Prompt Generation System

Abstract

To generate a prompt for a text generation model.SOLUTION: A system for generating a prompt for a text generation model for generating an answer sentence includes one or more computing devices and one or more storage devices. The one or more computing devices acquire a query sentence from a user, generate the prompt for the text generation model based on the query sentence, transmit the generated prompt to the text generation model via an open network, and perform filtering processing to exclude confidential information specified in the generation of the prompt.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to generating prompts for text generation models. [Background technology]

[0002] The development and use of generative AI (Artificial Intelligence) is becoming more and more popular. However, since products using generative AI are generated based on search results from past data accumulation, complex prompts must be generated to respond to the user's real-time environmental situation. This makes it difficult to quickly obtain the desired results.

[0003] Due to the nature of generative AI, when searching for manuals, etc., generative AI cannot utilize real-time environmental conditions, such as the questioner's location, temperature, and weather.

[0004] General search engines add additional information to search results based on user account and location information to perform searches that match user preferences. However, it is difficult to incorporate this information into the general LLMs (Large Language Models) used in generative AI.

[0005] For example, when a malfunction occurs in the field, the user searches the manual and prepares a solution. However, because environmental information cannot be added to a general LLM, the general LLM generates proposals based only on the input information. As a result, the general LLM may not be able to provide timely proposals. For example, when a car's air conditioner malfunctions, the priority of the items to be addressed may differ depending on the ambient temperature. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Patent No. 7416508 Summary of the Invention [Problem to be solved by the invention]

[0007] For example, a local LLM (Local LLM) can be constructed as a system that collects document information within a company. Private LLMs are typically more robust than open LLMs when it comes to information specific to a company. However, due to a lack of parameters as a language model, they tend to have difficulty obtaining appropriate answers. The parameters of a private LLM are generally around 75B, significantly lower than the 355B of GPT-3.5, and less than one-thousandth of GPT-4.

[0008] On the other hand, when using Open LLM, there is a risk of internal company information being leaked during fine tuning. Also, since Open LLM generates answers from a large amount of information, it can be difficult to generate the desired information. [Means for solving the problem]

[0009] One aspect of the present invention is a system for generating a prompt for a text generation model that generates an answer sentence, the system including one or more computing devices and one or more storage devices, wherein the one or more computing devices acquire a query sentence from a user, generate a prompt for the text generation model based on the query sentence, transmit the generated prompt to the text generation model via an open network, and perform a filtering process to exclude specified confidential information in generating the prompt. [Effects of the Invention]

[0010] According to one aspect of the present invention, appropriate prompts can be generated for a language model that returns an answer. [Brief explanation of the drawings]

[0011] [Figure 1] 1 shows an overview of a dialogue system according to an embodiment of the present specification. [Figure 2] 1 shows an example of the hardware configuration of a prompt generation system. [Figure 3] This indicates the information that the prompt generation system acquires from an external source in order to generate an answer sentence. [Figure 4] 1 is a block diagram showing a schematic functional configuration of a prompt generation system for generating a prompt for an open LLM in response to a query from a user. [Figure 5] 10 shows an example of text in which information has been supplemented in response to a query from a user terminal, and an example of text in which filtering for anonymity has been performed. [Figure 6] FIG. 1 is a block diagram illustrating the processing of a prompt generation system for an open LLM. [Figure 7] This shows the response received from Open LLM and the text in which the confidential information has been restored from that response. [Figure 8] The answer received from Open LLM and the learning data generated based on that answer are shown below. [Figure 9] 1 is a flowchart illustrating the processing of a prompt generation system. DETAILED DESCRIPTION OF THE INVENTION

[0012] In the following, when necessary for convenience, the description will be divided into multiple sections or examples, but unless otherwise specified, they are not unrelated to each other, and one is related to the other as a partial or complete modification, detail, supplementary explanation, etc. Furthermore, in the following, when the number of elements, etc. (including the number, numerical value, amount, range, etc.) is mentioned, it is not limited to that specific number, and may be more or less than the specific number, unless otherwise specified or when it is clearly limited in principle to a specific number, etc.

[0013] The processor realizes a predetermined function by executing a program stored in a primary storage device. The primary storage device stores the program executed by the processor and data necessary for the execution of the program. The program includes an operating system (OS) (not shown) as well as other programs. The processor may include multiple chips and multiple packages.

[0014] A program is executed by a processor to perform a predetermined process using a storage device and a communication port (communication device). Therefore, in this and other embodiments, a description that uses a program as the subject may also use a processor as the subject. Alternatively, the process executed by a program is a process performed by the computer and computer system on which the program runs.

[0015] A processor operates as a functional unit (means) that realizes a predetermined function by operating according to a program. For example, a processor functions as a unit (means) by operating according to a program, and as a unit (means) by operating according to a program. The same applies to other programs. Furthermore, a processor also operates as a functional unit (means) that realizes each of the multiple processes executed by each program. A computer and a computer system are devices and systems that include these functional units (means).

[0016] 1 shows an overview of a dialogue system according to an embodiment of the present specification. The dialogue system according to an embodiment of the present specification generates a response to a query from a user and presents the response to the user. The dialogue system includes a prompt generation system 10 and an external open Large Language Model (LLM) 30.

[0017] The prompt generation system 10 generates a prompt in response to an inquiry input from a user terminal 40 and transmits the prompt to the open LLM 30. The prompt generation system 10 performs necessary processing on the answer from the open LLM 30 and then transmits the processed answer to the user terminal 40.

[0018] In the example shown in Figure 1, a user sends a query from a user terminal 40 to the prompt generation system 10 about how to deal with a malfunction of a beer server installed in the local environment 1. The user terminal 40 and the prompt generation system 10 can communicate via a local area network and / or a wide area network such as the Internet. In addition to the query entered by the user, the user terminal 40 may also send location information of the user terminal to the prompt generation system 10.

[0019] The prompt generation system 10 obtains information related to an inquiry received from a user terminal 40 from a device group 45 and a database in the local environment 1. Information from the device group 45 is sent to the prompt generation system 10 as IoT (Internet of Things) information. IoT information can include, for example, management information stored in the beer server indicating the temperature, status, and location of the beer server, as well as information on the surrounding environment such as temperature and humidity. The prompt generation system 10 can obtain information about the target beer server from a database stored within the system.

[0020] The prompt generation system 10 synthesizes (S1) IoT information acquired from an external source and / or information acquired from a database with a query received from a user terminal 40, and generates a prompt (S2). The prompt generation S2 can utilize an in-region LLM.

[0021] A local LLM is an LLM accessible only to authorized users and exists on a private network (local network). An open LLM is an LLM available to many users, including users of the local LLM, and exists on an open network (wide area network). Compared to an open LLM, a local LLM is inferior in the number of parameters (representing strength or performance), the amount of training data, and the freshness of the training data, but it has learned more information specific to the region.

[0022] The prompt generation system 10 performs a filtering process on the generated prompt (S3). The filtering S3 removes predetermined information, including information that should be kept secret from the outside, from the prompt to be sent to the open LLM 30. In this way, the prompt generation system 10 includes the functionality of a proxy server. This reduces the possibility that information that should be kept secret will be leaked from the prompt to the open LLM 30. The prompt generation S2 and filtering S3 may be performed simultaneously. For example, the local LLM can generate a prompt from input information, excluding confidential information.

[0023] The open LLM 30 receives prompts from the prompt generation system 10 via an open wide area network such as the Internet. The open LLM 30 generates answers to the received prompts and returns them to the prompt generation system 10.

[0024] The prompt generation system 10 generates (S6) learning data from the answers received from the open LLM 30 and stores it in a database. The learning data is used to train (learn) the local LLM.

[0025] The prompt generation system 10 performs a predetermined conversion process on the answer received from the open LLM 30 and generates an answer to be returned to the user terminal 40 (S5). The prompt generation system 10 transmits the prompt that has been subjected to filtering S3 to the open LLM 30. Therefore, the answer from the open LLM 30 is likely to be an inappropriate answer for the omitted information.

[0026] The prompt generation system 10 can generate a more appropriate response to a user's query by complementing the response from the open LLM 30 with the information removed by filtering S3. The response conversion S5 may be performed using the local LLM.

[0027] As described above, the dialogue system according to an embodiment of the present specification uses the prompt generation system 10 that utilizes the local LLM and the external open LLM 30 to generate a response to a query from a user.

[0028] If information such as the specific target device, installation location, and surrounding conditions is not explicitly specified, the Open LLM30 will return a generic answer, and it is not possible to expect an answer that is suited to a specific situation. The Local LLM can reflect information from IoT devices and provide answers that have learned information within the region, but the Local LLM has a small number of parameters, so it may not be able to obtain the expected answer.

[0029] The dialogue system of one embodiment of this specification can generate more appropriate responses by utilizing the local LLM and the open LLM 30. Furthermore, the prompt generation system 10 combines information from various IoT devices installed in the local environment 1 with the local LLM to generate prompts to instruct the open LLM 30 from the input text, environmental information, and information stored in the local LLM. This allows information that cannot be disclosed to the public to be blocked while supplementing it with other usable terms and expressions, making it possible to obtain real-time information that incorporates local knowledge.

[0030] In this system, information within the region is stored in the regional LLM, preventing it from being stored in the Open LLM30. Using the proxy function allows for pre-filtering of information that should not be made public. The performance of the regional LLM is generally based on the learning content at the time of construction, so the freshness of the accumulated information can deteriorate over time. By reflecting the results of the Open LLM30 generation in the learning of the regional LLM, this system allows for continuous improvement of the regional LLM.

[0031] Note that some of the above processing by the prompt generation system 10 may be omitted. For example, prompts may be generated without using IoT information. Also, one or both of the processing steps S6 for generating learning data for the local LLL from the response from the open LLM 30 and S5 for converting the response from the open LLM 30 using confidential information may be omitted.

[0032] 2 shows an example of the hardware configuration of prompt generation system 10. Fig. 2 shows an example having a general computer configuration, which includes an arithmetic unit 321, a main memory unit 322, an auxiliary memory unit 323, an input unit 324, an output unit 325, and a network interface 327.

[0033] The components of the prompt generation system 10 are communicatively connected to one another via a communication means such as a bus 326. Note that the entire configuration of the prompt generation system 10 may be realized by virtual resources such as a cloud server.

[0034] The arithmetic unit 321 is configured using a CPU (Central Processing Unit), an MPU (Micro Processing Unit), a GPU (Graphics Processing Unit), etc. The arithmetic unit 321 reads and executes programs stored in the main memory device 322, thereby realizing the functions of the prompt generation system 10.

[0035] The main memory device 322 is a device that stores programs and data, and is a ROM (Read Only Memory), RAM (Random Access Memory), NVRAM (Non-Volatile RAM), or the like.

[0036] The auxiliary storage device 323 is, for example, an NVRAM such as an SSD (Solid State Drive), an SD memory card, an optical storage device such as a CD (Compact Disc) or a DVD (Digital Versatile Disc), an HDD (Hard Disc Drive), or a storage area of ​​a cloud server. The auxiliary storage device 323 includes a non-transitory storage medium for storing programs and data. The programs and data stored in the auxiliary storage device 323 are loaded into the main storage device 322 as needed.

[0037] The input device 324 is an interface that accepts input of information, such as a keyboard, a mouse, a touch panel, a card reader, a microphone, etc. Alternatively, the prompt generation system 10 may be configured to accept input of information between itself and another device via some kind of communication means.

[0038] The output device 325 is an interface that outputs various types of information, and may be, for example, a screen display device such as a liquid crystal monitor, LCD (Liquid Crystal Display), or graphic card, a printer, or an audio output device such as a speaker. Alternatively, the prompt generation system 10 may be configured to output information to and from other devices via some kind of communication means.

[0039] The network interface 327 is a device that allows the prompt generation system 10 to communicate with other devices. Some of the components shown in Figure 2 may be omitted, and other components may be added.

[0040] Prompt generation system 10 can be configured with one or more computers. Thus, prompt generation system 10 can include one or more processors and one or more storage devices. The one or more processors operate as predetermined functional units by executing programs stored in the one or more storage devices.

[0041] The processing by the prompt generation system 10 is described in detail below. FIG. 3 shows information that the prompt generation system 10 acquires from an external source in order to generate an answer sentence. The prompt generation system 10 receives a query sentence from a user via a user terminal 40. In the example shown in FIG. 3, the query sentence asks what to do when a beer server does not serve beer.

[0042] The prompt generation system 10 acquires information (IoT information) from devices installed within the area and stores it in an internal database, in addition to queries from user terminals 40. For example, the prompt generation system 10 acquires information from sensors installed within the area, as well as from devices used by users within the area, such as beer servers, and stores the information in a database.

[0043] 4 is a block diagram showing a schematic functional configuration of the prompt generation system 10, which is related to the process of generating a prompt (query statement) for the open LLM 30 in response to a query statement from a user. The prompt generation system 10 includes an analysis unit 111, a device information complementation unit 112, an additional information acquisition unit 113, an IoT information complementation unit 114, a filtering unit 115, and an in-region LLM 121. These can be implemented by the calculation device 321 executing corresponding programs.

[0044] The prompt generation system 10 includes several databases. In the configuration example shown in Fig. 4, the prompt generation system 10 includes a device database (DB) 151 and an IoT database 152. The databases are stored in, for example, the secondary storage device 323. The device database 151 stores information about devices used by users within the area.

[0045] The device database 151 may include information such as the name, model number, and installation location of each device. The IoT database 152 stores data acquired from devices, including sensors, installed within the area. Data may be collected from the devices periodically and also when a user inquires about them.

[0046] The prompt generation system 10 complements information in a query received from a user terminal 40 and removes confidential information from a prompt to the open LLM 30. First, the analysis unit 111 performs a primary analysis of the query received from the user terminal 40. Specifically, the analysis unit 111 analyzes the input query using the local LLM 121 and extracts complementation elements contained in the query. The complementation elements are targets for which information should be complemented in order to generate a prompt for the open LLM 30.

[0047] An example of an input prompt to the local LLM 121 is, "Please extract complementary elements related to the device from the following inquiry: Question: 'Beer is not coming out of the installed beer server. Please tell me how to fix this.'" An example of an answer obtained from the local LLM 121 is, "The elements that can be supplemented are beer server, beer, and installation." The local LLM 121 has been trained using a large amount of training data that includes combinations of prompts and complementary elements, including query sentences from users. Note that complementary elements may be extracted by pre-defining complementary elements and extracting them through morphological analysis without using a machine learning model.

[0048] Next, the analysis unit 111 generates a prompt for completing information for the extracted completion element. A method for completing information for each completion element may be defined in advance for the analysis unit 111. The completion method specifies the type of information to be completed for the completion element and the source for obtaining that type of information.

[0049] The analysis unit 111 collects necessary complementary information from a database within the system and / or other systems. If the analysis unit 111 cannot acquire some or all of the necessary complementary information from a database within the system or other systems, it queries the user for the necessary complementary information. The information queried may be the complementary information itself or information necessary for collecting complementary information within the system.

[0050] Here, the analysis unit 111 determines to obtain information about the beer server from the device database 151 and generates a query for that purpose. As an example, the user terminal 40 transmits the location information of the user terminal 40 to the prompt generation system 10 together with the query entered by the user.

[0051] The analysis unit 111 generates a query to obtain, for example, the device name and device model number of the beer server (device type) at the received installation location from the device database 151. The query may be generated using the local LLM 121. The analysis unit 111 generates a prompt to obtain the query and sends it to the local LLM 121. The local LLM 121 has been trained using a large amount of training data consisting of prompt and query combinations. Note that the query may be generated without using a machine learning model, according to the information to be supplemented and its source information.

[0052] For example, the prompt to the local LLM121 to generate a query is "Please generate the query required to obtain in-house equipment information related to beer servers. When generating the query, please respond with the results in the following format: Search destination:, Search query:". The expected response is "The relevant information will be searched for using the following conditions: Search destination: Equipment information DB, Search query: select equipment name, equipment model, installation location... from Equipment information DB where equipment type = "beer server"". Note that the record structure of equipment information has already been learned in advance.

[0053] The device information complementing unit 112 acquires complementary information from the device database 151 using the query generated by the analyzing unit 111. The device information complementing unit 112 can acquire information on one or more beer servers from the device database 151. The acquired information is sent to the analyzing unit 111.

[0054] If the analysis unit 111 cannot collect the information necessary for completion from within the system or from another system, it requests the additional information acquisition unit 113 to make an inquiry to the user. This makes it possible to request details of the failure situation, device identification information, etc. from the user. For example, if information on multiple devices is acquired from the device database 151 and the device being queried cannot be identified, the additional information acquisition unit 113 prompts the user to perform a confirmation operation to complete the information. For example, the additional information acquisition unit 113 can request the reading of a barcode attached to the device, as exemplified in FIG. 4, or request the transmission of detailed information about the device.

[0055] The analysis unit 111 can acquire further necessary information based on device information acquired within the system or from a user. For example, the analysis unit 111 specifies a device that provides the target device with necessary complementary information, such as a sensor within the target device or a sensor located nearby, and requests the IoT information complementation unit 114 to acquire data. The analysis unit 111 may generate a query using the local LLM 121 and pass it to the IoT information complementation unit 114. Note that query generation may be performed rule-based without using a machine learning model.

[0056] The analysis unit 111 repeatedly collects information within the system and queries the user one or more times to collect necessary complementary information about the device that is the subject of the query. When analyzing the query statement from the user terminal 40, the analysis unit 111 may determine the number of repetitions based on the number of devices included in the query statement, the type of each device, the installation location of each device, etc.

[0057] The filtering unit 115 generates a prompt for the open LLM 30 from the query statement and the collected supplementary information from the user terminal 40. In generating the prompt, the filtering unit 115 excludes confidential information from the prompt. For example, the filtering unit 115 deletes prohibited character strings and formats the text (including summarization), thereby deleting the prohibited character strings or replacing them with general expressions. This makes it possible to exclude content that is prohibited from being output externally from the prompt.

[0058] The filtering unit 115 uses the local LLM 121 to generate a prompt to be sent to the open LLM 30 from the query sentence and the completion information from the user terminal 40. The filtering unit 115 uses the local LLM 121 to generate a prompt by deleting prohibited expressions or converting them into general expressions in the query sentence and the completion information received from the user terminal 40. The local LLM 121 has been trained in advance to, for example, format input text by removing prohibited expressions and generate a prompt. The training data includes many combinations of input prompts indicating input text and output prompts.

[0059] The filtering unit 115 may generalize prohibited expressions and then generate prompts from the generalized data. Alternatively, instead of using a machine learning model, predetermined prohibited expressions may be deleted or converted to general expressions based on rules. By using LLM to remove prohibited expressions from the target text and create a prompt, it becomes possible to format the text into a smooth expression even after removing prohibited expressions.

[0060] 5 shows an example of text in which information has been supplemented for a query from the user terminal 40, and an example of text (prompt) that has been filtered for anonymity. Text 401 is a query sent from the user terminal 40 to the prompt generation system 10. Text 402 shows text in which device information has been supplemented for the query 401 sent from the user terminal 40. As described above, the device ID and model obtained from the device database 151 have been added to the query 401.

[0061] Text 403 represents text that has been supplemented with additional information or IoT information obtained from an external source, as compared to text 402. As described above, the text may be supplemented with information obtained from an external system, information obtained from the user, and information obtained from the IoT database 152, for example.

[0062] For ease of understanding, prohibited expressions are underlined in text 403. In other words, the device ID, the brand of beer used, the last maintenance date, and the name of the store where the device is installed are included in the confidential information. The prompt generation system 10 generates text 404 from text 403 by deleting or generalizing the prohibited expressions and formatting the text. Text 404 is a prompt, which is a query sent to the open LLM 30.

[0063] Next, we will explain how the prompt generation system 10 processes a response sentence (generation result) to a prompt 404 received from the open LLM 30. Figure 6 is a block diagram for explaining the processing of the prompt generation system 10 for the open LLM 30. The prompt generation system 10 includes a generation result receiving unit 117, an information conversion unit 118, and a training data generation unit 119. These can be implemented by the calculation device 321 operating in accordance with a predetermined program.

[0064] The prompt generation system 10 includes a training data database 154. The training data database 154 is stored, for example, in the secondary storage device 323. The training data database 154 stores the training data of the local LLM 121.

[0065] The generation result receiving unit 117 receives the answer sentence, which is the generation result in response to the prompt 404, from the open LLM 30. The generation result receiving unit 117 transmits the received answer sentence to the information conversion unit 118. The information conversion unit 118 converts a part of the answer sentence into an internal representation and also formats it. Specifically, the information conversion unit 118 restores the prohibited expressions (anonymization information) in the prompt 404 to the answer sentence. This makes it possible to return a more specific answer to the user. Note that the restoration of the anonymization information may be omitted.

[0066] The information conversion unit 118 can obtain information on the anonymized forbidden expressions from the filtering unit 115 and restore the anonymized information in the answer sentence based on that information. The information on the deleted portion may be included in the session information stored in the main memory device 322. The information conversion unit 118 can restore the confidential information removed from the prompt in the answer sentence using the in-domain LLM 121. The in-domain LLM 121 has been trained with training data including many combinations of inputs of pre-formatted answer sentences and deleted expressions and formatted answer sentences generated from them.

[0067] The training data generation unit 119 generates training data for the local LLM 121 from the answer sentences received from the open LLM 30 and the prompts (queries) sent to the open LLM 30. The generated training data is stored in the training data database 154. The generated training data is data for training the local LLM 121 so that it can generate appropriate answer sentences on behalf of the open LLM 30. This allows the training data for the local LLM 21 to be generated efficiently. Note that the generation of training data may be omitted.

[0068] The training data generation unit 119 extracts prompts for training data from the prompts sent to the open LLM 30, and extracts answers for training data from the answer sentences of the open LLM 30. One or more answers may be extracted from the answer sentences of the open LLM 30. The training data generation unit 119 may execute prompt and answer extraction for training data using the local LLM 121. The local LLM 121 has been trained using training data of combinations of full prompts and extracted prompts, and training data of combinations of full answer sentences and extracted answer sentences. Note that training data may be extracted from the full prompts and full answer sentences without using the local LLM 121.

[0069] 7 shows a response message 451 received from the open LLM 30 and text 452 in which confidential information has been restored from the response message 451. The text 452 is a response message presented to the user who made the inquiry. As can be seen by comparing the texts 451 and 452, confidential information such as the brand used, the last maintenance date, and the store where the product was installed has been restored in the text 452.

[0070] 8 shows a response sentence 451 received from the open LLM 30 and training data 454 generated based on the response sentence 451. The training data 454 shows a prompt extracted from the prompt 404 sent to the open LLM 30, "What if the DBF-120WD doesn't dispense beer?", along with multiple responses (completions) to the prompt. This training data 454 is used to train the local LLM 121 by a training program (not shown).

[0071] Next, the processing of the prompt generation system 10 will be described with reference to the flowchart shown in FIG. 9. First, the analysis unit 111 receives a query from a user terminal (S101). The analysis unit 111 analyzes the received query and extracts complement elements (S102). The analysis unit 111 then identifies a data complement method corresponding to the extracted complement elements (S103). The data complement method indicates the type of data to be complemented and its source.

[0072] The analysis unit 111 acquires necessary complementary information from the device database 151 and the IoT database 152 using the device information complementation unit 112 and the IoT information complementation unit 114 (S104). If additional information is required for the searched complementation method, the additional information acquisition unit 113 can acquire the complementary information from a user or an external system (S105). The additional information acquisition unit 113 generates an inquiry to the inquiry target (user or other system) and executes the inquiry. The additional information can be acquired by making an inquiry to the user in a chat format, or by linking with another system via API, etc.

[0073] The analysis unit 111 sequentially acquires each element of the complementary information, and when a predetermined amount of information is acquired, the analysis unit 111 integrates the acquired complementary information with the query from the user to generate query information (S106). For example, in the example shown in Fig. 9, after steps S104 and S105 are each repeated a predetermined number of times, the complementary information is integrated into the query (S106).

[0074] Next, the filtering unit 115 removes the information to be kept secret from the inquiry information based on the information to be kept secret, and generates a prompt (query sentence) to be sent to the open LLM 30 (S107). The filtering unit 115 sends the generated prompt to the open LLM 30 (S108).

[0075] The generation result receiving unit 117 receives the answer sentence from the open LLM 30 (S109). The information converting unit 118 complements the confidential information missing in the filtering S107 in the received answer sentence (S110). The information converting unit 118 transmits the text with the confidential information complemented to the user terminal 40 as the inquiry result (S111).

[0076] The answer sentence received from the open LLM 30 is used as training data for the local LLM 121. The training data generation unit 119 extracts a prompt that succinctly indicates the content of the inquiry from the full prompt sent to the open LLM 30, and further extracts one or more different texts that indicate the answer to the inquiry from the full answer sentence received from the open LLM 30. The training data generation unit 119 stores the extracted combinations of prompt and answer in the training data database 154.

[0077] The present invention is not limited to the above-described embodiments and includes various modifications. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and are not necessarily limited to those including all of the described configurations. Furthermore, it is possible to replace part of the configuration of one embodiment with the configuration of another embodiment, or to add the configuration of another embodiment to the configuration of one embodiment. Furthermore, it is possible to add, delete, or replace part of the configuration of each embodiment with other configurations.

[0078] Furthermore, the above-mentioned components, functions, processing units, etc. may be realized in part or in whole by hardware, for example, by designing them as integrated circuits. Furthermore, the above-mentioned components, functions, etc. may be realized in software by a processor interpreting and executing a program that realizes each function. Information such as the programs, tables, and files that realize each function can be stored in memory, a storage device such as a hard disk or SSD, or a storage medium such as an IC card or SD card.

[0079] In addition, the control lines and information lines shown are those that are considered necessary for the explanation, and do not necessarily show all the control lines and information lines in the product. In reality, it can be assumed that almost all components are interconnected. [Explanation of symbols]

[0080] 10 Prompt Generation System 30 Open LLM 40 User terminals 111 Analysis Department 112 Device Information Complementary Unit 113 Additional information acquisition unit 114 IoT Information Complementary Department 115 Filtering section 117 Generation result receiving unit 118 Information Conversion Unit 119 Learning Data Generation Unit 121 Regional LLM 321 Arithmetic equipment 322 Main storage 323 Subsidized Memory Device

Claims

1. 1. A system for generating prompts for a text generation model that generates answer sentences, comprising: one or more computing devices; one or more storage devices; the one or more computing devices, Obtain a query from the user, generating a prompt for a first text generation model based on the query; transmitting the generated prompt to the first text generation model over an open network; The system performs a filtering process to exclude designated confidential information in generating the prompt.

2. 10. The system of claim 1, the one or more storage devices store the confidential information excluded in the filtering process; the one or more computing devices, receiving a response sentence to the prompt from the first text generation model; Complementing the secret information in the answer text to generate a complemented answer text; The system returns the completed answer sentence to the user.

3. 3. The system according to claim 1 or 2, the one or more storage devices store a database containing information collected from one or more external devices; the one or more computing devices, Extracting a complementary element from the query; acquiring pre-specified complementary information for the complementary element from the database; The system generates the prompt based on the query and the complementary information.

4. 3. The system according to claim 1 or 2, the one or more computing devices, Extracting a complementary element from the query; querying the user for complementary information for the complementary element; The system generates the prompt based on the complementary information received from the user and the query.

5. 3. The system according to claim 1 or 2, the one or more storage devices store a second text generation model that has been trained to return a response sentence in response to an input prompt; The one or more computing devices generate the prompt using the second text generation model.

6. 6. The system of claim 5, the one or more computing devices, receiving a response sentence to the prompt from the first text generation model; The system generates training data for the second text generation model from the prompt and the answer sentence and stores the training data in a training data database.

7. 1. A method for generating prompts for a text generation model that generates answer sentences, comprising: The system receives a query from a user, the system generates a prompt for a first text generation model based on the query; the system transmitting the generated prompt to the first text generation model over an open network; The method wherein the system performs a filtering process to exclude designated confidential information in generating the prompt.

Images