods provided in the embodiments of this application is shown. Figure 2As shown, the intelligent big data retrieval system 100 integrating multimodal big models may include at least one processor 102 (processor 102 may be, but is not limited to, a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, a transmission device 106 for communication functions, and a controller 108. Those skilled in the art will understand that... Figure 2 The structure shown is for illustrative purposes only and does not limit the structure of the intelligent big data retrieval system 100 that integrates multimodal big models. For example, the intelligent big data retrieval system 100 that integrates multimodal big models may also include more than Figure 2 The more or fewer components shown, or having the same Figure 2 The different configurations shown.

[0226] The memory 104 can be used to store software programs and modules of application software, such as the program instructions corresponding to the method embodiments described above in this application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby realizing the above-described intelligent big data retrieval method that integrates multimodal big models. The transmission device 106 is used to acquire or send data via a network.

[0227] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

Claims

1. An intelligent big data retrieval method integrating multimodal large models, characterized in that, The method includes: The dynamic scene elements in the customized development and research scenario of the integrated application of data acquisition terminal are obtained. The dynamic scene elements include development stage characteristics, terminal hardware characteristics, application adaptation requirements and data association requirements. The dynamic scene elements are input into a deep-trained multimodal large model to generate a search configuration set specifically for the R&D scene. The search configuration set specifically for the R&D scene includes the definition of search dimensions, the method of element association, and the interaction interface of the multimodal large model. Through the multimodal large model interaction interface, the real-time state features of the R&D scenario are input into the multimodal large model. Based on the adjustment instructions output by the multimodal large model, the search configuration set dedicated to the R&D scenario is updated to obtain the adapted and adjusted search configuration set dedicated to the R&D scenario. Based on the adapted and adjusted R&D scenario-specific retrieval configuration set, a multi-path collaborative retrieval method is initiated to perform targeted data mining operations on the R&D-related big data resource library, thereby obtaining the initial data set corresponding to the targeted data mining operation results. The initial dataset is input into the multimodal large model, and the multimodal large model performs scenario-based filtering and optimization on the initial dataset to generate the final retrieval result set.

2. The intelligent big data retrieval method based on the fusion of multimodal large models according to claim 1, characterized in that, The step of inputting the dynamic scene elements into a deep-trained multimodal large model to generate a search configuration set specific to the R&D scene includes: Extract the development stage features from the dynamic scene elements, break them down to obtain development process node information, and each development process node information corresponds to a specific implementation link in the customized development of the data acquisition terminal fusion application. Extract the terminal hardware characteristic information from the dynamic scene elements, and break it down to obtain hardware structure association information. Each piece of hardware structure association information corresponds to a hardware component and functional implementation path of the data acquisition terminal. The application adaptation requirement information is extracted from the dynamic scene elements and broken down to obtain application compatibility association information. Each application compatibility association information corresponds to the adaptation direction and association logic between the data acquisition terminal and the target application. Extract data association requirement information from the dynamic scene elements, and break it down to obtain data interaction association information. Each data interaction association information corresponds to the flow and collaboration logic of different types of data in the R&D process. The development process node information, the hardware structure association information, the application compatibility association information, and the data interaction association information are input into a deep-trained multimodal large model, and a multi-dimensional association modeling operation of the R&D scenario is performed to obtain the result of the multi-dimensional association modeling operation of the R&D scenario. Based on the results of the multi-dimensional correlation modeling operation of the R&D scenario, the core search dimensions exclusive to the R&D scenario are mapped, and each core search dimension corresponds to a key R&D requirement direction. Based on the results of the multi-dimensional association modeling operation of the R&D scenario, the interactive association logic between the core search dimensions is generated, and the collaborative search rules for data of different dimensions are defined. Based on the development interface standard of the data acquisition terminal, a multimodal large model interaction interface is generated between the multimodal large model and the retrieval execution link, defining the data transmission format and linkage triggering conditions; By integrating the core search dimensions, the interaction logic between the core search dimensions, and the multimodal large model interaction interface, a basic structure for a search configuration set specific to the R&D scenario is formed. Based on historical search case data of R&D scenarios, scenario-adaptability enhancement operations are performed on the basic structure of the R&D scenario-specific search configuration set to obtain the R&D scenario-specific search configuration set. Each component of the R&D scenario-specific search configuration set corresponds to the actual needs of the R&D scenario.

3. The intelligent big data retrieval method based on the fusion of multimodal large models according to claim 1, characterized in that, The process involves inputting the real-time state features of the R&D scenario into the multimodal large model via the multimodal large model interaction interface. Based on the adjustment instructions output by the multimodal large model, the R&D scenario-specific search configuration set is updated to obtain an adapted and adjusted R&D scenario-specific search configuration set, including: Through the R&D status monitoring stage of the data acquisition terminal, real-time development progress information of the R&D scenario is collected. The real-time development progress information includes the currently executed development process node and the completion status of the development process node. Through the R&D status monitoring stage of the data acquisition terminal, real-time hardware adaptation information of the R&D scenario is collected. The real-time hardware adaptation information includes the current hardware debugging results and hardware parameter adjustment status of the data acquisition terminal. Through the R&D status monitoring stage of the data acquisition terminal, real-time application docking information of the R&D scenario is collected. The real-time application docking information includes the temporary docking effect between the data acquisition terminal and the target application and feedback on adaptation issues. The real-time development progress information, the real-time hardware adaptation information, and the real-time application docking information are input into the deep-trained multimodal large model through the multimodal large model interaction interface. Based on the deep-trained multimodal large model, feature extraction operations are performed on the real-time development progress information, the real-time hardware adaptation information, and the real-time application docking information to obtain the real-time status features of the R&D scenario. The real-time status features of the R&D scenario include the current R&D stage identifier, hardware configuration parameters, and application interface status. The real-time status features of the R&D scenario are correlated with the existing components of the R&D scenario-specific retrieval configuration set to identify the core retrieval dimensions and interactive correlation logic in the R&D scenario-specific retrieval configuration set that do not match the real-time status features of the R&D scenario. Based on the current R&D stage identifier and hardware configuration parameters in the real-time status characteristics of the R&D scenario, update the coverage parameters and priority values ​​of the core search dimensions that differ in the R&D scenario-specific search configuration set to obtain the updated core search dimensions. Based on the application interface status in the real-time status characteristics of the R&D scenario, the collaborative retrieval rules of the interaction association logic that differ in the R&D scenario-specific retrieval configuration set are adjusted to obtain the adjusted interaction association logic. Based on the changes in data transmission requirements in the real-time status characteristics of the R&D scenario, the data transmission format and linkage triggering conditions of the multimodal large model interaction interface in the R&D scenario-specific retrieval configuration set are adjusted to obtain the adjusted multimodal large model interaction interface. By integrating the updated core search dimensions, the adjusted interactive association logic, and the adjusted multimodal large model interaction interface, a set of search configurations adapted to the R&D scenario is obtained.

4. The intelligent big data retrieval method based on the fusion of multimodal large models according to claim 1, characterized in that, Based on the adapted and adjusted R&D scenario-specific search configuration set, a multi-path collaborative search method is initiated to perform targeted data mining operations on the R&D-related big data resource library, obtaining an initial data set corresponding to the targeted data mining operation results, including: Analyze the core search dimensions in the adapted and adjusted R&D scenario-specific search configuration set, and determine the target data type and data association attributes corresponding to each core search dimension; Analyze the interactive logic in the adapted and adjusted R&D scenario-specific search configuration set to determine the collaborative search order and data interaction method between different core search dimensions; Based on the target data type and data association attributes corresponding to each core search dimension, multiple independent search paths are generated, and each search path contains a set of search statements for a core search dimension. Based on the interactive association logic in the adapted and adjusted R&D scenario-specific search configuration set, collaborative triggering conditions are configured for each search path, and the data sharing timing and linkage execution rules between different search paths are defined. Input the retrieval parameters of each retrieval path and the aforementioned collaborative triggering conditions into the retrieval execution stage to initiate parallel retrieval operations for multiple retrieval paths; During the parallel retrieval process, the retrieval progress and preliminary data fragments of each retrieval path are obtained in real time through the multimodal large model interaction interface in the adapted and adjusted R&D scenario-specific retrieval configuration set. Based on the aforementioned collaborative triggering conditions, the preliminary data fragments from different retrieval paths are cross-associated to extract the association information between the data from different paths, resulting in a cross-association data set. This cross-association data set supplements the shortcomings of single-path retrieval. According to the interaction association logic in the adapted and adjusted R&D scenario-specific retrieval configuration set, the cross-association data set is integrated to form a structured intermediate data set; The structured intermediate data set is subjected to path redundancy data removal operation, and the intermediate data set after redundancy removal is used as the initial data set obtained by targeted data mining operation. The initial data set not only covers all core search dimensions, but also reflects the collaborative correlation value of data in different dimensions.

5. The intelligent big data retrieval method based on the fusion of multimodal large models according to claim 1, characterized in that, The process of inputting the initial data set into the multimodal large model, and then using the multimodal large model to perform scenario-based filtering and optimization on the initial data set to generate a final retrieval result set includes: The initial dataset is input into the deep-trained multimodal large model through the multimodal large model interaction interface in the adapted and adjusted R&D scenario-specific retrieval configuration set, thereby initiating the scenario-based screening process for the multimodal large model. Based on a deep-trained multimodal large model, the development process node association information corresponding to each data entry in the initial dataset is extracted, and the correlation between each data entry and the current R&D stage is analyzed. Based on a deep-trained multimodal large model, hardware parameter association information corresponding to each data entry in the initial dataset is extracted, and the matching degree between each data entry and the current hardware state of the data acquisition terminal is analyzed. Based on a deep-trained multimodal large model, the application docking association information corresponding to each data entry in the initial dataset is extracted, and the matching degree between each data entry and the target application adaptation requirements is analyzed. Based on the relevance of each data entry to the current R&D stage, the matching degree with the current hardware status of the data acquisition terminal, and the matching degree with the adaptation requirements of the target application, a comprehensive scenario adaptation evaluation operation is performed on each data entry to obtain a comprehensive score result for each data entry. Based on the comprehensive score of each data item, data items with scores higher than a preset threshold are selected to form the first-stage data set. Each data entry in the first stage of the filtered data set is fed back to the corresponding retrieval path, triggering a secondary supplementary retrieval operation in that retrieval path to obtain extended data associated with the core data entry; The extended data is merged with the first-stage filtered data set to generate a merged data set. The association weights between different data items in the merged data set are calculated and labeled using the multimodal large model to form the second-stage optimized data set. Based on the supporting role of data in R&D needs, the second-stage optimized data set is divided into a data hierarchy, namely, a core supporting data hierarchy, an auxiliary supporting data hierarchy, and a reference data hierarchy. Extract all data entries from the core support data layer and the auxiliary support data layer, and perform final structured organization according to the interaction association logic in the adapted and adjusted R&D scenario-specific retrieval configuration set to generate a final retrieval result set that meets the customized development requirements of the data acquisition terminal integrated application.

6. The intelligent big data retrieval method based on the fusion of multimodal large models according to claim 2, characterized in that, The results of the multi-dimensional correlation modeling operation based on the R&D scenario are mapped to the core retrieval dimensions specific to the R&D scenario, including: Extract the development process node association information from the multi-dimensional association modeling operation results of the R&D scenario, identify the key data requirement direction corresponding to each development process node, and each key data requirement direction corresponds to the data type that the development process node must obtain during the R&D process; Extract the hardware structure association information from the multi-dimensional association modeling operation results of the R&D scenario, identify the technical parameter data requirements corresponding to each hardware component, and identify the data content required for hardware debugging and optimization for each technical parameter data requirement. Extract application compatibility association information from the multi-dimensional association modeling operation results of the research and development scenario, identify the adaptation data requirements in the process of data acquisition terminal and target application docking, and each adaptation data requirement corresponds to the data support content to ensure the normal operation of the application; Extract data interaction and association information from the multi-dimensional association modeling operation results of the R&D scenario, identify collaborative data requirements in different types of data flow processes, and ensure data association content for each collaborative data requirement to ensure effective data interaction. The identified key data demand directions, technical parameter data demands, adaptation data demands, and collaborative data demands are input into a deep-trained multimodal large model, and a similar merging operation is performed on the demand dimensions to obtain the merged demand dimensions. Based on the overall development goals of the R&D scenario, the merged requirement dimensions are sorted by importance to obtain the sorted requirement dimensions, and the priority of each requirement dimension in the retrieval process is determined. Define a definite coverage area for each demand dimension in the sorted demand dimensions, and define the data types and data characteristics contained in each demand dimension to obtain demand dimensions with a definite coverage area. Mark the potential correlation points between each demand dimension and other demand dimensions in the demand dimensions with a defined coverage area to obtain the demand dimensions with marked potential correlation points; Based on the coverage, priority, and potential related points of the demand dimensions marked with potential related points, a retrieval execution standard is generated for each demand dimension, defining the retrieval method and judgment criteria for the data in that dimension, thus obtaining a demand dimension with a retrieval execution standard. The requirement dimension with retrieval execution standard is defined as the core retrieval dimension exclusive to the R&D scenario. Each core retrieval dimension exclusive to the R&D scenario can accurately correspond to a core data requirement of the R&D scenario.

7. The intelligent big data retrieval method based on the fusion of multimodal large models according to claim 3, characterized in that, The collaborative retrieval rules, which adjust the differing interactive association logics in the R&D scenario-specific retrieval configuration set based on changes in adaptation requirements in the real-time state characteristics of the R&D scenario, are used to obtain the adjusted interactive association logic, including: Extract adaptation requirement change information from the real-time status features of the R&D scenario, and identify the core retrieval dimensions and related logic involved in the adaptation requirement change information; The interaction logic that differs in the R&D scenario-specific search configuration set is analyzed, the specific content in the original collaborative search rules that conflicts with the information on the change in adaptation requirements is identified, and the causes of the conflict are analyzed and operated to distinguish the problems in the search order setting, the definition of data sharing conditions, or the setting of association judgment criteria. Based on the specific reasons for the conflict and the specific requirements in the adaptation needs change information, the collaborative retrieval order between the relevant core retrieval dimensions is redesigned to obtain a redesigned collaborative retrieval order, which is synchronized with the current R&D progress. Redefine the data sharing conditions between different core search dimensions, determine the triggering time, data format requirements and sharing scope limitations for data sharing, and obtain the redefined data sharing conditions, which meet the current adaptation requirements; The association judgment criteria between different core search dimensions are reset, the identification basis of data association, the expression of association strength and the processing method of association results are determined, and the reset association judgment criteria are obtained. The reset association judgment criteria improve the adaptability of association logic. The redesigned collaborative retrieval order, the redefined data sharing conditions, and the redefined association judgment criteria are input into a deep-trained multimodal large model. Logical validity verification is then performed using historical interaction data from the R&D scenario to obtain the logical validity verification results. Based on the results of the logical validity verification operation, the redesigned collaborative retrieval order, the redefined data sharing conditions, and the re-set association judgment criteria are adjusted in detail to eliminate logical loopholes and potential conflicts, resulting in the adjusted collaborative retrieval order, the adjusted data sharing conditions, and the adjusted association judgment criteria. The adjusted collaborative retrieval order, the adjusted data sharing conditions, and the adjusted association judgment criteria are integrated to form the adjusted interactive association logic; The adjusted interaction association logic forms an effective linkage with other components in the adapted and adjusted R&D scenario-specific search configuration set, so that the adapted and adjusted R&D scenario-specific search configuration set can respond collaboratively to changes in the adaptation requirements of the R&D scenario.

8. The intelligent big data retrieval method based on the fusion of multimodal large models according to claim 4, characterized in that, Based on the aforementioned collaborative triggering condition, the preliminary data fragments from different retrieval paths are cross-associated to extract the association information between the data from different paths, resulting in a cross-associated data set, including: Monitor the retrieval progress of each retrieval path, and trigger the cross-association process when any retrieval path meets the initial data fragment generation threshold set in the collaborative triggering conditions. Obtain preliminary data fragments generated by retrieval paths that meet the preliminary data fragment generation threshold, and parse the core data identifier and data attribute information in each preliminary data fragment to obtain the first core data identifier and first data attribute information; Obtain the preliminary data fragments currently generated by other parallel retrieval paths, parse the core data identifiers and data attribute information in them, and obtain the second core data identifiers and second data attribute information; Input the first core data identifier, the first data attribute information, the second core data identifier, and the second data attribute information into the association analysis stage to construct a data association index table; Based on the data association index table, data entries with consistent core data identifiers or complementary data attribute information are found in data segments of different retrieval paths to obtain relevant data entries; Based on the attribute matching degree between the relevant data items and the association requirements in the R&D scenario, an association strength analysis operation is performed on the relevant data items to determine the tightness of the data association between the relevant data items. Extract the association information between relevant data entries whose data association is close enough to meet the requirements. The association information includes the data source path, association attribute type, and association logical relationship. The extracted association information is bound to the corresponding related data entries to form association data groups. Each association data group contains related data entries from different search paths and their corresponding association information. Based on the logical relationships in the associated information, the missing associated data is obtained from the relevant search paths, and a completeness supplementation operation is performed on each associated data group to obtain the supplemented associated data group. Summarize all the supplemented related data groups to obtain a cross-related data set.

9. The intelligent big data retrieval method based on the fusion of multimodal large models according to claim 5, characterized in that, The process involves performing a comprehensive scenario adaptation evaluation on each data entry based on its relevance to the current R&D stage, its matching degree with the current hardware status of the data acquisition terminal, and its matching degree with the adaptation requirements of the target application. This yields a comprehensive score for each data entry, including: We define evaluation dimensions for the relevance of data entries to the current R&D stage. Each evaluation dimension corresponds to the specific data requirements of a key link in the R&D stage, resulting in the first set of evaluation dimensions. Set evaluation dimensions for the matching degree between data entries and the current hardware status of the data acquisition terminal. Each evaluation dimension corresponds to the technical parameter requirements of a component in the hardware structure for the data, thus obtaining the second set of evaluation dimensions. The evaluation dimensions for the matching degree between data entries and the adaptation requirements of the target application are set. Each evaluation dimension corresponds to the data compatibility requirements of an adaptation stage in the application docking process, resulting in a third set of evaluation dimensions. Based on the overall needs of the R&D scenario and historical evaluation data, specific evaluation execution standards are set for each evaluation dimension in the first set of evaluation dimensions, the second set of evaluation dimensions, and the third set of evaluation dimensions, resulting in a set of evaluation execution standards. The relevant information corresponding to each data entry is compared one by one with each evaluation execution standard in the set of evaluation execution standards, and the compliance status of each data entry in each evaluation dimension is recorded to obtain the dimension compliance record of each data entry. Based on the importance of each evaluation dimension to the current R&D scenario, a corresponding weight coefficient is assigned to each evaluation dimension in the first set of evaluation dimensions, the second set of evaluation dimensions, and the third set of evaluation dimensions to obtain a set of weight coefficients. Based on the dimensional compliance record of each data entry and the corresponding weight coefficient in the set of weight coefficients, a comprehensive calculation operation of the multi-dimensional evaluation results is performed to obtain the comprehensive evaluation score of each data entry. Based on the distribution range of the comprehensive evaluation scores of all data items, all data items are divided into different adaptation evaluation levels, and each adaptation evaluation level corresponds to a certain degree of adaptation to a scenario. Perform feature analysis on the data entries in each adaptation evaluation level, and extract the common features of the data entries in the adaptation evaluation level in terms of relevance, matching degree and matching degree to obtain the common features of the level. Based on the common characteristics of the levels, a secondary calibration operation is performed on the comprehensive evaluation score of each data item to obtain the comprehensive score result of each data item. The comprehensive score result is used to reflect the actual adaptation of the data item to the current needs of the R&D scenario.

10. An intelligent big data retrieval system integrating multimodal large models, characterized in that, The method includes a processor and a readable storage medium storing a program that, when executed by the processor, implements the intelligent big data retrieval method for fusing multimodal big models as described in any one of claims 1-9.