Question number result construction method based on AI large model in combination with process arrangement
By splitting complex business problems into independent nodes and combining with the process orchestration editor to process, the problem of insufficient processing capabilities of a single large language model in complex business scenarios is solved, and the modular and efficient processing of tasks is achieved.
Patent Information
- Application Number
- CN202411966989.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-30
- Publication Date
- 2025-04-25
AI Technical Summary
In complex business scenarios, it is difficult to effectively handle multi-level tasks, conditional judgment and cross-module collaboration in interactions that rely solely on large language model LLM, and the limitations of context length and model parameter scale lead to inability to deal with complex task chains.
The question result construction method based on AI large model combined with process orchestration is adopted to abstract and split complex business problems into multiple independent nodes and subprocesses, and user problems are handled separately through question-answer models and question-and-count models, and visual display and association mapping are used to use the process orchestration editor.
By disassembling complex tasks into task packages with single functions and independent functions, the modularity and reusability of tasks are achieved, the probability of errors under complex conditions is reduced, and the work efficiency and business processing capabilities of AI models are improved.
Smart Images
Figure CN119884148A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of AI big model intelligent mathematical technology, and in particular to a mathematical result construction method based on AI big model combined with process orchestration. Background Art
[0002] Although large language models (LLMs) have made significant progress in understanding and processing complex natural languages, there are still many problems in dealing with complex business scenarios when simply relying on the interaction between prompt words and large language models (LLMs). Complex businesses usually involve multi-level tasks, conditional judgments, and cross-module collaboration, which makes it difficult for simple natural language input to cover the entire process. The current capacity limitations of large language models (LLMs), especially the limitations on context length and the impact of model parameter scale, further make them unable to cope with complex task chains. Summary of the invention
[0003] In order to solve the above problems, the present invention provides a method for constructing question results based on an AI big model combined with process orchestration, which abstracts and splits complex business problems and converts them into multiple independent nodes and sub-processes, effectively simplifying the complexity of business processing.
[0004] The technical solution adopted by the present invention is to provide a method for constructing question and answer results based on a large AI model combined with process orchestration, including using a large language model to analyze and summarize user questions and then output question results, and also including, based on the large language model, respectively constructing a question-answering model and multiple question and answer models, each question and answer model contains databases in different fields. On this basis, the method for constructing question and answer results includes:
[0005] After the user inputs the question, a logical judgment is performed to determine whether the user's question is of the question-answer type or the question-number type;
[0006] When the judgment result is a question-answering model, the question-answering model receives the user's question, and after analysis and summary, outputs the question result;
[0007] When the judgment result is a question-number model, the field to which the user's question belongs is judged again, and the question-number model of the corresponding field is selected for analysis and summary, and then the question result is output.
[0008] The question-answering model is obtained after training based on a large language model, and the question-answering model is obtained after training based on a large language model using data sets from different fields to obtain question-answering models in multiple different fields.
[0009] The question model is divided according to industry fields.
[0010] Logical reasoning methods include:
[0011] Extract model keywords and model keywords of question-answering models and question-number models in multiple fields, and on this basis, determine whether the user question contains the same keywords as the model keywords. If so, select the question-answering model or the question-number model according to the model keywords. If not, determine again whether the user question starts or ends with the same keywords as the model keywords. If so, select the question-answering model or the question-number model according to the model keywords. If not, determine again whether the user question contains keywords similar to the model keywords. If so, select the question-answering model or the question-number model according to the similar model keywords.
[0012] The question model is analyzed and summarized, specifically including:
[0013] a. After the question-asking model receives the user's question, it uses the prompt word engineering to extract the prompt words for the user's question;
[0014] b. Configure multiple sets of tables, columns, and relationships related to the user's question based on the prompt words, and use a similarity algorithm to filter them to obtain SQL query statements;
[0015] c. Query the database according to the SQL query statement to obtain the result of the problem.
[0016] It also includes building a process orchestration editor and associating and mapping it with logical judgment, question-and-answer model, and question-number model.
[0017] In the process orchestration editor, logical judgment, question-and-answer model, and question-and-number model are all visualized using graphics containing text as nodes and are associated and mapped separately.
[0018] The beneficial effect of the present invention is that it provides a method for constructing numerical results based on an AI big model combined with process orchestration, and the AI big model is combined with workflow orchestration. By decomposing complex tasks into single-function, independent task packages, workflow orchestration can not only decompose complexity, but also realize the modularity and reusability of tasks. This decomposition method ensures that each task package can be optimized and executed individually, reducing the probability of errors under complex conditions. And it can provide flexible node customization and maintenance functions, and can customize process nodes that meet business scenarios according to business needs. To expand the diversity and flexibility of process flows. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION
[0020] like Figure 1As shown, the present invention provides a method for constructing question and answer results based on an AI big model combined with process orchestration, including using a big language model to analyze and summarize user questions and then outputting question results, and also including, based on the big language model, constructing a question and answer model and multiple question and answer models respectively, each question and answer model contains databases in different fields. On this basis, the question and answer result construction method includes,
[0021] After the user inputs the question, a logical judgment is performed to determine whether the user question is of the question-answer type or the question-number type;
[0022] When the judgment result is a question-answering model, the question-answering model receives the user's question, and after analysis and summary, outputs the question result;
[0023] When the judgment result is a question-number model, the field to which the user's question belongs is judged again, and the question-number model of the corresponding field is selected for analysis and summary, and then the question result is output.
[0024] like Figure 1 As shown, the question-answering model is obtained after training based on a large language model, and the question-answering model is obtained after training based on a large language model using data sets from different fields to obtain question-answering models in multiple different fields.
[0025] like Figure 1 As shown, the question model is divided according to industry fields.
[0026] like Figure 1 As shown, the method of logical judgment includes:
[0027] Extract model keywords and model keywords of question-answering models and question-number models in multiple fields, and on this basis, determine whether the user question contains the same keywords as the model keywords. If so, select the question-answering model or the question-number model according to the model keywords. If not, determine again whether the user question starts or ends with the same keywords as the model keywords. If so, select the question-answering model or the question-number model according to the model keywords. If not, determine again whether the user question contains keywords similar to the model keywords. If so, select the question-answering model or the question-number model according to the similar model keywords.
[0028] like Figure 1 As shown, the question model is analyzed and summarized, specifically including:
[0029] a. After the question-asking model receives the user's question, it uses the prompt word engineering to extract the prompt words for the user's question;
[0030] b. Configure multiple sets of tables, columns, and relationships related to the user's question based on the prompt words, and use a similarity algorithm to filter them to obtain SQL query statements;
[0031] c. Query the database according to the SQL query statement to obtain the result of the problem.
[0032] like Figure 1 As shown, it also includes building a process orchestration editor and associating and mapping it with logical judgment, question and answer model, and question and number model.
[0033] like Figure 1 As shown, in the process arrangement editor, the logical judgment, question-answer model, and question-number model are all visualized with graphics containing text as nodes and are associated and mapped respectively.
[0034] The present invention provides a method for constructing question and answer results based on a large AI model combined with process orchestration, which combines large language model question and answer with process orchestration, abstracts and splits complex business problems, and converts them into multiple independent nodes, effectively simplifying the complexity of business processing. Each node can flexibly run in parallel or serially to ensure efficient task collaboration. In this way, complex and changeable business processes can be clarified and modularized, thereby improving the accuracy and comprehensiveness of the AI large model in answering questions for users, increasing the maintainability and scalability of the AI large model, and significantly improving the work efficiency and business processing capabilities of the AI large model.
[0035] The present invention significantly improves the efficiency and accuracy of converting user questions into SQL parsing by gradually reasoning and disassembling the dependencies of complex table definitions, thereby significantly improving the response speed of the AI large model and the accuracy of the results.
[0036] The question-answering model and multiple question-answering models in the present invention are obtained after training with different data sets based on the large language model. The question-answering models are divided according to industry fields so that user questions can be processed more efficiently and accurately.
[0037] Compared with the traditional single AI large model, after the user inputs the question, the present invention will encapsulate the input into the sys.query variable and save it to the service's variable cache pool variable_pool so that subsequent nodes can quickly access and use it in the process.
[0038] At the same time, it enters the logic judgment node, which selects a suitable branch node from multiple migration paths connected to the node, that is, matches the user's question with a suitable large model. The logic judgment node performs logic judgment based on sys.query stored in variable_pool. The judgment rules usually include the following categories:
[0039] Contains / Not Contains: Determines whether sys.query contains or does not contain specific keywords.
[0040] Start With / End With: Determine whether sys.query starts or ends with the specified keyword.
[0041] Is / Not Is: Determines whether sys.query is equal to or not equal to a specific keyword.
[0042] By judging the above logical conditions, it is possible to determine along which migration line the workflow should continue to flow, that is, to assign user questions to the appropriate large model for processing.
[0043] When the user question enters the matched large model node, the node will form a model prompt based on sys.query in variable_pool and other user variables, and send an interaction request to the large language model LLM. The result returned by the large model will be written to the sys.answer variable and stored in variable_pool again to support the business logic execution and decision analysis of subsequent nodes.
[0044] The specific processing process of the question number model after receiving the user's question includes:
[0045] 1. Through multiple prompt words, select tables, columns, and relationships related to the problem from the database according to the configuration of the user's problem and key fields. Exclude database tables and columns that are not related to the problem to convert into SQL queries.
[0046] (1) Table Reasoning Phase
[0047] User questions are input into the question-answering model, which extracts the reference tables required to generate SQL queries. Drawing on the zero-shot-CoT method, the large language model (LLM) is required to explain the necessity of each table instead of just listing the table names, and sort them according to the field priority defined in the large language model.
[0048] (2) Column Reasoning Stage
[0049] Large Language Model (LLM) uses hints similar to table linking to extract the columns needed to translate the question into a SQL query. It only includes the table structure selected during the table linking phase, rather than the entire database schema.
[0050] 2. Combining CoT chain thinking and prompt word engineering, based on the question similarity method, we first screen out the top k questions in the training set of the large language model that are closest to the natural language question embedding of the test sample. In order to shield the database-related information, the question similarity method ignores the details in the database, allowing us to accurately evaluate the similarity of questions when generating similar queries, and generate the required SQL query statements after screening based on the similarity.
[0051] 3. Abnormal analysis
[0052] According to the generated SQL statements, pre-query analysis is performed in the database, and abnormal SQL statements are re-analyzed and reasoned. By combining abnormal SQL statements with CoT chain thinking, the large language model LLM is guided to re-analyze and judge abnormal problems, and assembled into new SQL statements for query processing to obtain the query results. The details are as follows:
[0053] (1) SQL pre-query analysis
[0054] Execute the generated SQL statements in the database and catch possible exceptions, such as syntax errors, table non-existence, data type mismatch, etc.
[0055] (2) Record abnormal SQL statements and corresponding error messages.
[0056] (3) Anomaly detection and reasoning
[0057] Using CoT technology and error information, the large language model (LLM) is guided to analyze the cause of the exception, for example, to determine whether it is a grammatical problem, missing table or field, inconsistent data type, etc.
[0058] When analyzing, you can try the following in order: check SQL syntax, confirm table structure, check field types, constraints, etc.
[0059] Based on the inferred abnormal reasons, gradually optimize the SQL statements, for example, add missing tables or fields, fix syntax errors, modify field types, etc.
[0060] About the Process Choreography Editor:
[0061] The present invention constructs a visual editor for associating and mapping with the question-answer model and the question-number model, namely, a process arrangement editor. In the editor, graphics with text are used to visualize the process nodes of the logical judgment, question-answer model, question-number model and individual models, and perform associative mapping.
[0062] The process orchestration editor adopts the workflow process orchestration framework, which can flexibly manage and configure each node, and process tasks based on dynamic logical judgment, question-and-answer model, and question-and-number model according to different business needs.
[0063] After the user inputs a question, the direction of the user's question is determined through logical judgment. If the question belongs to the question-answering category, the migration line is used to further determine whether it contains information such as pictures and tables, and a multi-dimensional summary and analysis is achieved through a large model. If the question belongs to the question-answering category, it enters the intelligent question-answering node through the migration line for processing.
Claims
1. A method for constructing question results based on a large AI model combined with process orchestration, including using a large language model to analyze and summarize user questions and then output question results, characterized in that: It also includes constructing a question-answering model and multiple question-number models based on the large language model, each question-number model contains databases in different fields. On this basis, the question-number result construction method includes: After the user inputs the question, a logical judgment is performed to determine whether the user question is of the question-answer type or the question-number type; When the judgment result is a question-answering model, the question-answering model receives the user's question, and after analysis and summary, outputs the question result; When the judgment result is a question-number model, the field to which the user's question belongs is judged again, and the question-number model of the corresponding field is selected for analysis and summary, and then the question result is output.
2. According to claim 1, a method for constructing numerical results based on an AI big model combined with process orchestration is characterized by: The question-answering model is obtained after training based on a large language model, and the question-answering model is obtained after training based on a large language model using data sets from different fields to obtain question-answering models in multiple different fields.
3. According to claim 2, a method for constructing numerical results based on an AI big model combined with process arrangement is characterized in that: The question model is divided according to industry fields.
4. According to claim 1, a method for constructing numerical results based on an AI big model combined with process orchestration is characterized by: Logical reasoning methods include: Extract model keywords and model keywords of question-answering models and question-number models in multiple fields, and on this basis, determine whether the user question contains the same keywords as the model keywords. If so, select the question-answering model or the question-number model according to the model keywords. If not, determine again whether the user question starts or ends with the same keywords as the model keywords. If so, select the question-answering model or the question-number model according to the model keywords. If not, determine again whether the user question contains keywords similar to the model keywords. If so, select the question-answering model or the question-number model according to the similar model keywords.
5. According to claim 1, a method for constructing numerical results based on AI big model combined with process arrangement is characterized by: The question model is analyzed and summarized, specifically including: a. After the question-asking model receives the user's question, it uses the prompt word engineering to extract the prompt words for the user's question; b. Configure multiple sets of tables, columns, and relationships related to the user's question based on the prompt words, and use a similarity algorithm to filter them to obtain SQL query statements; c. Query the database according to the SQL query statement to obtain the result of the problem.
6. According to claim 1, a method for constructing numerical results based on AI big model combined with process arrangement is characterized by: It also includes building a process orchestration editor and associating and mapping it with logical judgment, question-and-answer model, and question-number model.
7. According to claim 6, a method for constructing numerical results based on AI big model combined with process arrangement is characterized by: In the process orchestration editor, logical judgment, question-and-answer model, and question-and-number model are all visualized using graphics containing text as nodes and are associated and mapped separately.