A multi-source data workbench intelligent comprehensive analysis method and system
By unifying the organization of various data types through a multi-source data workbench and utilizing large language models for intelligent analysis, the problems of evidence correlation and experience accumulation in multimodal data fusion analysis are solved, and efficient and secure collaborative analysis of multi-source data is achieved.
Patent Information
- Application Number
- CN202611100477.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-23
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies cannot automatically correlate structured data anomalies with unstructured evidence in multimodal data fusion analysis. The analysis conclusions lack a complete chain of evidence, and user analysis experience cannot be recorded and accumulated, making it difficult to meet the needs of collaborative analysis of multi-source data with high security requirements.
By using a multi-source data workbench, structured data, semi-structured data, and unstructured information are organized in a unified manner. Intelligent analysis is performed using a large language model to generate analysis results and establish a mapping relationship between the results and the original data. This supports automated and natural language interactive analysis and forms reusable analytical skills.
It achieves a unified analysis space for data from multiple sources, reduces the cost of switching between systems, supports the accumulation and reuse of analytical experience, improves the interpretability and security of analytical conclusions, and meets the requirements of collaborative data analysis with high security requirements.
Smart Images

Figure CN123115178A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, and in particular to an intelligent comprehensive analysis method and system for multi-source data workbench. Background Technology
[0002] With the gradual application of large-scale modeling, semantic retrieval, intelligent question answering, and data analysis technologies in government affairs, finance, auditing, legal affairs, manufacturing, energy, healthcare, and enterprise management, intelligent comprehensive analysis has become an important means to improve business processing efficiency, discover risk clues, and support decision-making. In actual business operations, the objects of analysis typically include structured data, semi-structured data, and unstructured materials: structured data includes database tables, business ledgers, transaction records, and indicator data; semi-structured data includes tables, lists, forms, and ledger documents; and unstructured materials include contracts, policy documents, meeting minutes, reports, scanned copies, images, PDFs, and Word documents. These multi-source business data collectively form the important foundation for business personnel to judge the nature of problems, verify factual evidence, and form analytical conclusions. How to unify and organize the above-mentioned multi-source data and conduct comprehensive analysis in the same analytical task is a current technical challenge facing intelligent analysis systems.
[0003] For example, patent CN121389009A discloses a method and system for cross-media data fusion analysis based on multimodal LLM. This scheme includes: receiving multimodal data from different sources, including text, image, and audio data; inputting the multimodal data into the unified encoder of the multimodal LLM model for joint encoding to generate semantic feature representations with cross-modal associations; inputting the semantic feature representations into the cross-modal alignment module of the multimodal LLM model to generate spatially consistent cross-media embedding representations; using a hierarchical fusion network in the multimodal LLM model to perform multi-scale feature interaction on the cross-media embedding representations to generate semantically enhanced unified fusion features; and inputting the unified fusion features into the multi-task prediction head of the multimodal LLM model to output a comprehensive analysis result with cross-modal consistency. While the above scheme can achieve unified semantic representation and fusion analysis of different media data through multimodal LLM, it still has the following problems when applied to practical business analysis scenarios:
[0004] (1) Existing technologies can only achieve the output of multimodal data fusion analysis. When users discover an anomaly in a certain business through structured data, they still need to manually switch to the document system, knowledge base or local data to find relevant unstructured supporting materials. It is impossible to achieve automatic association between anomaly discovery and evidence verification, and the analysis conclusion lacks a complete chain of evidence.
[0005] (2) In the existing technology, the judgment rules, comparison logic and verification experience formed by users in the analysis process cannot be recorded and accumulated, which leads to repeated configuration and analysis in similar business projects, and cannot form a continuous accumulation of analytical capabilities.
[0006] (3) Existing technologies are designed for single-point analysis tasks and lack the governance capabilities required for organization-level private deployment, such as permission isolation, controlled tool invocation, execution sandbox and cross-task memory management, which makes it difficult to meet the needs of multi-source data collaborative analysis in scenarios with high security requirements.
[0007] Therefore, how to build a comprehensive analysis platform and method that can automatically correlate structured data anomalies with unstructured evidence and support the reuse of analytical experience based on multimodal data fusion analysis has become a technical problem that urgently needs to be solved in this field. Summary of the Invention
[0008] The purpose of this invention is to address the shortcomings of existing technologies by providing an intelligent comprehensive analysis method and system for multi-source data workbench.
[0009] To achieve the above objectives, the present invention adopts the following technical solution:
[0010] A multi-source data workbench intelligent comprehensive analysis method is applied to the workbench, and the analysis method includes:
[0011] S1. Obtain multi-source business data corresponding to the business analysis task, and register the obtained multi-source business data to the workbench;
[0012] S2. Configure the analytical skills required for business analysis tasks and associate the analytical skills with the workbench;
[0013] S3. Within the same workbench context, invoke the large language model and perform intelligent analysis on the business data registered to the workbench according to the configured analysis skills, and generate analysis results; among which intelligent analysis includes automated analysis based on analysis skills and dialogue analysis based on natural language interaction;
[0014] S4. Store the analysis results and establish a mapping relationship between the analysis results and the corresponding business data registered in the workbench.
[0015] Furthermore, the automated analysis based on analytical skills in step S3 specifically includes:
[0016] Acquire the data reading rules, information extraction requirements, and comparison rules configured in the analysis skills;
[0017] The data to be analyzed is retrieved from the business data registered to the workbench according to the data retrieval rules;
[0018] The data to be analyzed, information extraction requirements, comparison rules, and historical analysis results and runtime memory in the workbench are encapsulated into a comprehensive business analysis context;
[0019] The large language model is invoked to generate analysis results that meet the output requirements based on the comprehensive business analysis context.
[0020] Furthermore, the dialogue analysis based on natural language interaction in step S3 specifically includes:
[0021] Receive analysis questions from users via natural language input;
[0022] Based on the analysis problem, unstructured fragments, structured query results, historical analysis results, runtime memory, and analysis skill rules in the business data registered to the workbench are encapsulated into a comprehensive business analysis context;
[0023] The large language model is invoked to generate analysis results based on the comprehensive business analysis context.
[0024] Furthermore, step S3, following the dialogue analysis based on natural language interaction, also includes:
[0025] Transform the analytical logic formed in dialogue analysis into new analytical skills, and store the new analytical skills in the skill library.
[0026] Furthermore, the transformation of analytical logic formed in dialogue analysis into new analytical skills specifically includes:
[0027] Identify reusable analysis objects, areas of interest, decision logic, and output requirements from the results of dialogue analysis;
[0028] Generate a draft of analytical skills based on the identification results;
[0029] Receive user confirmation of the analysis skills draft;
[0030] Based on the user's confirmation, the confirmed analysis skills are saved as available analysis skills for the current workbench and / or added to the skill library.
[0031] Furthermore, step S3, following the automated analysis based on analytical skills, also includes:
[0032] The analysis results obtained from automated analysis are used as input for dialogue analysis, which is then used to further analyze the dialogue results.
[0033] Furthermore, step S3, following the dialogue analysis based on natural language interaction, also includes:
[0034] The analytical logic formed in the dialogue analysis is transformed into new analytical skills, and the new analytical skills are used to automatically analyze the business data registered in the workbench to generate new automated analysis results.
[0035] Furthermore, the specific steps in step S4 of establishing the mapping relationship between the analysis results and the corresponding business data already registered in the workbench are as follows:
[0036] Record the source document identifier and location information within the document for each piece of evidence cited in the analysis results;
[0037] Store records linking evidence citations to source document identifiers and location information within the documents;
[0038] In response to the user's selection of evidence citations in the analysis results, the system locates and displays the corresponding original data fragments based on the associated records.
[0039] Furthermore, the multi-source business data in step S1 includes: unstructured data, semi-structured data, and structured data.
[0040] Correspondingly, a multi-source data workbench intelligent comprehensive analysis is also provided, including:
[0041] The acquisition module is used to acquire multi-source business data corresponding to business analysis tasks and register the acquired multi-source business data to the workbench.
[0042] The configuration module is used to configure the analytical skills required for business analysis tasks and associate the analytical skills with the workbench.
[0043] The analysis module is used to call the large language model and perform intelligent analysis on the business data registered to the workbench within the same workbench context, based on the configured analysis skills, and generate analysis results; the intelligent analysis includes automated analysis based on analysis skills and dialogue analysis based on natural language interaction;
[0044] The mapping module is used to store the analysis results and establish a mapping relationship between the analysis results and the corresponding business data registered in the workbench.
[0045] Compared with the prior art, the present invention has the following beneficial effects:
[0046] (1) Unified analysis space: Through the workbench mechanism, structured data, semi-structured data, unstructured information, analysis skills, AI dialogue process and analysis results are organized in the same business task space to reduce cross-system switching and manual processing costs;
[0047] (2) Multi-source comprehensive analysis: Through large-scale model semantic understanding, data parsing and contextual reference mechanism, the database, knowledge base, document data and existing results are accessed and comprehensively analyzed in a unified manner;
[0048] (3) Reusable analysis methods: Through the two-layer structure of skill library and workbench skills, the analysis rules and experience formed by business personnel are precipitated into reusable assets, reducing the cost of repeated configuration in similar projects;
[0049] (4) Enhanced human-machine collaboration: Through the linkage of skill execution analysis driven by large models and AI dialogue analysis, it supports both the batch execution of standardized skills and interactive follow-up questions and manual judgment by business personnel based on specific issues;
[0050] (5) Results and evidence closed loop: Through the results accumulation and evidence traceability mechanism, the intelligent analysis conclusions are linked with the original data, structured data and analysis skills to improve the interpretability and verifiability of the analysis conclusions;
[0051] (6) Data security and controllability: Through private online platform, permission isolation, execution sandbox and model call governance, sensitive business data can be intelligently analyzed in a controlled environment, reducing the risk of data leakage and unauthorized access;
[0052] (7) Continuous accumulation and optimization: By summarizing the conclusions of AI dialogue analysis into skills, the effective experience in a single analysis is transformed into reusable capabilities, forming a continuous accumulation of business knowledge and analysis methods. Attached Figure Description
[0053] Figure 1 This is a flowchart of an intelligent comprehensive analysis method for multi-source data workbench provided in Embodiment 1. Detailed Implementation
[0054] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.
[0055] The purpose of this invention is to address the shortcomings of existing technologies by providing an intelligent comprehensive analysis method and system for multi-source data workbench.
[0056] Example 1
[0057] This embodiment provides an intelligent comprehensive analysis method for multi-source data workbench, such as... Figure 1As shown, it includes:
[0058] S1. Obtain multi-source business data corresponding to the business analysis task, and register the obtained multi-source business data to the workbench;
[0059] S2. Configure the analytical skills required for business analysis tasks and associate the analytical skills with the workbench;
[0060] S3. Within the same workbench context, invoke the large language model and perform intelligent analysis on the business data registered to the workbench according to the configured analysis skills, and generate analysis results; among which intelligent analysis includes automated analysis based on analysis skills and dialogue analysis based on natural language interaction;
[0061] S4. Store the analysis results and establish a mapping relationship between the analysis results and the corresponding business data registered in the workbench.
[0062] The analysis method in this embodiment is applied to the workbench, which is an analysis space constructed based on business projects, business matters, or special analysis topics. It is used to uniformly associate and store business data, analysis skills, analysis sessions, runtime memories, and analysis results involved in the same business analysis task from multiple sources.
[0063] The workbench context integrates business objects that were originally scattered across document systems, data analysis systems, knowledge base systems, and AI analysis tools, forming a unified analysis space oriented towards business tasks. Through the workbench, business personnel can complete the entire analysis process—from document retrieval, data exploration, and anomaly detection to evidence verification and conclusion formation—within a single interface, without switching between multiple independent systems. Specifically:
[0064] This embodiment generates a unique workbench identifier for each business analysis task, binding the data, skills, sessions, memories, and results under that task to the same workbench context. For example, in a compliance review task for a procurement project, a corresponding workbench is created, linking the project's tender documents, scanned copies of contracts, supplier qualification materials, financial payment records, meeting minutes, and other data, along with pre-configured contract compliance review skills, price comparison skills, and other analytical skills, along with the business personnel's analysis session records, system operation memories, and analysis conclusions, to that workbench.
[0065] In step S1, the multi-source business data corresponding to the business analysis task is obtained, and the obtained multi-source business data is registered to the workbench.
[0066] This process involves acquiring multi-source business data corresponding to business analysis tasks, and uniformly integrating and registering business data from different sources and in different formats into the corresponding workbench. This provides a data foundation for subsequent retrieval, citation, comparison, analysis, and tracing. The multi-source business data includes unstructured, semi-structured, and structured data, which may be stored in various media, including but not limited to documents, tables, database tables, knowledge bases, knowledge graphs, and business application programming interfaces (APIs).
[0067] Unstructured data includes text or image materials such as contracts, policy documents, meeting minutes, rectification materials, business working papers, scanned copies, pictures, PDF files, and Word documents. These materials do not have a fixed field structure and require the use of technologies such as OCR recognition, visual model understanding, and text extraction to extract their content.
[0068] Semi-structured data includes tables, lists, forms, ledgers, and other documents that have a certain field structure but are not entirely uniform in format. This type of data is between structured and unstructured data, and it is necessary to identify the row and column structure, field meanings, and key records.
[0069] Structured data includes business system database tables, ledger data, cash flow records, project lists, and external knowledge base data. This type of data has clear field definitions and data types, and can be directly obtained through database integration, API calls, or knowledge base connections.
[0070] After acquiring the aforementioned multi-source business data, it is registered in the workbench for unified management. Once registration is complete, the association between the workbench and each data item is established, enabling subsequent analysis to retrieve the corresponding data within the workbench's scope. For example, when a user performs intelligent analysis in this workbench, relevant content is automatically retrieved from the registered data, eliminating the need for the user to manually specify file paths or repeatedly upload files.
[0071] This embodiment, while registering data to the workbench, uses appropriate parsing methods to uniformly transform unstructured, semi-structured, and structured data into data content that can be read, referenced, and analyzed by the large model, providing a data foundation for subsequent AI dialogue analysis and skill execution. Specifically:
[0072] For unstructured data, OCR recognition, visual model understanding, and text extraction capabilities are combined to extract the main text, layout structure, table areas, seals and signatures, image descriptions, and key evidence fragments from scanned documents, images, PDFs, Word documents, etc. For example, for a scanned contract document, OCR recognition is used to convert the image into editable text, visual model understanding is used to understand the layout structure of the contract (such as the position of the title, body text, and signature area), and text extraction is used to extract key clauses in the contract (such as amount, term, liability for breach of contract, etc.).
[0073] For semi-structured data, data is directly extracted from tables, lists, forms, etc., and the row and column structure, field meanings, business objects, and key records are identified. For example, for an Excel-formatted supplier list, the correspondence between the header (such as supplier name, registered capital, qualification level) and the data in each row is identified, and each supplier record is extracted.
[0074] For structured data, we obtain structured content such as database table data, business ledgers, cash flow records, and project lists through database integration, API calls, or knowledge base connections; for example, we can retrieve all payment records for a project from a financial database through SQL queries.
[0075] After parsing, the parsed content of different forms of data is uniformly formatted to form text, tables, field descriptions, data fragments, and evidence fragments that can be read by the large language model. At the same time, a mapping relationship between the parsing results and the original data is established, so that the large language model can reference the corresponding data in subsequent analysis and support necessary evidence backtracking.
[0076] In step S2, the analysis skills required for the business analysis task are configured and associated with the analysis skills to the workbench.
[0077] Analytical skills are abstractions of business analysis experience, forming analytical method objects that can be understood and executed by a large language model. Each analytical skill describes its applicable scenarios, analytical objectives, and required data scope. Analytical skills are used to constrain the data reading scope, analytical objectives, judgment logic, tool invocation methods, and output formats of the large model, and support their flow and reuse between the workbench and the skill library. For example, a contract compliance review skill can be configured. The applicable scenario for this skill is the review of contracts for procurement projects. The analytical objective is to determine whether the contract terms comply with the company's procurement system. The required data scope includes contract documents, procurement system documents, and supplier qualification materials.
[0078] Within the skill object, users can configure data reading rules, information extraction requirements, verification requirements, output requirements, doubt judgment logic, comparison rules, and large model reasoning prompt rules.
[0079] Within the analysis skills, the seven core configurable information points are: First, data retrieval rules, used to define the scope and path for large-scale model identification, filtering, and parsing of target multi-source documents, ensuring the accuracy of data entry (e.g., limiting the reading of only key contract sections and OCR parsing); second, information extraction requirements, used to guide the large-scale model to accurately extract specific business entities and metadata from unstructured text (e.g., extracting contract amount, payment method, and supplier name); third, verification requirements, used to set a compliance checklist that must be verified item by item for a single document (e.g., verifying whether the contract is missing necessary attachments and whether the supplier's qualifications are valid); and fourth, output requirements, used to standardize the format, structure, and presentation of the final analysis results, ensuring seamless integration with downstream systems (e.g., outputting structured JSON). The review report); fifth, the doubt judgment logic, used to pre-set anomaly identification mechanisms, enabling the large model to keenly intercept potential business risks (such as judging excessively high down payment ratios or biased disclaimer clauses as major doubts); sixth, the comparison rules, used to define the consistency cross-validation logic between multi-source documents (such as comparing whether the contract signing entity and the business license name are completely consistent); seventh, the large model reasoning prompt rules, used to configure the Prompt template and thinking chain strategy to guide the large model's thinking path, thereby constraining the reasoning process, reducing illusions, and ultimately achieving high-precision, standardized business analysis.
[0080] After configuration, the analysis skills will be associated with the workbench. There are two ways to associate them: one is for the user to reference an existing analysis skill from the skill library to the current workbench; the other is for the user to directly configure a new analysis skill in the current workbench.
[0081] The skill library is a reusable skill storage space that supports skill reuse across workbenches. For example, if an audit department configures a supplier qualification review skill and saves it to the skill library, other workbenches can directly reference the skill from the skill library when performing similar audit tasks without reconfiguration.
[0082] In step S3, within the same workbench context, the large language model is invoked and intelligent analysis is performed on the business data registered to the workbench based on the configured analysis skills to generate analysis results.
[0083] Intelligent analytics includes automated analytics based on analytical skills and dialogue analytics based on natural language interaction. Both approaches are performed within the same workbench context, sharing data, skills, conversations, memories, and the context of results and evidence within the workbench.
[0084] Large language models are generative language models based on the Transformer architecture and pre-trained on massive corpora. They have natural language understanding, reasoning, and generation capabilities and support private deployment methods, such as Qwen and DeepSeek.
[0085] Interact with the large language model through API interface or local call method. When calling, pass the encapsulated comprehensive business analysis context as input to the large language model. The large language model performs inference based on the context content and generates analysis results.
[0086] Automated analysis refers to the process by which a large language model automatically performs analysis according to pre-configured analysis skills. Specifically, automated analysis based on analysis skills includes:
[0087] First, the data reading rules, information extraction requirements, and comparison rules configured in the analysis skill are obtained. These configurations are determined during the skill configuration phase and stored in the skill object. Next, according to the data reading rules, the data to be analyzed is read from the business data registered in the workbench. For example, the data reading rule for the contract compliance review skill specifies reading the contract amount field from the contract file. The system automatically locates the contract file registered in the workbench and extracts the contract amount value. Then, the data to be analyzed, the information extraction requirements, the comparison rules, and the historical analysis results and runtime memory in the workbench are collectively encapsulated into a comprehensive business analysis context. Encapsulation refers to concatenating them into a long text string written in Markdown specifications, which serves as the system prompt word part of the large model prompt words.
[0088] In this embodiment, historical analysis results refer to the conclusions generated by analyses previously performed on the same workbench. For example, if a price analysis was previously performed on this workbench and the result showed that the contract price was 20% higher than the market average, this result will be incorporated into the context of the current automated analysis. Runtime memory refers to the reasoning and evidentiary relationships accumulated on the workbench. During the analysis process, the task memory is updated and the necessary reasoning and evidentiary relationships are retained. For example, the reasoning that the supplier has a relationship was derived in a previous analysis will be recorded in the runtime memory for reference in subsequent analyses.
[0089] Finally, the large language model is invoked, taking the encapsulated comprehensive business analysis context as input. The large language model performs inference based on the context and generates analysis results that meet the output requirements. For example, for the contract compliance review skill, after receiving the encapsulated context, the large language model performs the following inference: reads the contract amount (e.g., 1.5 million yuan), reads the limit standard in the procurement system (e.g., exceeding 1 million yuan requires the general manager's approval), and checks whether there is a general manager's signature in the approval record. If not, it is determined that it has not been approved. The generated analysis result is that the contract amount of 1.5 million yuan exceeds the system limit, but there is no general manager's approval record attached, which is determined to be a compliance risk.
[0090] Dialogue analytics refers to the process by which business users interact with a large language model using natural language, and the large language model answers user questions in real time. Specifically, dialogue analytics based on natural language interaction includes:
[0091] First, the system receives analysis questions input by users via natural language. Users can enter any question related to the current business analysis task in the workbench's dialog interface, such as "Please analyze which clauses in this contract may have compliance risks." Then, based on the user's analysis question, the system automatically organizes relevant content from the business data registered in the workbench, including:
[0092] Unstructured fragments: Extracting fragments relevant to the user's question from unstructured materials such as contracts, policy documents, and meeting minutes. For example, if a user asks about payment terms in a contract, locate and extract the paragraph containing the payment terms from the contract document.
[0093] Structured query results: Retrieve records related to the user's question from structured data such as database tables, ledgers, and transaction records. For example, if the user asks about the supplier's historical transaction records, query the supplier's ledger and return the supplier's historical transaction data.
[0094] Historical analysis results: The conclusions drawn from previous analyses performed in this workbench. For example, if a user asks about the compliance risks identified in previous analyses, relevant information can be extracted from the historical analysis results.
[0095] Running memory: The reasoning and evidentiary relationships accumulated in the workbench.
[0096] Analysis Skill Rules: The judgment logic and comparison rules contained in the analysis skills associated with the workbench. For example, if a user asks whether this contract has a problem according to the standards of the compliance review skill, the judgment logic in the contract compliance review skill will be invoked to answer the question.
[0097] The above content is then encapsulated into a comprehensive business analysis context; finally, the large language model is called, taking the encapsulated comprehensive business analysis context as input, and the large language model generates a response based on the context.
[0098] If a user inputs, "Please help me analyze which clauses in this contract may have compliance risks," the large language model, after receiving the context, generates the following response: "Based on analysis, this contract has the following compliance risks: The payment terms stipulate payment within 30 days of delivery, but the company's procurement policy stipulates payment within 15 days of delivery and acceptance, which is inconsistent; the penalty for breach of contract is 0.1% per day, lower than the company's standard template of 0.5%, which may indicate a risk of improper transfer of benefits."
[0099] In step S4, the analysis results are stored, and a mapping relationship is established between the analysis results and the corresponding business data registered in the workbench.
[0100] The automated analysis results generated by automated analysis and / or the dialogue analysis results generated by dialogue analysis are stored as the analysis conclusions of this workbench. During storage, the relationships between the analysis results and source materials, analysis skills, and analysis processes are recorded. For example, if an analysis result is generated by the contract compliance review skill through automated analysis, the relationship between the analysis result and the contract compliance review skill is recorded; if the analysis result references contract documents and policy documents, these reference relationships are recorded.
[0101] Establish a mapping relationship between the analysis conclusions and the corresponding evidence fragments in the business data already registered in the workbench, specifically as follows:
[0102] The record analysis conclusions should include the source document identifier and the document location information for each piece of evidence. The source document identifier is used to uniquely identify an original document, such as document ID, file path, database table name, etc. The document location information is used to locate the specific position within the document. For text documents, page number, paragraph number, line number, or character offset can be recorded. For tabular documents, worksheet name, row number, and column number can be recorded. For scanned documents, page number and area coordinates can be recorded.
[0103] Store records linking evidence citations to source document identifiers and location information within the documents. These links can be stored in a database associated with the workbench.
[0104] When a user selects evidence citations in the analysis conclusion, the corresponding original data fragments are located and displayed based on the aforementioned associated records.
[0105] This embodiment allows business personnel to trace back from the analysis results to the corresponding original data fragments, table records, business data, or knowledge base content; it also supports the review, revision, sharing, and archiving of the accumulated analysis results.
[0106] Compared with existing technologies, this embodiment features a unified organization mechanism for multi-source business data centered on a workbench, binding structured data, semi-structured data, unstructured data, skills, memories, conversations, and results within the same business task context; a multi-source context comprehensive analysis mechanism based on a large model, combining document fragments, structured data, historical results, runtime memories, and analysis skill rules into a unified analysis context; a linkage mechanism between large model-driven skill execution analysis and AI dialogue analysis sharing the same context, maintaining the same set of workbench data, skills, memories, and results between standardized skill execution and natural language dialogue analysis; a mapping and traceability mechanism between analysis results and evidence data, establishing a traceable association between evidence identifiers in the result text and original data fragments, structured data, or business records; and a mechanism for converting AI dialogue analysis conclusions into skill assets: extracting analysis objects, rules, and output requirements from the dialogue-generated content to form reusable analysis skills.
[0107] Accordingly, this embodiment also provides an intelligent integrated analysis system for multi-source data workbench, including:
[0108] The acquisition module is used to acquire multi-source business data corresponding to business analysis tasks and register the acquired multi-source business data to the workbench.
[0109] The configuration module is used to configure the analytical skills required for business analysis tasks and associate the analytical skills with the workbench.
[0110] The analysis module is used to call the large language model and perform intelligent analysis on the business data registered to the workbench within the same workbench context, based on the configured analysis skills, and generate analysis results; the intelligent analysis includes automated analysis based on analysis skills and dialogue analysis based on natural language interaction;
[0111] The mapping module is used to store the analysis results and establish a mapping relationship between the analysis results and the corresponding business data registered in the workbench.
[0112] Example 2
[0113] The difference between the intelligent comprehensive analysis method for multi-source data workbench provided in this embodiment and Embodiment 1 is as follows:
[0114] In this embodiment, the automated analysis and dialogue analysis do not run in isolation, but share the same workbench context and can call each other and enhance each other in a loop.
[0115] Step S3, following the automated analysis based on analytical skills, also includes: using the analysis results obtained from the automated analysis as input for dialogue analysis, to further analyze the results in dialogue, specifically:
[0116] The results generated from the skill execution analysis path can be further processed into the AI dialogue analysis path for follow-up questions, explanations, verification, and supplementary analysis. For example, if automated analysis finds that a contract worth 1.5 million yuan was not approved by the general manager, after outputting this result, the user can continue to ask in the dialogue analysis: Where are the regulations regarding the general manager's approval authority? Please show the relevant policy clauses. The large language model locates the relevant clauses based on the policy documents in the workbench and returns: According to the company's procurement management system, contracts exceeding 1 million yuan require the general manager's approval. You can refer to the policy documents.
[0117] For example, a user can request verification in the dialogue analysis: Please verify whether this contract amount really exceeds the system limit. The large language model rereads the contract amount and system limit, compares them, and confirms: Verification is correct, the contract amount is 1.5 million yuan, the system limit is 1 million yuan, and it does exceed 500,000 yuan.
[0118] Step S3, following the dialogue analysis based on natural language interaction, also includes: transforming the analysis logic formed in the dialogue analysis into new analysis skills, storing the new analysis skills in the skill library, and using the new analysis skills to automatically analyze the business data registered to the workbench to generate new automated analysis results.
[0119] Effective analytical approaches, judgment logic, or analytical processes formed during AI dialogue analysis can be distilled into analytical skills for reuse in subsequent skill execution analysis paths. High-value answers, analytical steps, or judgment logic generated during AI dialogue analysis can be summarized into new analytical skills. For example, if a user summarizes a method for judging contract compliance through a series of questions and answers during dialogue analysis—that contracts exceeding 1 million yuan require general manager approval, contracts exceeding 5 million yuan require board approval, and that approval documents must be submitted before the contract signing date—this analytical logic can be identified and transformed into a new analytical skill: a contract hierarchical approval and review skill, stored in a skill library. Users can then directly call this skill for automated analysis when performing similar tasks in the future.
[0120] In this embodiment, the transformation process into new analytical skills is as follows:
[0121] Identify the analysis objects (i.e., the business entities involved in the dialogue analysis, such as contract amount and approval limit), the scope of attention (i.e., the attribute range or numerical range corresponding to the analysis object, such as contract amount > 1 million yuan), the judgment logic (i.e., the conditional judgment statements or reasoning rules in the dialogue analysis, such as if a contract amount > 1 million yuan and there is no approval record, it is judged as a compliance risk), and the output requirements (i.e. the conclusion format or presentation method of the analysis results, such as risk level and rectification suggestions) from the analysis results.
[0122] Based on the large model, an analytical skills draft is generated according to the identified analytical objects, scope of concern, judgment logic and output requirements. The skills draft adopts the same data structure as the existing analytical skills.
[0123] The generated skill draft is presented to the user for confirmation. The user can view the various configurations of the skill draft on the interface, and can modify the skill name, adjust the judgment logic, and add output requirements, etc.
[0124] In response to user confirmation, the confirmed analytical skill is saved as an available analytical skill for the current workbench and / or added to the skill library. Saving it as an available analytical skill for the current workbench means that the skill takes effect immediately and can be invoked to perform automated analysis within the current workbench; adding it to the skill library means that the skill is stored in the system-level skill library, making it available for reference and reuse by other workbench. The system also retains the association between the skill and its source analytical process, facilitating subsequent traceability and optimization.
[0125] The two-tiered reuse mechanism for analytical skills in this embodiment supports referencing, copying, embedding, and cross-project reuse between workbench skills and skill library entries.
[0126] Example 3
[0127] The intelligent comprehensive analysis method for multi-source data workbench provided in this embodiment differs from the medical ethics method in that:
[0128] This embodiment can also construct a large model execution sandbox and an analysis process governance mechanism.
[0129] The process of data analysis, skill execution, AI dialogue generation, tool invocation, result saving, and skill accumulation is managed in a unified manner. A sandbox is used to constrain and govern the operation of large models. Specifically:
[0130] Record the execution status of tasks such as data analysis, skill execution, AI dialogue analysis, and tool invocation; for example, when a user triggers the execution of an analysis skill, the system records the start time, current progress, and execution status (waiting, executing, completed, failed, etc.) of the task and displays it to the user on the interface.
[0131] For time-consuming tasks, process feedback is provided; for example, when the analysis skills require batch review of a large number of contract documents, the system displays the current processing progress to the user (e.g., 3 / 10 of the contracts have been completed), so that the user understands the progress of the analysis.
[0132] For interruptible intelligent generation processes, the system provides the ability to pause or terminate them; if the user finds that the analysis direction is incorrect or that it is no longer necessary to continue, the user can pause or terminate the ongoing analysis task through the interface.
[0133] By implementing a sandbox, the scope of data that large models can access, the scope of tools they can invoke, and the scope of external connections they can execute are restricted. For example, the system can configure sandbox rules to stipulate that large models can only access business data registered in the current workbench and cannot access data in other workbenches; they can only invoke whitelisted analysis tools (such as data query tools and document retrieval tools) and cannot invoke unauthorized tools; they can only connect to allowed external data sources and cannot access the external Internet.
[0134] Historical analysis sessions, tool call records, and execution processes are saved for later recovery, review, and record keeping. For example, users can view the complete analysis history of a workbench, including the input parameters, output results, the version of the large model called, and the execution time for each skill execution.
[0135] It provides unified management capabilities for data, skills, and results, ensuring the sustainable maintenance of the analysis process and outputs; for example, users can edit configured skills, delete unused data, and archive the analysis results of completed workbenches.
[0136] The large model execution sandbox mechanism in this embodiment constrains the range of data that the large model can access, the range of tools that can be invoked, and the range of external connections that can be executed, thereby achieving controlled intelligent analysis.
[0137] Example 4
[0138] The intelligent comprehensive analysis method for multi-source data workbench provided in this embodiment differs from the previous embodiment in that:
[0139] This embodiment can also construct a private online platform and a data security guarantee mechanism.
[0140] By leveraging the deployment and management capabilities of a private online platform, external data sources, permission scopes, model configurations, and runtime configurations are maintained, providing a secure and controllable operating environment for intelligent analysis of multi-source business data. Specifically:
[0141] It supports deployment of the platform in an organization's private environment or dedicated network environment; for example, a government agency can deploy the system in its internal government cloud, where all business data is processed in the internal network environment without going through the public Internet, thereby reducing the risk of sensitive business data being leaked.
[0142] Maintain the access configuration of databases, knowledge bases, file storage, or other external data sources, and verify their availability; for example, the administrator configures the connection information of the financial database (IP address, port, authentication method), configures the API interface address of the knowledge base, and configures the path of the configuration file storage. The system verifies whether these data sources can be accessed normally when it starts up.
[0143] Access can be isolated by organization, tenant, role, workspace, or workbench to manage the data, skills, results, and model capabilities that users can access, such as:
[0144] Organizational dimension: Data between different organizations (such as different subsidiaries under a group) is isolated from each other. Users of company A cannot access the workbench and data of company B.
[0145] Tenant dimension: In the SaaS deployment model, the data of different tenants are isolated from each other.
[0146] Role Dimension: Within the same workbench, auditors can configure skills and perform analyses, audit assistants can only view analysis results and cannot modify skills, and audit supervisors can review and confirm analysis conclusions.
[0147] Workspace dimension: The same user has different permissions in different workspaces. For example, a user may have full permissions in the workspace they created, but only read-only permissions in the shared workspace.
[0148] Workbench dimension: Users can only access authorized workbenches; unauthorized workbenches are not visible.
[0149] By managing data sources, controlling access, implementing execution sandboxes, recording operations, and governing model calls, we ensure data boundaries and access security during business collaboration. For example, we record each user's operation logs (when they accessed which workbench, what skills they used, and which large model they called), which facilitates post-event auditing and traceability.
[0150] The private deployment and permission isolation mechanism in this embodiment controls the access boundaries of business data, analytical skills, model capabilities, and analytical results through dimensions such as organization, tenant, role, workspace, or workbench.
[0151] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.
Claims
1. A multi-source data workbench intelligent comprehensive analysis method, characterized in that, Applied to the workbench, the analysis methods include: S1. Obtain multi-source business data corresponding to the business analysis task, and register the obtained multi-source business data to the workbench; S2. Configure the analytical skills required for business analysis tasks and associate the analytical skills with the workbench; S3. Within the same workbench context, invoke the large language model and perform intelligent analysis on the business data registered to the workbench according to the configured analysis skills, and generate analysis results; among which intelligent analysis includes automated analysis based on analysis skills and dialogue analysis based on natural language interaction; S4. Store the analysis results and establish a mapping relationship between the analysis results and the corresponding business data registered in the workbench.
2. The method according to claim 1, wherein, The automated analysis based on analytical skills in step S3 specifically includes: Acquire the data reading rules, information extraction requirements, and comparison rules configured in the analysis skills; The data to be analyzed is retrieved from the business data registered to the workbench according to the data retrieval rules; The data to be analyzed, information extraction requirements, comparison rules, and historical analysis results and runtime memory in the workbench are encapsulated into a comprehensive business analysis context; The large language model is invoked to generate analysis results that meet the output requirements based on the comprehensive business analysis context.
3. The method of claim 1, wherein, The dialogue analysis based on natural language interaction in step S3 specifically includes: Receive analysis questions from users via natural language input; Based on the analysis problem, unstructured fragments, structured query results, historical analysis results, runtime memory, and analysis skill rules in the business data registered to the workbench are encapsulated into a comprehensive business analysis context; The large language model is invoked to generate analysis results based on the comprehensive business analysis context.
4. The method of claim 1, wherein, Step S3, following the dialogue analysis based on natural language interaction, also includes: Transform the analytical logic formed in dialogue analysis into new analytical skills, and store the new analytical skills in the skill library.
5. The method of claim 4, wherein, The specific steps of transforming the analytical logic formed in dialogue analysis into new analytical skills include: Identify reusable analysis objects, areas of interest, decision logic, and output requirements from the results of dialogue analysis; Generate a draft of analytical skills based on the identification results; Receive user confirmation of the analysis skills draft; Based on the user's confirmation, the confirmed analysis skills will be saved as available analysis skills for the current workbench and / or added to the skill library.
6. The method of claim 1, wherein, The automated analysis based on analytical skills in step S3 also includes: The analysis results obtained from automated analysis are used as input for dialogue analysis, which is then used to further analyze the dialogue results.
7. The method of claim 1, wherein, The dialogue analysis based on natural language interaction in step S3 also includes: The analytical logic formed in the dialogue analysis is transformed into new analytical skills, and the new analytical skills are used to automatically analyze the business data registered to the workbench to generate new automated analysis results.
8. The method of claim 1, wherein, The specific steps in step S4 of establishing the mapping relationship between the analysis results and the corresponding business data already registered in the workbench are as follows: Record the source document identifier and location information within the document for each piece of evidence cited in the analysis results; Store records linking evidence citations to source document identifiers and location information within the documents; In response to the user's selection of evidence citations in the analysis results, the system locates and displays the corresponding original data fragments based on the associated records.
9. The method of claim 1, wherein, The multi-source business data in step S1 includes: unstructured data, semi-structured data, and structured data.
10. An analysis system based on the intelligent comprehensive analysis method of multi-source data workbench according to any one of claims 1-9, characterized in that, include: The acquisition module is used to acquire multi-source business data corresponding to business analysis tasks and register the acquired multi-source business data to the workbench. The configuration module is used to configure the analytical skills required for business analysis tasks and associate the analytical skills with the workbench. The analysis module is used to call the large language model and perform intelligent analysis on the business data registered to the workbench in the same workbench context, based on the configured analysis skills, and generate analysis results. Intelligent analysis includes automated analysis based on analytical skills and dialogue analysis based on natural language interaction; The mapping module is used to store the analysis results and establish a mapping relationship between the analysis results and the corresponding business data registered in the workbench.
Citation Information
Patent Citations
Cross-media data fusion analysis method and system based on multi-modal LLM
CN121389009A
