Data resource retrieval method and system based on large model technology
By employing large model technology and RAG retrieval, the problem of insufficient semantic understanding in traditional data resource retrieval is solved, achieving more accurate and efficient data resource retrieval, and is applicable to multiple operating systems and databases.
Patent Information
- Application Number
- CN202511100498.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-07
- Publication Date
- 2025-11-18
AI Technical Summary
Traditional data resource retrieval methods rely on keyword matching, which makes it difficult to handle users' implicit semantics and complex logical relationships, resulting in inaccurate or incomplete search results and a lack of semantic understanding capabilities.
Employing large-scale model technology, combined with multi-turn dialogue processing and RAG retrieval, it generates complete questions, splits questions, retrieves data resource catalog information, and generates SQL query statements, supporting multiple database systems.
It improves the accuracy and efficiency of data resource retrieval, supports multiple operating systems and databases, meets the requirements of domestic IT innovation systems, and has high data security.
Smart Images
Figure CN120973902A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of large model technology, specifically to a method and system for data resource retrieval based on large model technology. Background Technology
[0002] Driven by the digital wave, data resources across industries are experiencing explosive growth. Enterprises have accumulated massive amounts of business data, covering customer information, transaction records, market dynamics, and more; the scientific research field has generated a wealth of experimental data, research reports, and academic literature; and the internet is teeming with hundreds of millions of web pages and social media posts. These data resources are not only vast in quantity but also diverse in type, including structured data (such as tabular data in databases), semi-structured data (such as data in XML and JSON formats), and unstructured data (such as text, images, audio, and video).
[0003] Faced with such a massive amount of data resources, quickly and accurately retrieving the required information has become a key need for both businesses and users. However, traditional data resource retrieval methods are mainly based on keyword matching and simple Boolean logic operations. Users need to accurately input keywords when searching, and the system returns results based on the occurrence of those keywords in the data. But this method has serious shortcomings. Users often find it difficult to accurately grasp the keywords that describe their needs, resulting in inaccurate or incomplete search results, with a large amount of relevant information being missed. At the same time, traditional retrieval lacks semantic understanding capabilities and cannot handle the implicit semantics and complex logical relationships in user queries, making it difficult to provide satisfactory answers to fuzzy or ambiguous queries. Summary of the Invention
[0004] The purpose of this invention is to provide a method and system for data resource retrieval based on large model technology, so as to solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for data resource retrieval based on large model technology, comprising the following core process steps:
[0006] The request receiving process involves receiving the user request, verifying the token's validity, and parsing the request parameters.
[0007] The step of reading configuration information involves reading process configuration information, including the URL hyperlink address of the data resource directory system and the data resource database connection information.
[0008] The problem preprocessing steps involve preprocessing the problem, removing illegal characters, and determining whether the length of the processed problem meets the minimum length α. If not, the user is prompted that there is insufficient available information for the problem.
[0009] Multi-round dialogue processing step, judge whether the current session has historical information, if yes, judge whether the current user question is an extension of the historical user question by the large model, if the return result contains yes, the large model generates a complete question according to the historical user question content and the current user question;
[0010] Application scenario recall step, send the final user question processed by the large model to the application scenario knowledge base for recall processing, take the TOP1 entry with the highest similarity when the similarity is less than β, if there is no entry with a similarity higher than β, call the large model to rewrite the question, use the split question to perform knowledge recall again, if still none, prompt the user that no relevant data can be found;
[0011] Use purpose extraction step, according to the application scenario, call the large model to extract possible use directions;
[0012] Data resource directory recall step, convert the data recalled by the application scenario into an array, and loop the data resource directory knowledge recall, the recall result is a UUID unique identifier;
[0013] Database query step, according to the obtained unique identifiers, generate a SQL query statement, query the data resource directory information, support MYSQL, POSTGRE and Han Gao database;
[0014] Answer organization step, give corresponding prompts according to the similarity threshold range;
[0015] Answer return step, provide a blocking one-time return of the answer result or an SSE character stream return similar to the typewriter method.
[0016] Preferably, in the question preprocessing step, the rules for eliminating illegal characters are: [’·°–!"#$%&'()*+,-. / :;<=>?@,。?★、…【】()《》?“”‘’![\]^_`{|}~]+?.
[0017] Preferably, in the multi-round dialogue processing step, when the large model judges whether the current user question is an extension of the historical user question, part of the prompt is: you are a dialogue system for processing a single task, the current task is: judge whether the current user question is an extension of the historical user question; the output content needs to include an analysis and a judgment result; if the analysis conclusion is that the current question is an extension of the historical question, the judgment result must contain yes; if the current question is not an extension of the historical question, the judgment result is no; if the large model returns a result containing yes, the core prompt for the large model to generate a complete question according to the historical user question content and the current user question is: you are a dialogue system for processing a single task, please generate a complete question according to the historical user question content and the current user question.
[0018] Preferably, in the application scenario recall step, the core prompt for calling the large model to rewrite the question is: <instruction> please organize refined language to generate the split result according to the known information. The format must be task={‘task_i’: question split result}. It is used to split multiple questions or requirements into single query questions. Please output the question split result of the last piece of known information in strict accordance with the example format. If it cannot be converted into NewInstruction, please say “according to the known information, the question cannot be answered”; the answer should be in Chinese.
[0019] Preferably, in the use extraction step, the core prompt for calling the large model to extract possible use directions is: please organize refined language to infer the possible use of
known information
known information
[0020] A system for a data resource retrieval method based on large model technology, comprising:
[0021] A request receiving module is configured to receive a user request, verify whether a token is legal, and parse request parameters.
[0022] A configuration information reading module is connected to the request receiving module and is configured to read process configuration information. The configuration information includes, but is not limited to, data resource directory system URL hyperlink addresses and data resource database connection information.
[0023] A question preprocessing module is connected to the configuration information reading module and is configured to preprocess the question, remove illegal characters according to a preset rule, and determine whether the length of the processed question meets a minimum length α. If not, the user is prompted that there is not enough information in the question.
[0024] A multi-round dialogue processing module is connected to the question preprocessing module and is configured to determine whether there is historical information in the current session. If there is, the large model is called to determine whether the current user question is an extension of the historical user question. If the large model returns a result containing an identifier indicating an extension, the large model is required to generate a complete question according to the historical user question content and the current user question, and output the final user question processed by the large model.
[0025] The application scenario recall module is connected with the multi-turn dialogue processing module, and is configured to send the final user question processed by the large model to an application scenario knowledge base for recall processing. When the similarity is less than β, the TOP1 entry with the highest similarity is taken. If there is no entry with a similarity higher than β, the large model is called to rewrite the question, the question is split and used to perform knowledge recall, and the TOP1 entry with the highest similarity is taken. If the similarity is still less than β, the user is prompted that relevant data cannot be found.
[0026] The use purpose extraction module is connected with the application scenario recall module, and is configured to call the large model to extract possible use purposes according to the application scenario.
[0027] The data resource directory recall module is connected with the use purpose extraction module, and is configured to convert the data recalled by the application scenario into an array and perform data resource directory knowledge recall in a loop. The recall result is a UUID unique identifier.
[0028] The database query module is connected with the data resource directory recall module, and is configured to generate a SQL query statement according to the obtained unique identifiers, query the data resource directory information, and support MYSQL, POSTGRE and Han Gao databases.
[0029] The answer organization module is connected with the database query module, and is configured to give corresponding prompts according to the similarity threshold range.
[0030] The answer return module is connected with the answer organization module, and is configured to provide a blocking one-time return of an answer result or an SSE character-by-character stream return in a typewriter mode.
[0031] Preferably, in the multi-turn dialogue processing module, when the large model is called to determine whether the current user question is an extension of the historical user question, the prompt input to the large model is: you are a dialogue system for processing a single task, and the current task is: determining whether the current user question is an extension of the historical user question. The output content needs to include an analysis and a determination result. If the analysis conclusion is that the current question is an extension of the historical question, the determination result must include yes. If the current question is not an extension of the historical question, the determination result is no.
[0032] Preferably, in the application scenario recall module, when calling the large model to rewrite the question, the prompt input to the large model is: <Instruction> Please organize refined language to generate the splitting result based on the known information. The format must be task = {'task_i': question splitting result}, used to split multiple questions or requirements into a single question or requirement query question. Please strictly follow the example format to output the question splitting result of the last known information; if it cannot be converted into NewInstruction, please say "This question cannot be answered based on the known information"; please use Chinese for the answer.
[0033] Preferably, in the usage extraction module, when calling the large model to extract possible usage directions, the prompt input to the large model is: Please organize and refine language based on the [known information] to reason about the possible uses of the [known information]. Only the reasoning result of the last existing information needs to be output. With the help of the professional knowledge of the language model, the usage reasoning is performed from the fields, contents, categories, and attribute information that the table may contain.
[0034] Preferably, in the answer organization module, the prompts given based on the similarity threshold range are as follows: High similarity prompts "Based on the above analysis, the following matching data assets have been found:"; Medium similarity prompts "Based on the above analysis, the following data may meet some requirements:"; Low similarity prompts "Unable to find data for you. We suggest you obtain it through other means. If you encounter any problems during use, please contact the administrator."
[0035] Compared with the prior art, the beneficial effects of the present invention are:
[0036] This invention proposes a method and system for data resource retrieval based on large model technology, which creatively applies large model technology to the field of data resource catalog retrieval, improving data utilization efficiency. Combined with RAG retrieval enhancement generation technology, it solves the problems of general large models in terms of knowledge limitations, illusions, data security, training costs, and timeliness. It offers high data security, ensuring local knowledge remains within the domain. It supports CPUs such as Intel x86, AMD64, Hygon x86, and Kunpeng ARM architectures; common Linux systems such as Ubuntu and CentOS; and domestic operating systems such as Kylin. It also supports common databases such as MySQL and PostgreSQL, as well as the domestic HANGGUO database, meeting the requirements of domestic IT innovation systems. Attached Figure Description
[0037] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0038] To make the objectives, technical solutions, and advantages of the present invention clear and complete, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only some, not all, embodiments of the present invention, and are merely illustrative of the embodiments of the present invention. They are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0039] Example 1, please refer to Figure 1 This invention provides a technical solution: a method for data resource retrieval based on large model technology, comprising the following core process steps:
[0040] S1, Request to receive
[0041] Receive user requests, verify the validity of the token, and parse the request parameters.
[0042] S2, Read configuration information
[0043] Read process configuration information, including but not limited to: data resource catalog system URL hyperlink address, data resource database connection information, etc.
[0044] S3, Problem Preprocessing
[0045] The problem is preprocessed using the following rules to remove illegal characters:
[0046] ['·°–! "#$%&\'()*+,-. / :;<=>?@,。?★、…【】()《》?“”''![\\]^_`{|}~]+?
[0047] Determine if the length of the processed question meets the minimum length α. If not, prompt the user that there is insufficient available information for the question.
[0048] S4, Multi-turn Dialogue Processing
[0049] Determine if the current session has historical information. If so, the larger model is required to determine if the current user's question is an extension of a historical user's question. Part of the prompt is as follows (some content has been omitted):
[0050] You are a dialogue system that handles a single task. The current task is to determine whether the current user's question is an extension of a previous user's question.
[0051] The output should include an analysis and judgment result.
[0052] If the analysis conclusion is that the current problem is an extension of a historical problem, then the determination result must include "yes"; if the current problem is not an extension of a historical problem, the determination result is "no".
[0053] The following is omitted...
[0054] If the result returned by the large model contains "yes", then the large model is required to generate a complete question based on the historical user question content and the current user question. The core prompt is as follows (some content is omitted):
[0055] You are a dialogue system for handling single tasks. Please generate a complete question based on the historical user question content and the current user question;
[0056] The following are examples;
[0057] The following is omitted...
[0058] Output the final user question processed by the large model.
[0059] S5. Application scenario recall
[0060] Send the final user question processed by the large model to the application scenario knowledge base for recall processing, requiring the similarity to be less than β, and take the TOP1 entry with the highest similarity.
[0061] If there is no entry with a similarity higher than β, call the large model to rewrite the question again. The core prompt is as follows (some content is omitted):
[0062] <Instruction>Please organize and refine the language according to the known information to generate the splitting result. The format must be task = {'task_i': the splitting result of the question}, which is used to split multiple questions or requirements into single questions and requirements for querying questions. Please strictly output the splitting result of the last known information according to the example format. If it cannot be transformed into NewInstruction, say "The question cannot be answered based on the known information." The answer should be in Chinese. Please do not output the example content.< / Instruction>
[0063] The following is omitted...
[0064] Use the split questions to perform knowledge recall again, take the TOP1 entry with the highest similarity. If the similarity is still less than β, prompt the user that no relevant data can be found.
[0065] S6. Usage extraction
[0066] According to the application scenario, call the large model to extract possible usage directions. The core prompt is as follows (some content is omitted):
[0067] Based on the [known information], please organize and refine your language to reason about the possible uses of the [known information]. You only need to output the reasoning result of the last piece of existing information. Use the expertise of the language model to reason about the uses from the information such as the fields, content, categories, and attributes that the table may contain.
[0068] [Note]: 1. No reasoning process is required; simply output the answer according to the given format. 2. Do not include serial numbers in the output. 3. The example I provided is for formatting reference only; do not output it. 4. Please use Chinese characters for your answer.
[0069] The following is omitted…
[0070] S7, Data Resource Catalog Recall
[0071] The data retrieved from the application scenario is converted into an array, and the data resource catalog knowledge is retrieved in a loop. The retrieved result is a UUID unique identifier.
[0072] S8, Database Query
[0073] Based on the unique identifiers obtained in the previous step, a SQL query statement is generated to query data resource directory information, supporting MySQL, POSTGRE and HANGGUO databases.
[0074] S9, Answering Organization
[0075] Based on the similarity threshold range, the following approximate suggestions are given:
[0076] High similarity: Based on the above analysis, the following data assets have been found to match your criteria:
[0077] Medium similarity: Based on the above analysis, the following data may meet some of the requirements:
[0078] Low similarity: No data could be found for you. We suggest you obtain the data through other means. If you encounter any problems, please contact the administrator.
[0079] One example of how questions and answers are organized:
[0080] [Question]
[0081] Student college entrance examination related data
[0082]
answer
[0083] To analyze student college entrance examination data, you typically need to collect the following data to complete the analysis:
[0084] 1. Student College Entrance Examination Related Data: Based on the [known information] you provided, I infer the following possible uses: understanding the number of college entrance examination applicants, score distribution, admission status, etc., as well as contacting relevant staff.
[0085] Based on the above analysis, the following data may meet some of the students' needs related to the college entrance examination:
[0086] 1. Student Learning Big Data: Raw Score and Assigned Score Records for Subject Selection in High-Level Groups: ID|Grade|Exam ID|Student ID|Student's Raw Score in High-Level Groups|Student's Assigned Score in High-Level Groups|Raw Score Subject Code for High-Level Groups|Assigned Score Subject Code for High-Level Groups
[0087] Directory Name: <a href＝"https: ip path detail.html？cata_id＝xxxx&token＝xxxxx"target＝"blank">Student high score group elective original score and assignment record data
[0088] Table of Contents Description: Student Learning Big Data: Raw Scores and Assigned Scores for Subject Selection in High-Level Groups
[0089] Database Resources: Student Learning Big Data: Raw Scores and Assigned Scores for Subject Selection in High-Level Groups; Department: Municipal Education Bureau
[0090] Contact person for application: Municipal Education Bureau
[0091] 2. Student Learning Big Data Analysis of City Exam Papers: ID | Grade | Exam ID | Exam Paper ID | Subject Code | Number of Students Taking the Exam | Number of Students Absent | Number of Students Missing the Exam | Number of Students Scoring Zero | Highest Raw Score | Lowest Raw Score | Average Raw Score | Standard Deviation | Chart Data | Average Score of Objective Questions | Average Score of Subjective Questions
[0092] Directory Name:<ahref="https: / / IP / path / detail.html?cata_id=xxxx&token=xxxxx"target="blank"> Big Data Analysis of Student Learning and Examination Papers
[0093] Table of Contents Description: Big Data Analysis of Student Learning and City-Level Exam Papers
[0094] Database resources: Big data on student learning, city-wide exam paper analysis data resources
[0095] Department: Municipal Education Bureau
[0096] Contact person for application: Municipal Education Bureau
[0097] ...
[0098] S10, Answer Return
[0099] Provides either a blocking, one-time response or a typewriter-like, word-by-word streaming response from SSE.
[0100] Example 2, based on Example 1, proposes a system for data resource retrieval based on large model technology, including:
[0101] The request receiving module is used to receive user requests, verify the validity of the token, and parse the request parameters.
[0102] The configuration information reading module, connected to the request receiving module, is used to read process configuration information, including but not limited to the URL hyperlink address of the data resource directory system and the data resource database connection information.
[0103] The problem preprocessing module, connected to the configuration information reading module, is used to preprocess the problem, remove illegal characters according to preset rules, and determine whether the length of the processed problem meets the minimum length α. If it does not meet the minimum length α, the user is prompted that there is insufficient available information for the problem.
[0104] The multi-turn dialogue processing module, connected to the question preprocessing module, is used to determine whether the current conversation has historical information. If so, it calls the large model to determine whether the current user question is an extension of a historical user question. If the large model returns a result containing an identifier indicating an extension, it is required to generate a complete question based on the content of the historical user question and the current user question, and output the final user question processed by the large model. When calling the large model to determine whether the current user question is an extension of a historical user question, the prompt input to the large model is: You are a dialogue system that processes a single task. The current task is: determine whether the current user question is an extension of a historical user question. The output content needs to include an analysis and judgment result. If the analysis conclusion is that the current question is an extension of a historical question, the judgment result must contain "yes". If the current question is not an extension of a historical question, the judgment result is "no".
[0105] The application scenario retrieval module, connected to the multi-turn dialogue processing module, sends the end-user question processed by the large model to the application scenario knowledge base for retrieval. When the similarity is less than β, the top-ranked entry with the highest similarity is selected. If no entry with a similarity higher than β is found, the large model is invoked to rewrite the question, and the split question is used for knowledge retrieval again, selecting the top-ranked entry with the highest similarity. If the similarity is still less than β, the user is prompted that no relevant data can be found. When invoking the large model to rewrite the question, the prompt input to the large model is: <Instruction> Please organize refined language to generate the splitting result based on the known information. The format must be task = {'task_i': question splitting result}. This is used to split multiple questions or requirements into a single question or requirement query. Please strictly follow the example format to output the question splitting result of the last known information. If it cannot be converted into a New Instruction, please say "This question cannot be answered based on the known information"; please use Chinese for the answer.
[0106] The purpose extraction module, connected to the application scenario recall module, is used to extract possible purpose directions from the large model based on the application scenario. When the large model is called to extract possible purpose directions, the prompt input to the large model is: Please organize and refine language based on the [known information] to reason about the possible uses of the [known information]. Only the reasoning result of the last existing information needs to be output. With the help of the professional knowledge of the language model, the purpose reasoning is performed from the fields, contents, categories, and attribute information that the table may contain.
[0107] The data resource catalog retrieval module, connected to the usage extraction module, is used to convert the data retrieved from the application scenario into an array, and perform a cyclical retrieval of data resource catalog knowledge. The retrieval result is a unique UUID identifier.
[0108] The database query module, connected to the data resource catalog retrieval module, is used to generate an SQL query statement based on several unique identifiers obtained, and to query data resource catalog information. It supports MySQL, POSTGRE and HANGGUO databases.
[0109] The response organization module, connected to the database query module, provides corresponding prompts based on similarity threshold ranges. Specifically, the prompts based on the similarity threshold range are as follows: High similarity: "Based on the above analysis, the following matching data assets have been found:"; Medium similarity: "Based on the above analysis, the following data may meet some of your needs:"; Low similarity: "No data could be found for you. We suggest you obtain it through other means. If you encounter any problems during use, please contact the administrator."
[0110] The answer return module, connected to the answer organization module, is used to provide either a blocking one-time return of the answer result or a typewriter-like SSE word-by-word stream return.
[0111] Example 3, based on Example 2, proposes a method for compiling and configuring programs that is compatible with domestic IT innovation systems. Its features include support for CPUs such as Intel x86, AMD64, Hygon x86, and Kunpeng ARM architecture; support for conventional Linux systems such as Ubuntu and CentOS, as well as domestic operating systems such as Kylin; and support for conventional databases such as MySQL and PostgreSQL, as well as the domestic HANGGUO database, thus meeting the requirements of domestic IT innovation systems.
[0112] Its core implementation idea is as follows:
[0113] S1 provides a Docker image (D1) for Intel x86, AMD64, and Hygon x86 CPU platforms, compiled under an x86 base image and software environment. It can run on Intel x86 and AMD64 CPU platforms, and under Ubuntu x86, CentOS x86, and Kylin x86 operating systems.
[0114] S2 provides a Docker image (D2) for the Kunpeng ARM platform, using an arm64 base image and basic software, within an ARM compilation environment. It can run on the Kunpeng ARM platform and the Galaxy Kylin arm64 operating system.
[0115] S3. The system itself supports PostgreSQL and Hage databases. When using Hage database, this system provides an external adapter file, which is mounted via Docker volume to override the original database driver file in site-packages / sqlalchemy / dialects / postgresql / base.py, thus enabling support for Hage database through an external plugin.
[0116] S4. It should be noted that although the images provided in this invention have not undergone rigorous testing, the x86 images are theoretically applicable to all types of CPUs in the x86 architecture, and the arm64 images are applicable to all types of CPUs in the arm64 architecture.
[0117] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for data resource retrieval based on large model technology, characterized in that: The core process steps include the following: The request receiving process involves receiving the user request, verifying the token's validity, and parsing the request parameters. The step of reading configuration information involves reading process configuration information, including the URL hyperlink address of the data resource directory system and the data resource database connection information. The problem preprocessing steps involve preprocessing the problem, removing illegal characters, and determining whether the length of the processed problem meets the minimum length α. If not, the user is prompted that there is insufficient available information for the problem. The multi-turn dialogue processing steps determine whether the current session has historical information. If so, the big model determines whether the current user's question is an extension of the historical user's question. If the returned result contains "yes", the big model generates a complete question based on the content of the historical user's question and the current user's question. The application scenario recall step involves sending the end-user question processed by the large model to the application scenario knowledge base for recall processing. If the similarity is less than β, the TOP1 entry with the highest similarity is selected. If there is no entry with a similarity higher than β, the large model is called to rewrite the question and the knowledge is recalled again using the split question. If there is still no similarity, the user is prompted that no relevant data can be found. The usage extraction step involves calling a large model to extract possible usage directions based on the application scenario. The data resource catalog retrieval step involves converting the data retrieved from the application scenario into an array, looping through the data resource catalog knowledge retrieval, and the retrieval result is a unique UUID identifier. The database query process involves generating an SQL query statement based on several unique identifiers obtained, and querying data resource directory information. It supports MySQL, POSTGRE, and HANGGUO databases. Answer the organizational steps and provide corresponding prompts based on the similarity threshold range; The response return step provides either a blocking one-time response or a typewriter-like SSE word-by-word stream response.
2. The data resource retrieval method based on large model technology according to claim 1, characterized in that: In the problem preprocessing step, the rule for removing illegal characters is: ['·°–! "#$%&'()*+,-. / :;<=>?@,。?★、…【】()《》?"”''![\]^_`{|}~]+?.
3. The data resource retrieval method based on large model technology according to claim 2, characterized in that: In the multi-turn dialogue processing steps, when the large model determines whether the current user's question is an extension of a previous user's question, part of the prompt is: You are a dialogue system that processes a single task. The current task is: to determine whether the current user's question is an extension of a previous user's question; the output should include an analysis and judgment result. If the analysis concludes that the current problem is an extension of a historical problem, then the judgment result must include "yes". If the current question is not an extension of a historical question, the result is no; if the large model returns a result containing yes, the core prompt when the large model generates a complete question based on the content of historical user questions and the current user question is: You are a dialogue system that processes a single task. Please generate a complete question based on the content of historical user questions and the current user question.
4. The data resource retrieval method based on large model technology according to claim 3, characterized in that: In the application scenario recall step, the core prompt when rewriting the question by calling the large model is: <Instruction> Based on the known information, please organize refined language to generate the split results. The format must be task = {'task_i': question split result}, used to split multiple questions or requirements into a single question or requirement query. Please strictly follow the example format to output the question split result of the last known information; if it cannot be converted into a New Instruction, please say "Based on the known information, this question cannot be answered"; please use Chinese for the answer.
5. The data resource retrieval method based on large model technology according to claim 4, characterized in that: In the usage extraction step, the core prompt when calling the large model to extract possible usage directions is: Based on the [known information], organize and refine language to reason about the possible uses of the [known information]. Only the reasoning result of the last existing information needs to be output. With the help of the professional knowledge of the language model, use the usage reasoning from the fields, content, categories and attribute information that the table may contain.
6. A system for using the data resource retrieval method based on large model technology according to claim 5, characterized in that: include: The request receiving module is used to receive user requests, verify the validity of the token, and parse the request parameters. The configuration information reading module, connected to the request receiving module, is used to read process configuration information, including but not limited to the URL hyperlink address of the data resource directory system and the data resource database connection information. The problem preprocessing module, connected to the configuration information reading module, is used to preprocess the problem, remove illegal characters according to preset rules, and determine whether the length of the processed problem meets the minimum length α. If it does not meet the minimum length α, the user is prompted that there is insufficient available information for the problem. The multi-turn dialogue processing module, connected to the question preprocessing module, is used to determine whether the current session has historical information. If so, it calls the large model to determine whether the current user question is an extension of the historical user question. If the large model returns a result containing an identifier indicating an extension, it is required to generate a complete question based on the content of the historical user question and the current user question, and output the final user question processed by the large model. The application scenario retrieval module, connected to the multi-turn dialogue processing module, sends the end-user questions processed by the large model to the application scenario knowledge base for retrieval. When the similarity is less than β, the top-1 item with the highest similarity is selected. If there is no item with a similarity higher than β, the large model is called to rewrite the question, and the knowledge is retrieved again using the split question. The top-1 item with the highest similarity is selected. If the similarity is still less than β, the user is prompted that no relevant data can be found. The purpose extraction module is connected to the application scenario recall module. It is used to extract possible purpose directions from the large model based on the application scenario. The data resource catalog retrieval module, connected to the usage extraction module, is used to convert the data retrieved from the application scenario into an array, and perform a cyclical retrieval of data resource catalog knowledge. The retrieval result is a unique UUID identifier. The database query module, connected to the data resource catalog retrieval module, is used to generate an SQL query statement based on several unique identifiers obtained, and to query data resource catalog information. It supports MySQL, POSTGRE and HANGGUO databases. The answer organization module, connected to the database query module, is used to provide corresponding prompts based on the similarity threshold range. The answer return module, connected to the answer organization module, is used to provide either a blocking one-time return of the answer result or a typewriter-like SSE word-by-word stream return.
7. The system according to claim 6, characterized in that: In the multi-turn dialogue processing module, when calling the large model to determine whether the current user's question is an extension of a historical user's question, the prompt input to the large model is: You are a dialogue system that processes a single task. The current task is: to determine whether the current user's question is an extension of a historical user's question; the output should include an analysis and judgment result. If the analysis concludes that the current problem is an extension of a historical problem, then the judgment result must include "yes". If the current problem is not an extension of a historical problem, the result is no.
8. The system according to claim 7, characterized in that: In the application scenario recall module, when rewriting the question by calling the large model, the prompt input to the large model is: <Instruction> Please organize refined language to generate the splitting result based on the known information. The format must be task = {'task_i': question splitting result}, used to split multiple questions or requirements into a single question or requirement query. Please strictly follow the example format to output the question splitting result of the last known information; if it cannot be converted into a New Instruction, please say "This question cannot be answered based on the known information"; please use Chinese for the answer.
9. A system according to claim 8, characterized in that: In the usage extraction module, when calling the large model to extract possible usage directions, the prompt input to the large model is: Please organize and refine language based on the [known information] to reason about the possible uses of the [known information]. Only the reasoning result of the last existing information needs to be output. With the help of the professional knowledge of the language model, the usage reasoning is performed from the fields, contents, categories, and attribute information that the table may contain.
10. A system according to claim 9, characterized in that: In the answer organization module, the specific prompts given based on the similarity threshold range are as follows: High similarity prompts "Based on the above analysis, the following matching data assets have been found:"; Medium similarity prompts "Based on the above analysis, the following data may meet some requirements:"; Low similarity prompts "No data could be found for you. We suggest you obtain it through other means. If you encounter any problems during use, please contact the administrator."