Large language model business query method, product, storage medium and equipment

Abstract

The invention provides a large language model business query method, a product, a storage medium and equipment, and the method comprises the steps: firstly searching for the relationships among various businesses, then converting the relationships into vector business digital tracking links, then searching for the needed business digital tracking links through employing vector business query conditions, and finally, carrying out the business query. And finally, converting the searched service digital tracking chain into the service data required to be queried and the service relationship of the service data required to be queried. Therefore, the purposes of reducing derivation cost and improving data security are achieved while the problems of incapability, inaccuracy and incompleteness in business query are solved.

Description

Technical Field

[0001] The present application relates to the field of information processing technology, and in particular to a large language model business query method, product, storage medium and device. Background Art

[0002] In the enterprise management system, it is often necessary to conduct various business inquiries to understand the handling status of various businesses, and form a global visible digital business tracking chain based on the query results to achieve progress deviation management and compliance audit management of the business. When conducting business inquiries, especially when inquiring business data related to a certain business, it is often impossible to accurately and timely obtain the required business data and business relationship information due to insufficient and incomplete business query conditions.

[0003] In some existing solutions, improvements are made in terms of the efficiency of business queries and the breadth of returned business data to increase the total amount of business data that can be obtained per unit time. However, improving the efficiency and breadth of the returned results of business queries does not solve the problem of inability to find, inaccurate, and incomplete business data due to incomplete and insufficient business query conditions. In some existing solutions, the problem of inability to find, inaccurate, and incomplete business data is solved by automatically supplementing and improving business query conditions and supplementing business feature codes. However, the automated processing mechanism that is highly related to a certain business increases the coupling between the business query system and a single business, thereby reducing the scalability and adaptability of the business query system in other businesses, and ultimately pushing up development costs and operation and maintenance costs. In addition, supplementing business feature codes and promoting and mastering business query methods based on new feature codes require additional costs. In other existing solutions, a customized large language model is used to extract and summarize key business features of business query conditions in voice or text form to improve the transfer learning ability of the business query system in different businesses, thereby achieving the purpose of reducing various costs. However, the big language model is interactively applied based on natural language, and the knowledge mastered by the big language model is expressed in the form of natural language. In addition, the big language model cannot directly distinguish between different users. This leads to the use of methods based on specific induced language to generate business query conditions in some cases to obtain sensitive data that exceeds the original authority. Summary of the invention

[0004] The technical problem to be solved by this application is that in the prior art, the business query conditions are incomplete and insufficient due to various reasons, resulting in the inability to find, inaccurate and incomplete business data, as well as the cost increase and sensitive data security protection issues derived from solving the above problems. A large language model business query method, system, device and equipment are provided.

[0005] In a first aspect, the technical solution of the present application provides a large language model business query method, comprising:

[0006] Acquire business data, where the business data is collected from different business systems;

[0007] Parsing the business relationships of the business data in different business scenarios and generating a business relationship sample pair data set; the business relationship sample pair data set includes a business relationship positive feedback sample set and a business relationship negative feedback sample set;

[0008] The business data in the business relationship sample data set is converted into a business vector by using an initial vector model; the relevant business vectors are searched by a vector comparison method, and the business data corresponding to the relevant business vectors are combined to form a business digital tracking chain;

[0009] According to different application scenarios, the similarity distance of business vectors in business query conditions, business relationship positive feedback sample sets, and business relationship negative feedback sample sets is obtained, the required business relationship sample pair data training set is generated, and the initial vector model is subjected to reinforcement learning and continuous fine-tuning to obtain an iterative large language vector model;

[0010] After obtaining the business query condition and enhancing and supplementing the business query condition through the large language model, the business query condition vector is obtained through the large language vector model;

[0011] The business digital tracking chain related to the business query condition vector is searched through a vector comparison method, and the business data and business relationship corresponding to the related business digital tracking chain are returned as query results.

[0012] In some solutions, the large language model business query method, the business data is obtained, and the business data is collected from different business systems, including:

[0013] Using different types of data interfaces to collect the business data from different types of business systems, the data interfaces include but are not limited to enterprise service bus interfaces, data service interfaces and table import interfaces;

[0014] The required business data is obtained through precise query, a business index is compiled and saved on the corresponding business data through precise update; the precise query means that when performing a query operation on the data, the returned data items must be based on completely matching features and conditions; the precise update means that when performing an update operation on the data, the updated data items must be based on completely matching features and conditions.

[0015] The large language model business query method described in some solutions, the business relationship of the business data in different business scenarios is parsed, and the business relationship sample pair data is generated:

[0016] Generating a business relationship sample pair data set according to the business data and the business relationship by enumeration, including:

[0017] Obtain the business title and business index of the business data through precise query, enumerate all business titles and their business indexes as business query conditions, use the business titles and their business indexes with business relationships as business relationship positive feedback sample sets, use the business titles and their business indexes without business relationships as business relationship negative feedback sample sets, and form a business relationship sample pair data set of business titles in business query conditions;

[0018] Obtain the business paragraphs and business indexes of business data through precise query, enumerate all business paragraphs and their business indexes, use the business paragraphs and their business indexes with business relationships as the business relationship positive feedback sample set, use the business paragraphs and their business indexes without business relationships as the business relationship negative feedback sample set, and form a business relationship sample pair data set of the business paragraphs in the business query condition;

[0019] The business relationship artificial negative feedback sampling and expansion method and the difficult negative business relationship sample mining method are used to expand the size of the business relationship negative feedback sample set.

[0020] In some solutions, the large language model business query method, the method of searching for relevant business vectors by vector comparison, and combining business data corresponding to the relevant business vectors to form a business digital tracking chain, includes:

[0021] According to the business relationship samples, the business index in the data set is precisely queried to obtain the corresponding business features and business rules;

[0022] The business features in the business data are sorted according to the order of different business features in the business rules, and the re-sorted business index is used as the business relationship chain; a set of business relationship chains is accurately updated to save the specified business relationship sample pair data set.

[0023] The large language model business query method described in some solutions uses an initial vector model to convert the business data in the business relationship sample data set into a business vector, including:

[0024] Through precise query, we obtain the data set of business sample pairs, use the big language model to organize the text format and complete the implicit content of the business titles, business paragraphs and business features in the business data, and then use the big language vector model to convert the business data into business vectors, including converting business titles into business title vectors, converting business paragraphs into business paragraph vectors, and converting business features into business feature vectors.

[0025] The large language model business query method described in some solutions obtains the business query conditions, the similarity distance of the business vectors in the business relationship positive feedback sample set and the business relationship negative feedback sample set according to different application scenarios, generates the required business relationship sample pair data training set, performs reinforcement learning and continuous fine-tuning on the initial vector model, and obtains an iterative large language vector model, including:

[0026] Obtain business rules through precise query and enumerate business features in business rules;

[0027] Obtaining a data set of business relationship sample pairs that match business features of business data in business query conditions through precise query;

[0028] According to the positive feedback threshold defined in the business rules, a vector comparison calculation is performed on the business vectors in the business query conditions in the business relationship sample pair data set and the business vectors in the business relationship positive feedback sample set, and those with vector similarity lower than the positive feedback threshold are used as the business relationship sample pair data training set; according to the negative feedback threshold defined in the business rules, a vector comparison calculation is performed on the business vectors in the business query conditions in the business relationship sample pair data set and the business vectors in the business relationship negative feedback sample set, and those with vector similarity higher than the negative feedback threshold are used as the business relationship sample pair data training set.

[0029] In some embodiments, the large language model business query method, after obtaining the business query condition and enhancing and supplementing the business query condition through the large language model, obtains the business query condition vector through the large language vector model, including:

[0030] Obtain business query conditions and data encryption certificates;

[0031] Submit the business query conditions and the implied account roles to the large language model;

[0032] The large language model reorganizes the business query condition text to obtain a new business query condition text, which includes both the business query condition and the account role.

[0033] Generate business query condition vectors through the large language vector model.

[0034] The method of searching for a business digital tracking chain related to the business query condition vector by using a vector comparison method, and returning business data and business relationships corresponding to the related business digital tracking chain as a query result, includes:

[0035] Obtain the required business digital tracking chain through similarity query; obtain relevant business data through precise query based on the business digital tracking chain, and then encrypt the relevant business data and business relationships;

[0036] Obtain the data decryption certificate, decrypt the encrypted business data and business relationship according to the data decryption certificate, and then return it.

[0037] In a second aspect, the technical solution of the present application provides a computer-readable storage medium, in which program information is stored. After the computer reads the program information, the computer executes the steps of the large language model business query method described in any one of the first aspects.

[0038] In a third aspect, the technical solution of the present application provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the large language model business query method described in any one of the first aspects.

[0039] In a fourth aspect, the technical solution of the present application provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the large language model business query method described in any one of the first aspects.

[0040] Compared with the prior art, the above technical solution provided by this application has the following technical effects:

[0041] The large language model business query method, product, storage medium and device provided by the present application first find the relationship between various businesses, then convert these relationships into a digital business tracking chain after vectorization, then use the business query conditions after vectorization to find the required digital business tracking chain, and finally convert the found digital business tracking chain into the required business data and its business relationship. Thus, the problem of not being able to find, not being able to find accurately and not being able to find completely in business query is solved while reducing the derivation cost and improving data security. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a flowchart of a large language model business query method according to an embodiment of the present application;

[0043] Figure 2 This is a schematic diagram of the structure of the logical relationship framework of the large language business query method according to an embodiment of the present application;

[0044] Figure 3A schematic diagram of data transmission relationship for implementing a service query method according to an embodiment of the present application;

[0045] Figure 4 This is a schematic diagram of the structure of a data set of business relationship samples according to an embodiment of the present application;

[0046] Figure 5 A schematic diagram of the hardware connection relationship of an electronic device for implementing a service query method according to an embodiment of the present application. DETAILED DESCRIPTION

[0047] The specific implementation of the present application is further described below with reference to the accompanying drawings.

[0048] It is easy to understand that according to the technical solution of the present application, without changing the essential spirit of the present application, a variety of structural modes and implementation modes that can be replaced by those skilled in the art can be replaced with each other. Therefore, the following specific implementation modes and drawings are only exemplary descriptions of the technical solution of the present application, and should not be regarded as the entirety of the present application or as a limitation or restriction on the technical solution of the application.

[0049] The directional terms such as up, down, left, right, front, back, front, back, top, bottom, etc. mentioned or may be mentioned in this specification are defined relative to the structures shown in the drawings. They are relative concepts and may change accordingly according to different positions and different usage conditions. Therefore, these or other directional terms should not be interpreted as restrictive terms.

[0050] This embodiment provides a large language model service query method, such as Figure 1 As shown, including:

[0051] S100: Acquire business data, where the business data is collected from different business systems.

[0052] In specific implementation, various types of data interfaces (including but not limited to enterprise service bus, data service interface API, table import, etc.) are used to collect business data from various business systems.

[0053] S200: parsing the business relationship of the business data in different business scenarios, and generating a business relationship sample pair data set; the business relationship sample pair data set includes a business relationship positive feedback sample set and a business relationship negative feedback sample set.

[0054] In the specific implementation, the business relationship of business data in different business scenarios is analyzed, the business relationship is extracted from the business data, and the business relationship sample pair data set is enumerated. Furthermore, the business relationship artificial negative feedback sampling amplification method and the difficult negative business relationship sample mining method are used to expand the scale of the business relationship negative feedback sample set.

[0055] S300: using an initial vector model to convert the business data in the business relationship sample pair data set into a business vector; searching for relevant business vectors by a vector comparison method, and combining the business data corresponding to the relevant business vectors to form a business digital tracking chain.

[0056] In the specific implementation, the large language vector model is used to convert the business data in the business relationship sample pair data set into business vectors. At the same time, the business query conditions that do not meet the positive feedback threshold and negative feedback threshold conditions and the vector similarity in the business relationship positive feedback sample set and the business relationship negative feedback sample set are used as the business relationship sample pair data training set. The vector comparison method is used to find the relevant business vectors as business relationships, and the business data is combined to form a business digital tracking chain.

[0057] S400: According to different application scenarios, obtain the similarity distance of business vectors in the business query conditions, business relationship positive feedback sample set and business relationship negative feedback sample set, generate the required business relationship sample pair data training set, perform reinforcement learning and continuous fine-tuning on the initial vector model, and obtain an iterative large language vector model.

[0058] In the specific implementation, according to different application scenarios, the similarity distance of business vectors in business query conditions, business relationship positive feedback sample sets and business relationship negative feedback sample sets is calculated, and the required business relationship sample pair data training set is selected to generate, and an iterative large language vector model is obtained through reinforcement learning and continuous fine-tuning.

[0059] S500: After obtaining a business query condition and enhancing and supplementing the business query condition through a large language model, a business query condition vector is obtained through the large language vector model.

[0060] In specific implementation, after the business query conditions are enhanced and supplemented by using the large language model, the business query condition vector is obtained through the large language vector model.

[0061] S600: searching for a business digital tracking chain related to the business query condition vector by a vector comparison method, and returning business data and business relationships corresponding to the related business digital tracking chain as a query result.

[0062] In the specific implementation, the relevant business digital tracking chain is found through the vector comparison method, and the required business data and business relationships are returned according to the data access certificate.

[0063] The above scheme of the present application first uses a customized and trained large language model to supplement the business data and business query conditions respectively, so that they contain sufficient business classification, business scenarios and business feature descriptions, and ensure the consistency of the grammatical format. Then, use a customized and trained large language vector model to convert these supplemented business data or business query conditions into high-dimensional vectors respectively. Finally, the required business vector is found by comparing and calculating the vector similarity between the business query condition vector and the business vector, thereby obtaining the required business data. This solution achieves the purpose of reducing derivative costs and improving data security while solving the problems of inability to find, inaccurate and incomplete search in business queries.

[0064] Specifically, combined Figure 2 and Figure 3 As shown in the structural diagram, the above method can be run in a large language model business query system, and the system can be loaded in a computer device 9, which includes a business management system 7 and a large language model query device 8. The query device includes a data acquisition module 1, a relationship analysis module 2, a model training module 3, a stock data update module 4 and a business query module 5. The business management system 7 includes: a business tracking chain management module 701, a business statistics display module 702, a business compliance inspection module 703, a business early warning module 704, a business alarm module 705, a business checklist dispatching module 706, and a business management system database 707. Business database 101, business data collection interface 102, business annotation service 103, business specification configuration service 104, business index compiler 105, business relationship database 201, manual annotation relationship analyzer 202, business title relationship analyzer 203, business paragraph relationship analyzer 204, business relationship chain analyzer 205, large language model library 301, business specification database 302, large language model fine-tuning trainer 303, enhanced training data set generator 304, business chain vector library 401, business tracking chain generator 402, stock business updater 403, business relationship extractor 404, unchained business relationship analyzer 405, business query service 501, large language model service 502, large language vector model service 503, business query log library 504, business query audit gateway 505, business data encryption service 506 and business management system 7. The above components, databases and services can be implemented in the form of program modules.

[0065] In some schemes, in step S100: different types of data interfaces are used to collect the business data from different types of business systems, and the data interfaces include but are not limited to enterprise service bus interfaces, data service interfaces and table import interfaces; the required business data is obtained through precise query, and a business index is compiled and saved on the corresponding business data through precise update; the precise query means that when querying the data, the returned data items must be based on completely matching features and conditions; the precise update means that when updating the data, the updated data items must be based on completely matching features and conditions. Specifically, the business data collection interface 102 collects business data from the business management system 7, and then uses the precise update network to save the business data to the business database 101. The business management system 7 submits the business relationship manual data annotation information to the business annotation service 103, and saves it to the business database 101 through the precise update network. In the figure:

[0066] Precise update network: refers to when updating the data in the database, the updated data items must be based on completely matching features and conditions. Precise query network: refers to when querying the data in the database, the returned data items must be based on completely matching features and conditions. Similar query network: refers to when querying the data in the database, the returned data items are based on the value range calculated by mathematical formulas or algorithms. Its basic similarity query network can be used but not limited to Portions' PostgreSQL pgvector vector database product. Model update network: refers to the network that saves the model features such as the name, fine-tuning dataset, pre-order model, and basic model of the large language model (including the large language vector model) to the large language model library. At the same time, the large language model update message is released to the model scheduling network. Model scheduling network: refers to subscribing to the large language model update message according to the model features such as the name, fine-tuning dataset, and source model of the large language model (including the large language vector model). When receiving a new large language model update message, it provides the large language model download service to the downstream application and automatically completes the quantitative adjustment and format conversion of the large language model. Its basic framework can adopt but is not limited to the Ollama large model scheduling framework. Large language model: It is a large language model that is fine-tuned and trained to organize the text format of business descriptions. Its basic model can adopt but is not limited to Meta's LLama3 70B large language model. Large language vector model: It is a large language vector model that is used to convert the title, paragraph and feature description text of the business into vectors in a high-dimensional space according to business relationships. Its basic model can adopt but is not limited to MokaAI's M3E Large large language vector model.

[0067] In some scenarios, business relationship samples are paired with data sets such as Figure 4As shown, the business relationship of the business data in different business scenarios is parsed, and a business relationship sample pair data set is generated. The business relationship sample pair data set is generated according to the business data and the business relationship by enumeration, including: obtaining the business title and business index of the business data by precise query, enumerating all business titles and their business indexes as business query conditions, taking the business title and its business index with business relationship as the business relationship positive feedback sample set, taking the business title and its business index without business relationship as the business relationship negative feedback sample set, and forming the business relationship sample pair data set of the business title in the business query condition; obtaining the business paragraph and business index of the business data by precise query, enumerating all business paragraphs and their business indexes, taking the business paragraph and its business index with business relationship as the business relationship positive feedback sample set, taking the business paragraph and its business index without business relationship as the business relationship negative feedback sample set, and forming the business relationship sample pair data set of the business paragraph in the business query condition; adopting the business relationship artificial negative feedback sampling amplification method and the difficult negative business relationship example mining method to expand the scale of the business relationship negative feedback sample set.

[0068] Specifically, combined Figure 3 and Figure 4 As shown, the required business rules are obtained from the business specification database 302 through the precise query network and displayed in a visual manner on the business specification configuration service 104, and the edited and modified new business rules are saved to the business specification database 302 through the precise update network. Further, the business index compiler 105 obtains the required business data through the precise query network, compiles and generates a business index for it, and saves it on the corresponding business data in the business database 101 through the precise update network. The business title relationship analyzer 203 obtains the business title 1101 and business index 1006 of the business data 1005 from the business database 101 through the precise query network, enumerates all the business titles 1101 and their business indexes 1006 as the business query conditions 1002, takes the business titles 1101 and their business indexes 1006 with business relationships as the business relationship positive feedback sample set 1003, takes the business titles 1101 and their business indexes 1006 without business relationships as the business relationship negative feedback sample set 1004, and constitutes the business relationship sample pair data set 1001 of the business title 1101 in the business query conditions 1002, which is saved in the business relationship database 201 through the precise update network.

[0069] Further, the business paragraph relationship analyzer 204 obtains the business paragraph 1102 and business index 1006 of the business data 1005 from the business database 101 through the precise query network, enumerates all business paragraphs 1102 and their business indexes 1006, takes the business paragraph 1009 with business relationship and its business index 1006 as the business relationship positive feedback sample set 1003, takes the business paragraph 1102 without business relationship and its business index 1006 as the business relationship negative feedback sample set 1004, constitutes the business relationship sample pair data set 1001 of the business paragraph 1009 in the business query condition 1002, and saves it in the business relationship database 201 through the precise update network. The manual annotation relationship analyzer 202 adopts the business relationship manual negative feedback sampling and amplification method and a difficult negative business relationship sample mining method to expand the scale of the business relationship negative feedback sample set. First, the business relationship manual negative feedback sampling and amplification method is adopted to obtain the business data 1005 and the business index 1006 with the business relationship manual data annotation business characteristics from the business database 101 through the precise query network. Alternatively, a difficult negative business relationship sample mining method is used, and business data 1005 and business index 1006 that meet these keyword conditions are obtained from the business database 101 through the precise query network according to the keywords in the difficult negative sample manual annotation set. Secondly, the relevant business relationship sample pair data set is found in the precise query network through the business index, and it is determined whether this business index is consistent with the expected business relationship feedback sample set. If it is inconsistent, the business data is migrated and kept unique to the expected business relationship feedback sample set, so as to form the business relationship sample pair data set 1001. Otherwise, no processing is performed. Finally, it is saved in the business relationship database 201 through the precise update network.

[0070] Furthermore, the method of searching for relevant business vectors by vector comparison and combining business data corresponding to the relevant business vectors to form a business digital tracking chain includes: obtaining corresponding business features and business rules by accurately querying the business index in the data set according to the business relationship sample; sorting the business features in the business data according to the sequence of different business features in the business rules, and using the re-sorted business index as the business relationship chain; and saving a specified business relationship sample data set by accurately updating a set of business relationship chains.

[0071] Combination Figure 3 and Figure 4As shown, the business relationship chain analyzer 205 first obtains the business relationship sample pair data set 1001 from the business relationship database 201 through the precise query network. Then, according to the business index 1006 in the business relationship sample pair data set 1001, the corresponding business feature 1103 is obtained from the business database 101 through the precise query network. Secondly, the business rules are obtained from the business specification database 302 through the precise query network. Thirdly, the business features 1010 in the business data 1005 are sorted according to the sequence of different business features 1103 in the business rules, and then the re-sorted set of business indexes 1006, i.e., the business relationship chain 1007, is saved to the specified business relationship sample pair data set 1001 in the business relationship database 201 through the precise update network.

[0072] Preferably, in the above scheme, an initial vector model is used to convert the business data in the business relationship sample pair data set into a business vector, including: obtaining a business sample pair data set through precise query, using a large language model to organize the text format and complete the implicit content of the business title, business paragraph and business features in the business data, and then converting the business data into a business vector through a large language vector model, including converting the business title into a business title vector, converting the business paragraph into a business paragraph vector and converting the business feature into a business feature vector.

[0073] Combination Figure 3 and Figure 4 As shown, the enhanced training data generator 304 first obtains the latest large language model and large language vector model from the large language model library 301 through the model scheduling network and loads and runs. Next, the business sample pair data set 1001 is obtained from the business relationship database 201 through the precise query network. Secondly, the business title 1101, business paragraph 1102 and business feature 1103 in the business data 1005 are sorted into text format and implicit content is completed using the large language model, and then the business data 1005 is converted into a business vector 1008 through the large language vector model, including converting the business title 1101 into a business title vector 1104, converting the business paragraph 1102 into a business paragraph vector 1105, and converting the business feature 1103 into a business feature vector 1106, which are saved in the business relationship database 201 through the precise update network.

[0074] In the above scheme, according to different application scenarios, the similarity distance of business vectors in business query conditions, business relationship positive feedback sample sets and business relationship negative feedback sample sets is obtained, the required business relationship sample pair data training set is generated, the initial vector model is subjected to reinforcement learning and continuous fine-tuning, and an iterative large language vector model is obtained, including: obtaining business rules through precise query and enumerating business features in business rules; obtaining a business relationship sample pair data set that matches the business features of business data in business query conditions through precise query; performing vector comparison calculation on the business vectors in the business query conditions in the business relationship sample pair data set and the business vectors in the business relationship positive feedback sample set according to the positive feedback threshold defined in the business rules, and using the vectors with a similarity lower than the positive feedback threshold as the business relationship sample pair data training set; performing vector comparison calculation on the business vectors in the business query conditions in the business relationship sample pair data set and the business vectors in the business relationship negative feedback sample set according to the negative feedback threshold defined in the business rules, and using the vectors with a similarity higher than the negative feedback threshold as the business relationship sample pair data training set.

[0075] Combination Figure 3 and Figure 4As shown, the enhanced training data generator 304 first obtains business rules from the business specification database 302 through the precise query network. Next, the business features 1103 in the business rules are enumerated, and the business relationship sample pair data set 1001 matching the business features 1103 of the business data 1005 in the business query condition 1002 is obtained from the business relationship database 201 through the precise query network. Secondly, according to the positive feedback threshold defined in the business rules, the business vector 1008 in the business query condition 1002 in the business relationship sample pair data set 1001 and the business vector 1008 in the business relationship positive feedback sample set 1003 are compared and calculated, and the vectors with a vector similarity lower than the positive feedback threshold are used as the business relationship sample pair data training set. At the same time, according to the negative feedback threshold defined in the business rules, the business vector 1008 in the business query condition 1002 in the business relationship sample pair data set 1001 and the business vector 1008 in the business relationship negative feedback sample set 1004 are compared and calculated, and the vectors with a vector similarity higher than the negative feedback threshold are used as the business relationship sample pair data training set. Finally, the business relationship sample pair data training set is sent to the large language model fine-tuning trainer 303. The business tracking chain generator 403 first obtains the business relationship sample pair data set 1001 from the business relationship database 201 through the precise query network, finds similar business vectors as business relationships through the vector comparison method, and combines them to form a business digital tracking chain. Then, the business digital tracking chain is saved in the business chain vector library through the precise update network. The further unchained business data analyzer 404 first obtains the newly added business data from the business database through the precise query network, and then confirms whether the business data has been saved in the business relationship sample pair data set 1001 of the business relationship database 201 through the precise query network. If it is not saved in the data set, the business tracking chain generator 403 is called to associate the business data with the relevant business digital tracking chain, and is updated to the business chain vector library 401 through the precise update network.

[0076] In the above scheme, after obtaining the business query conditions and enhancing and supplementing the business query conditions through the large language model, the business query condition vector is obtained through the large language vector model, including: obtaining the business query conditions and data encryption certificates; submitting the business query conditions and the implied account roles to the large language model; the large language model reorganizes the business query condition text to obtain a new business query condition text, in which the new business query condition text contains both the business query conditions and the account roles; and generating the business query condition vector through the large language vector model. Combined with Figure 3 and Figure 4, which includes: first, the business management system 7 submits the business query conditions and the data encryption certificate to the business query service 501. Secondly, the business query service 501 submits the business query conditions and the implied account roles to the large language model service 502. Then the large language model service 502 reorganizes the business query condition text according to the preset prompt word template, so that the new business query condition text contains both the business query conditions and the account roles. Finally, the business vector is generated by the large language vector model service 503, and the required business digital tracking chain is obtained from the business chain vector library 401 through the similarity query network.

[0077] Preferably, the method of searching for the business digital tracking chain related to the business query condition vector by vector comparison, and returning the business data and business relationship corresponding to the relevant business digital tracking chain as the query result, includes: obtaining the required business digital tracking chain through similarity query; obtaining relevant business data through precise query according to the business digital tracking chain, and then encrypting the relevant business data and business relationship; obtaining the data decryption certificate, decrypting the encrypted business data and business relationship according to the data decryption certificate and returning it. Figure 3 and Figure 4 , which includes: first, the business query service 501 sends the data decryption certificate to the business data decryption service 506. At the same time, the business vector generated by the large language vector model service 503 is pushed to the business query audit gateway 505 through the business digital tracking chain obtained by the similarity query network. Then, the business query audit gateway 505 obtains relevant business data from the business database 101 through the precise query network according to the business digital tracking chain, and then encrypts the business data and business relationship and sends them to the business data encryption service 506. Finally, the business data encryption service 506 decrypts the encrypted business data and business relationship according to the received data decryption certificate and sends them to the business management system 7.

[0078] The above-mentioned scheme provided by the present application adopts a business relationship fine-tuning training algorithm in the fine-tuning of the large language vector model, and generates a business relationship positive and negative feedback sample set based on business data and its business relationships to fine-tune the vector model. The fine-tuning training algorithm constructs a business relationship sample pair data set 1001 based on the business relationship positive feedback sample set 1003 and the business relationship negative feedback sample set 1004 in the business query condition 1002 that has been corrected and supplemented by the large language model, thereby assisting the large language vector model to master the semantic matching tasks of various business relationships in actual business scenarios. At the same time, the scale of the business relationship negative feedback sample set 1004 is further expanded with the help of the business relationship artificial negative feedback sampling amplification method and the difficult negative business relationship sample mining method. Through the aforementioned steps, a better ability to conduct business queries through semantic vectors containing business relationships is achieved.

[0079] In order to solve the problem of rising development costs, operation and maintenance costs, and training costs derived from new business query methods, we first use a customized and trained big language vector model to convert business data into business vectors. Secondly, we use business relationships to classify different businesses, generate business vectors for different business scenarios and business characteristics, and generate business digital tracking chains. Secondly, for unlinked business relationships, we use the newly iterated big language vector model to generate business vectors, and integrate them into the relevant business digital tracking chains according to the business relationships, so that the business digital tracking chains contain business vectors generated by big language vector models of different iterative versions. In this way, after using big language vector models of different iterative versions to convert business query conditions into vectors, we can find the business digital tracking chain for the query we want. Then we can obtain the business data we want to query, and achieve the goal of not increasing development costs, operation and maintenance costs, and training costs.

[0080] In order to solve the problem of sensitive data leakage caused by large language models, firstly, the business digital tracking chain only contains business vectors, business data index certificates and business relationship indexes, and does not store any sensitive natural language information. Secondly, when making a business query request, the business management system will submit a data encryption certificate to encrypt the business data, thereby ensuring the confidentiality of the business data during transmission.

[0081] In actual applications, this solution is highly integrated with the business management system, not only providing business query functions that include business features, business classifications, and semantic similarity of business scenarios, but also using business-related semantics, contexts, and scenarios to associate multiple business events (meetings, resolutions), affairs (procurement projects), and files (official documents), so that managers can grasp the entire process of business development from a global, visual perspective, and effectively achieve progress deviation management and compliance audit management of the business. The system runs well as a whole. In actual operation, this solution can run in a variety of computer equipment hardware environments, including cloud environments, central processing unit environments, hybrid processor environments, and distributed processor environments, and can achieve stable and reliable operation performance.

[0082] In some embodiments of the present application, a computer-readable storage medium is provided, in which program information is stored. After a computer reads the program information, the steps of the large language model business query method described in any one of the above embodiments are executed.

[0083] In some embodiments of the present application, a computer program product is provided, including a computer program / instruction, which, when executed by a processor, implements the steps of the large language model business query method described in any of the above embodiments.

[0084] In some embodiments of the present application, an electronic device is provided, such as Figure 5As shown, the electronic device includes at least one processor 51 and at least one memory 52, at least one of the memories 52 stores program information, and at least one of the processors 51 reads the program information and executes the large language model business query method described in any of the above method embodiments. The device may also include: an input device 53 and an output device 54. The processor 51, the memory 52, the input device 53 and the output device 54 can be connected in communication. The memory 52, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer executable programs and modules. The processor 51 executes various functional applications and data processing by running the non-volatile software programs, instructions and modules stored in the memory 52, that is, the large language model business query method provided by any of the above schemes is implemented. The memory 52 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required by at least one function; the data storage area may store data created according to the use of the large language model business query method, etc. In addition, the memory 52 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 52 may optionally include a memory remotely arranged relative to the processor 51, and these remote memories may be connected to a device for executing the large language model business query method through a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network and a combination thereof. The input device 53 may receive input user clicks, and generate signal inputs related to user settings and function controls of the large language model business query method. The output device 54 may include a display device such as a display screen. When the one or more modules are stored in the memory 52, when they are run by the one or more processors 51, the large language model business query method in any of the above-mentioned method embodiments is executed.

[0085] As needed, the above technical solutions can be combined to achieve the best technical effect.

[0086] The above are only the principles and preferred embodiments of the present application. It should be noted that, for ordinary technicians in this field, on the basis of the principles of the present application, several other modifications can be made, which should also be considered as the protection scope of the present application.

Claims

1. A large language model business query method, characterized in that: include: Acquire business data, where the business data is collected from different business systems; Parsing the business relationships of the business data in different business scenarios and generating a business relationship sample pair data set; the business relationship sample pair data set includes a business relationship positive feedback sample set and a business relationship negative feedback sample set; The business data in the business relationship sample data set is converted into a business vector by using an initial vector model; the relevant business vectors are searched by a vector comparison method, and the business data corresponding to the relevant business vectors are combined to form a business digital tracking chain; According to different application scenarios, the similarity distance of business vectors in business query conditions, business relationship positive feedback sample sets, and business relationship negative feedback sample sets is obtained, the required business relationship sample pair data training set is generated, and the initial vector model is subjected to reinforcement learning and continuous fine-tuning to obtain an iterative large language vector model; After obtaining the business query condition and enhancing and supplementing the business query condition through the large language model, the business query condition vector is obtained through the large language vector model; The business digital tracking chain related to the business query condition vector is searched through a vector comparison method, and the business data and business relationship corresponding to the related business digital tracking chain are returned as query results.

2. The large language model business query method according to claim 1, characterized in that: The acquiring of business data, wherein the business data is collected from different business systems, includes: Using different types of data interfaces to collect the business data from different types of business systems, the data interfaces include but are not limited to enterprise service bus interfaces, data service interfaces and table import interfaces; The required business data is obtained through precise query, a business index is compiled and saved on the corresponding business data through precise update; the precise query means that when performing a query operation on the data, the returned data items must be based on completely matching features and conditions; the precise update means that when performing an update operation on the data, the updated data items must be based on completely matching features and conditions.

3. The large language model business query method according to claim 2, characterized in that: The business relationship of the business data in different business scenarios is analyzed, and the business relationship sample pair data is generated: Generating a business relationship sample pair data set according to the business data and the business relationship by enumeration, including: Obtain the business title and business index of the business data through precise query, enumerate all business titles and their business indexes as business query conditions, use the business titles and their business indexes with business relationships as business relationship positive feedback sample sets, use the business titles and their business indexes without business relationships as business relationship negative feedback sample sets, and form a business relationship sample pair data set of business titles in business query conditions; Obtain the business paragraphs and business indexes of business data through precise query, enumerate all business paragraphs and their business indexes, use the business paragraphs and their business indexes with business relationships as the business relationship positive feedback sample set, use the business paragraphs and their business indexes without business relationships as the business relationship negative feedback sample set, and form a business relationship sample pair data set of the business paragraphs in the business query condition; The business relationship artificial negative feedback sampling and expansion method and the difficult negative business relationship sample mining method are used to expand the size of the business relationship negative feedback sample set.

4. The large language model business query method according to claim 3, characterized in that: The method of searching for relevant business vectors by vector comparison and combining business data corresponding to the relevant business vectors to form a business digital tracking chain includes: According to the business relationship samples, the business index in the data set is precisely queried to obtain the corresponding business features and business rules; The business features in the business data are sorted according to the order of different business features in the business rules, and the re-sorted business index is used as the business relationship chain; a set of business relationship chains is accurately updated to save the specified business relationship sample pair data set.

5. The large language model business query method according to claim 4, characterized in that: The business data in the business relationship sample pair data set is converted into a business vector using an initial vector model, including: Through precise query, we obtain the data set of business sample pairs, use the big language model to organize the text format and complete the implicit content of the business titles, business paragraphs and business features in the business data, and then use the big language vector model to convert the business data into business vectors, including converting business titles into business title vectors, converting business paragraphs into business paragraph vectors, and converting business features into business feature vectors.

6. The large language model business query method according to claim 5, characterized in that: According to different application scenarios, the similarity distance of business vectors in business query conditions, business relationship positive feedback sample sets, and business relationship negative feedback sample sets is obtained, the required business relationship sample pair data training set is generated, and the initial vector model is subjected to reinforcement learning and continuous fine-tuning to obtain an iterative large language vector model, including: Obtain business rules through precise query and enumerate business features in business rules; Obtaining a data set of business relationship sample pairs that match business features of business data in business query conditions through precise query; According to the positive feedback threshold defined in the business rules, a vector comparison calculation is performed on the business vectors in the business query conditions in the business relationship sample pair data set and the business vectors in the business relationship positive feedback sample set, and those with vector similarity lower than the positive feedback threshold are used as the business relationship sample pair data training set; according to the negative feedback threshold defined in the business rules, a vector comparison calculation is performed on the business vectors in the business query conditions in the business relationship sample pair data set and the business vectors in the business relationship negative feedback sample set, and those with vector similarity higher than the negative feedback threshold are used as the business relationship sample pair data training set.

7. The large language model business query method according to any one of claims 4 to 6, characterized in that: After acquiring the business query condition and enhancing and supplementing the business query condition through the large language model, obtaining the business query condition vector through the large language vector model includes: Obtain business query conditions and data encryption certificates; Submit the business query conditions and the implied account roles to the large language model; The large language model reorganizes the business query condition text to obtain a new business query condition text, which includes both the business query condition and the account role. Generate business query condition vectors through a large language vector model; The method of searching for a business digital tracking chain related to the business query condition vector by using a vector comparison method, and returning business data and business relationships corresponding to the related business digital tracking chain as a query result, includes: Obtain the required business digital tracking chain through similarity query; obtain relevant business data through precise query based on the business digital tracking chain, and then encrypt the relevant business data and business relationships; Obtain the data decryption certificate, decrypt the encrypted business data and business relationship according to the data decryption certificate, and then return it.

8. A computer-readable storage medium, characterized in that: The storage medium stores program information, and after the computer reads the program information, the computer executes the steps of the large language model business query method according to any one of claims 1 to 7.

9. A computer program product, characterized in that It includes a computer program / instruction, characterized in that when the computer program / instruction is executed by a processor, the steps of the large language model business query method described in any one of claims 1-7 are implemented.

10. An electronic device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the large language model business query method according to any one of claims 1 to 7.

Images