Generative AI support system and generative AI support program

Abstract

This disclosure relates to a generative AI support system, and aims to enable a generative AI to generate correct answers that require special non-public information. [Solution] When a user 18 inputs learning information, a learning adapter 32 generates learning vector data and a unique ID. A set of the learning information and the unique ID is registered in a local DB 36. The unique ID and learning vector data are registered in an external vector DB. When a user 18 inputs a question, a question adapter 34 converts the question into question vector data and queries a vector DB 22 for one or more unique IDs having learning vector data similar to the question vector data. The question adapter 34 retrieves learning information having the unique ID returned from the vector DB from the local DB 36, and poses the question to a generation AI 24 with a context including the learning information.

Description

[Technical field]

[0001] This disclosure relates to a generative AI support system and a program for supporting generative AI, and in particular to a generative AI support system and a program for supporting generative AI that are suitable for increasing the probability of a generative AI that has the function of responding to questions input in natural language. [Background technology]

[0002] Generative AI, including Chat GPT developed by OpenAI, is making remarkable progress. For example, the following non-patent document 1 provides an overview of how Chat GPT processes natural language. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] A rough understanding of ChatGPT's core technology "GPT", Oracle Japan https: / / qiita.com / ksonoda / items / b767cbd283e379303178 Summary of the Invention [Problem to be solved by the invention]

[0004] As described in Non-Patent Document 1 above, a generation AI that handles natural language converts text entered by a user into vector data. The converted vector data is provided to a trained language model that uses a huge amount of text data as learning material. The language model selects vector data similar to the input vector data according to the probability of the selection, and makes the words or phrases corresponding to the vector data part of the generated document. Thereafter, the words etc. that have been made part of the generated document are newly provided to the language model, and the above process is repeated. As a result, the generation AI generates document material that can serve as a response to the text entered by the user.

[0005] However, a generation AI generates responses within the scope of information that can be collected. For this reason, a generation AI may return an incorrect response to a question related to information that it cannot collect. In other words, if a question from a user is one for which the correct answer can be generated based on publicly available information, the existing generation AI can return the correct answer with a high probability, but if the question is related to special information, it will not be able to return the correct answer.

[0006] This disclosure has been made to solve the problems described above, and its primary objective is to provide a generative AI support system that allows an existing generative AI to generate correct answers that require special information while keeping that information confidential. A second objective of this disclosure is to provide a program for assisting a generation AI in causing an existing generation AI to generate a correct answer that requires special information while keeping that information confidential. [Means for solving the problem]

[0007] In order to achieve the above object, a first aspect is a generative AI support system that supports a generative AI that processes natural language, The generation AI has a function of receiving input in natural language and generating a response on a network accessible to an unspecified number of Internet users, A communication interface for accessing the generation AI and a vector DB present on the network; an input / output interface for accepting input from a user, the input including text information in a natural language, and for outputting responses to the user; a learning adapter for processing learning information included in the input; a query adapter for processing natural language directives contained in said input; a local DB capable of communicating with the learning adapter and the question adapter and having access to which is blocked from the network; The learning adapter includes: A process of generating learning vector data as vector data for processing text information included in the learning information as natural language; A process of generating a unique ID specific to the learning information; A process of registering a set of the learning information and the unique ID in the local DB; A process of registering the unique ID and the learning vector data in the vector DB, The query adapter: A process of generating question vector data as vector data for processing the command as a natural language; A process of querying the vector DB for one or more unique IDs having learning vector data similar to the query vector data; A process of querying the local DB for learning information having a unique ID returned from the vector DB; A process of issuing the command to the generation AI with a context including the learning information returned from the local DB; It is desirable that the system is configured to execute a process of outputting a response of the generation AI to the question to the user.

[0008] A second aspect is a program for assisting a generation AI that processes natural language, comprising: The generation AI has a function of receiving input in natural language and generating a response on a network accessible to an unspecified number of Internet users, A process for accessing the generation AI and a vector DB present on the network via a communication interface; receiving input from a user via an input / output interface, the input including text information in a natural language; outputting a response to the user via the input / output interface; A process for accessing a local DB to which access from the network is blocked; A process of generating learning vector data as vector data for processing text information included in the learning information included in the input as natural language; A process of generating a unique ID specific to the learning information; A process of registering a set of the learning information and the unique ID in the local DB; A process of registering the unique ID and the learning vector data in the vector DB; generating question vector data as vector data for processing a command in natural language included in the input as natural language; A process of querying the vector DB for one or more unique IDs having learning vector data similar to the query vector data; A process of querying the local DB for learning information having a unique ID returned from the vector DB; A process of issuing the command to the generation AI with a context including the learning information returned from the local DB; A process of outputting a response of the generation AI to the question to the user; It is preferable that the present invention includes a program for causing a computer to execute the above. Effect of the Invention

[0009] According to the first and second aspects, it is possible to have an existing generation AI generate a correct answer that requires special information while keeping that information secret. [Brief description of the drawings]

[0010] [Figure 1] FIG. 1 is a diagram illustrating a basic configuration of a generation AI support system according to a first embodiment of the present disclosure. [Diagram 2] FIG. 1 is a diagram illustrating an improved configuration of a generation AI support system according to a first embodiment of the present disclosure. [Diagram 3] FIG. 3 is a diagram for explaining the operation of the learning phase of the generation AI support system shown in FIG. 2. [Figure 4]4 is a flowchart for explaining a process flow for implementing the learning adapter shown in FIG. 3. [Diagram 5] FIG. 3 is a diagram for explaining the operation of the question phase of the generation AI support system shown in FIG. 2. [Figure 6] 6 is a flowchart for explaining the flow of processing for implementing the query adapter shown in FIG. 5 . [Figure 7] 7 is a flowchart for explaining a modified example of the process of step 122 shown in FIG. 6. [Figure 8] 7 is a flowchart for explaining a modified example of the process of step 124 shown in FIG. 6. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] Embodiment 1 [Basic configuration of the first embodiment] Fig. 1 shows a basic configuration of a generative AI support system according to a first embodiment of the present disclosure. The generative AI support system shown in Fig. 1 includes a local computer system 10 (hereinafter simply referred to as "PC system 10"). The PC system 10 includes hardware such as a CPU, a GPU, and a memory. The PC system 10 realizes the functions of a learning adapter 12 and a question adapter 14 by having processors such as the CPU and GPU execute programs stored in the memory.

[0012] The PC system 10 is equipped with an input / output interface (I / F) 16 including a keyboard, mouse, microphone, display, and speaker. A user 18 can input information to be used as learning material (hereinafter referred to as "learning information") or commands in natural language to the PC system 10 via the input / output I / F 16. The commands include questions and conditions for the image to be generated. The PC system 10 can output a response to the command to the user 18 via the input / output I / F 16. A detailed explanation of the case where the command is a question will be given below.

[0013] In this embodiment, user 18 does not mean a specific individual, but means one or more natural persons who have the authority to use the generation AI support system of this embodiment. For example, if the generation AI support system of this embodiment is used to share knowledge within a specific organization, one or more natural persons belonging to that organization correspond to user 18.

[0014] The PC system 10 also includes a communication I / F 20. The PC system 10 can communicate with an external network such as the Internet via the communication I / F 20. More specifically, the PC system 10 of this embodiment can communicate with a vector database (DB) 22 and a generation AI 24 provided on a cloud via the communication I / F 20.

[0015] [Operation of the basic configuration shown in Figure 1] [Learning Phase] Generation AI24 is a system that generates natural language responses to natural language input, as typified by Chat GPT released by OpenAI. Generation AI24 can collect and learn from various information available on the Internet, and generate responses to questions based on the collected information. For this reason, it may return incorrect answers to special questions that are not publicly available.

[0016] In the learning phase, the generative AI support system of this embodiment can learn special information, such as information that is not generally available or information that is strongly related to anticipated questions. An example of the former is the following information, which is intended to be shared among staff at Hospital A:

[0017] What to bring when transferring from Hospital A to Hospital B: ·Medical information form; ·Nursing communication form; · Rehabilitation summary; ·image Three-day prescription *If you use the stroke or femoral linkage pass, add the following: ·For Collaborative Pass Staff; - Linkage pass for patient families.

[0018] As an example of the latter, when questions such as "Is there a story about a monkey and a crab forming a friendship?", "Is there a story that teaches us not to be greedy or take advantage of others?", or "Is there a story that teaches us that even small people can overcome difficulties with wisdom and courage?" are anticipated, text information of various stories can be learned. More specifically, in response to the above questions, text information such as "The Battle of the Monkey and the Crab," "The Rolling of the Rice Ball," and "Issun-boshi" can be learned.

[0019] The user 18 provides the PC system 10 with learning information required to give a correct answer to the user's question via the input / output I / F 16 (FIG. 1, step 1). The learning information is not limited to text information, but may include image information or audio information together with the text information.

[0020] The learning information input by the user 18 is processed by the learning adapter 12. The learning adapter 21 is equipped with a trained language model. This language model has a function of generating vector data for processing text input as natural language. When the learning adapter 12 receives learning information, it uses this function to generate vector data for natural language processing of the text contained therein. More specifically, here, one vector data is generated for a group of learning information, not for each word or phrase.

[0021] When the learning adapter 12 generates the vector data, it provides the vector data to the vector DB 22 together with the learning information (FIG. 1, step 2). The vector DB 22 is a database that contains a trained natural language model and has a function of additionally learning the learning information with the vector data.

[0022] [Question Phase] A user 18 can input a question in natural language to the generation AI support system of this embodiment (FIG. 1, step 3). The question input by the user 18 is processed by the question adapter 14. The question adapter 14 has a built-in trained natural language model and has a function of generating vector data for processing the question as natural language. Here, similar to the Transformer provided in Chat GPT, for example, vector data is generated for each word and phrase. In addition, each piece of vector data includes position information indicating the position in the document that constitutes the question and attention information indicating the relationship with other words and phrases contained in the document.

[0023] The question adapter 14 can provide the question converted into vector data to the vector DB 22 via the communication I / F 20. When the vector DB 22 receives vector data from the question adapter 14 in this way, it extracts information having vector data similar to the received vector data from the stored information. In other words, the vector DB 22 processes the vector data corresponding to the question using a language model, thereby extracting one or more pieces of information having vector data appropriate as a response to the question according to the probability of occurrence. The extracted information is returned to the question adapter 14 as information related to the question (FIG. 1, step 4).

[0024] For example, if the user 18 inputs a question such as "What should I bring when I transfer from Hospital A to Hospital B?", the vector DB 22 extracts the above-mentioned learning information related to this "items to bring" and provides it to the question adapter 14. Also, if the user 18 inputs a question such as "Is there a story about a monkey and a crab becoming friends?", the vector DB 22 provides the question adapter 14 with text information of one or more stories including "The Battle of the Monkey and the Crab."

[0025] When the question adapter 14 receives the learning information from the vector DB 22, it attaches a context including the learning information and poses the question from the user 18 to the generation AI 24 (FIG. 1, step 5).

[0026] That is, the generation AI 24 is simultaneously provided with a question from the user 18 and the learning information necessary to correctly answer the question. In this case, the generation AI 24 refers to the provided learning information and generates a correct answer to the question with a high probability. The answer generated by the generation AI 24 is provided to the question adapter 14 via the communication I / F 20 (FIG. 1, step 6). This answer is then output to the user 18 via the input / output I / F 16. As a result, the user 18 can obtain a correct answer to the question with a high probability.

[0027] [Issues with the basic configuration shown in Figure 1] As described above, in the support system with the basic configuration shown in Fig. 1, it is necessary to store the learning information in the vector DB 24 that exists on the cloud. The vector DB 24 is a device that is assumed to be accessed by an unspecified number of Internet users. Therefore, if the learning information is confidential information such as personal information, the basic configuration shown in Fig. 1 is not necessarily optimal from the viewpoint of security.

[0028] [Improved structure of the first embodiment] Fig. 2 shows an improved configuration for solving the above problems. In Fig. 2, the same components as those shown in Fig. 1 are denoted by the same reference numerals, and the description thereof will be omitted or simplified.

[0029] The improved configuration shown in Fig. 2 includes a PC system 30. Like the PC system 10 shown in Fig. 1, the PC system 30 includes hardware such as a CPU, a GPU, and a memory, and the processors such as the CPU and the GPU execute programs in the memory to realize functions described below.

[0030] 2 includes a local DB 36 in addition to the learning adapter 32 and the query adapter 34. The local DB 36 is hardware built into the PC system 30, and is physically located inside the communication I / F 20. In other words, the local DB 36 is not limited to being housed in the same housing as the learning adapter 32 and the query adapter 34, but is accessible only by the user 18 and is placed in an environment separated from a network accessed by an unspecified number of users.

[0031] In the improved configuration shown in Fig. 2, the learning adapter 32 and the question adapter 34 operate to provide support similar to that of the generative AI support system of the basic configuration shown in Fig. 1, while storing special learning information such as confidential information in the local DB 36. Below, the operation of the generative AI support system of the improved configuration shown in Fig. 2 will be described in detail with reference to Figs. 3 to 6.

[0032] [Operation of the improved configuration shown in Figure 2] [Learning Phase] Fig. 3 is a diagram for explaining the operation of the learning phase of the generative AI support system shown in Fig. 2. In Fig. 3, elements similar to those shown in Fig. 1 are given the same reference numerals and their explanations are omitted or simplified. Also, in Fig. 3, the input / output I / F 16 and the communication I / F 20 of the PC system 30 are omitted. The input / output I / F 16 and the communication I / F 20 are similarly omitted in Fig. 5, which will be referred to later.

[0033] For the improved configuration shown in FIG. 2, the user 18 can input learning information in the same manner as for the basic configuration shown in FIG. 1 (FIG. 3, step 11).

[0034] The input learning information is processed by the learning adapter 32. The learning adapter 32, like the learning adapter 12 shown in FIG. 1, is equipped with a trained language model and has a function of converting the learning information into vector data. In addition to this function, the learning adapter 32 shown in FIG. 2 has a function of assigning a unique ID to the learning information. The unique ID is an identifier consisting of a predetermined number of characters, symbols, etc., and is assigned individually to each piece of learning information to prevent duplication.

[0035] The learning adapter 32 shown in FIG. 2 registers a set of the learning information and the unique ID in the local DB 36 (FIG. 3, step 12).

[0036] Furthermore, the learning adapter 32 registers a set of the unique ID and the vector data attached to the learning information in the external vector DB 22 (FIG. 3, step 13). As a result, only the unique ID and the vector data, not the learning information itself, are added to the vector DB 22 existing on the cloud.

[0037] Fig. 4 is a flowchart for explaining the flow of processing executed by the PC system 30 to realize the function of the learning adapter 32. In the routine shown in Fig. 4, first, it is determined whether or not learning information has been input by the user 18 (step 100).

[0038] When it is determined that the learning information has been input, a unique ID is assigned to the learning information (step 102). The unique ID is generated according to a predetermined rule, for example, by generating consecutive numbers in sequence.

[0039] Next, the learning information is input to the trained language model to generate vector data for handling the learning information as natural language (step 104). As in the case shown in FIG. 1, one vector data is generated for each group of learning information.

[0040] After the above process is completed, the set of the learning information and the unique ID is then registered in the local DB 36 (step 106).

[0041] Next, the set of the unique ID and the vector data is registered in the external vector DB 22 (step 108). When the above processing is completed, the learning phase for the current learning information is completed.

[0042] [Question Phase] Fig. 5 is a diagram for explaining the operation of the question phase of the generative AI support system shown in Fig. 2. As in the case shown in Fig. 1, a user 18 can input a question in natural language to the generative AI support system with an improved configuration (Fig. 5, step 14).

[0043] A question from the user 18 is processed by the question adapter 34. Like the question adapter 14 shown in FIG. 1, the question adapter 34 has a function of generating vector data for processing a question as natural language. Then, the question adapter 34 provides the vectorized question to the external vector DB 22. As described above, the vector DB 22 extracts vector data similar to the received vector data. Therefore, if the question posed by the user 18 is related to learning information, the vector DB extracts the vector data attached to the learning information. Thereafter, the vector DB 22 returns the unique ID registered as a set with the extracted vector data to the question adapter 34 (FIG. 5, step 15).

[0044] When the query adapter 34 receives a unique ID from the vector DB 22, it queries the local DB 36 for learning information paired with the unique ID. When the local DB 36 receives such a query, it provides the query adapter 34 with the learning information registered in combination with the received unique ID (FIG. 5, step 16).

[0045] Thereafter, like the question adapter 14 shown in Fig. 1, the question adapter 34 attaches a context including the received learning information and poses the question from the user 18 to the generation AI 24 (Fig. 5, step 17). For example, in response to the question "What should you bring when transferring from Hospital A to Hospital B?", the question adapter 34 adds the above learning information about this "things to bring" and executes a query to the generation AI 24. Alternatively, in response to the question "Is there a story about a monkey and a crab becoming friends?", the question adapter 34 attaches text information of one or more stories including "The Battle of the Monkey and the Crab" and executes a query to the generation AI 24 (Fig. 5, step 17).

[0046] The generation AI 24 generates a correct answer to the question with a high probability by referring to the provided learning information. The generated answer is provided to the question adapter 14 (FIG. 5, step 18), and further provided to the user 18 by any output method (FIG. 5, step 19). As a result, the user 18 can obtain the correct answer to the question with a high probability.

[0047] Fig. 6 is a flowchart for explaining the flow of processing executed by the PC system 30 to realize the function of the question adapter 34. In the routine shown in Fig. 6, first, it is determined whether or not a question has been input by the user 18 (step 110).

[0048] If it is determined that a question has been input, the text information constituting the question is converted into vector data (step 112).

[0049] Next, the vector data corresponding to the query is provided to the external vector DB 22. That is, a query is executed on the vector DB 22 using the vector data (step 114).

[0050] The vector DB 22 extracts one or more vector data similar to the received vector data based on the probability of relevance. Next, the unique ID registered as a set with the extracted vector data is read out. The read unique ID is then returned to the query adapter 34. During this time, the query adapter 34 waits for receipt of the unique ID from the vector DB 22 (step 116).

[0051] Upon receiving the unique ID, the inquiry adapter 34 then queries the local DB 36 for learning information registered in combination with the received unique ID (step 118).

[0052] Thereafter, the query adapter 34 remains in a waiting state until learning information is received from the local DB 36 (step 120).

[0053] Then, upon receiving the learning information, the question adapter 34 poses a question with a context including the learning information to the generation AI 24 (step 122).

[0054] Thereafter, when a response is received from the generation AI 24 (step 124), the question adapter 34 outputs the received response to the user 18 using any output method (step 126).

[0055] According to the above process, a question with the necessary information for answering can be posed to the generation AI 24. Therefore, according to the generation AI support system of this embodiment, it is possible to cause the generation AI 24 to return a correct answer to a question with a high probability. Furthermore, according to the improved configuration shown in FIG. 2, the learning information itself is stored in the local DB 36 and is therefore kept in a confidential state. Therefore, according to the improved generation AI support system, it is possible to achieve the effect of causing a generation AI shared by an unspecified number of Internet users to generate a correct answer to a special question with a high probability while keeping the learning information in a confidential state.

[0056] [Modification of the first embodiment] Hereinafter, with reference to Figs. 7 and 8, a modified example of the generation AI support system for increasing the probability that the generation AI will generate a correct answer will be described.

[0057] Fig. 7 is a flowchart for explaining a modified example of the process of step 122 shown in Fig. 6. Also, Fig. 8 is a flowchart for explaining a modified example of the process of step 124 shown in Fig. 6. The modified generation AI support system can be realized by having the improved PC system 30 shown in Fig. 2 execute the processes shown in Figs. 7 and 8 instead of steps 122 and 124.

[0058] 7 is executed after the question adapter 34 receives learning information from the local DB 36, as in the case shown in Fig. 6. Here, a contextual question is first generated according to the following rules (step 122-1).

[0059] A context is added to the user's question, including the following information: (1) Learning information (2) A prompt to "Limit the sources of information used to generate a response to the attached learning information" (3) A prompt to "provide a source of information that supports your response"

[0060] When a question with a context is generated by the above process, the question is sent to the generation AI 24 (step 122-2). Note that the context may further include a prompt specifying an out, as shown below. (4) A prompt stating, "If you don't know the answer, just reply 'I don't know.'"

[0061] The generation AI 24 generates a response according to the prompt attached to the question. Therefore, when the process shown in Fig. 7 is executed, the generation AI 24 generates a response based on the learning information. Furthermore, the generation AI 24 returns the information source on which the response is based to the question adapter 34 together with the response.

[0062] In the modified generation AI support system, step 124 shown in Fig. 8 is then executed. First, it is determined whether or not a response and a reason have been received from the generation AI 24 (step 124-1).

[0063] Upon receiving the response and the evidence, the question adapter 34 again sends a prompt to the generation AI 24 to check the consistency of the response and the evidence. For example, if the response is "A" and the source of the evidence is "B", the question "Does the information of B contain the information of A?" is sent to the generation AI 24 (step 124-2).

[0064] When the generation AI 24 receives such a question, it judges whether or not the information of A is included in the information of B, and returns the result. If the returned result is "True", it can be determined that the response generated by the generation AI 24 is trustworthy. On the other hand, if the result is "Fault", it can be determined that the response from the generation AI 24 is not trustworthy.

[0065] Following the processing of step 124-2 above, the query adapter 34 determines whether the result received from the generation AI 24 is "True" (step 124-3).

[0066] If the result is "True," the question adapter 34 outputs the response generated by the generation AI 24 to the user 18 (step 124-4).

[0067] On the other hand, if the result is not "True," question adapter 34 outputs a message such as "Cannot answer" to user 18 (step 124-5).

[0068] According to the above process, when the generation AI 24 generates an incorrect answer, a message such as "I cannot answer" is output to the user 18, thereby preventing the incorrect answer from being output to the user 18. Therefore, according to the above modification, it is possible to increase the probability that the correct answer will be output to the user 18, and effectively solve the problem of incorrect answers that is inherent to the generation AI 24.

[0069] Incidentally, in the above description of the modified example, the context attached to the question always includes a combination of a prompt to limit the information source (above (2)) and a prompt requesting the basis for the reply (above (3)), but the present disclosure is not limited to this. The context may include only the prompt to limit the information source. Alternatively, this prompt may be removed and the context may include only the prompt requesting the basis for the reply. In the latter case, since the basis for the reply is not presented by the generation AI 24, it is sufficient to execute only the process described with reference to FIG. 6 in step 124.

[0070] In the above modified example, the prompt countermeasures described with reference to Figures 7 and 8 are combined with the improved configuration shown in Figure 2. However, the present disclosure is not limited to this, and the above prompt countermeasures may be combined with the basic configuration shown in Figure 1. Specifically, at least one of the prompt countermeasures (2) and (3) above may be combined with the generation AI support system of the basic configuration shown in Figure 1. In this case, the learning information cannot be kept secret, but a novel effect of using the learning information to cause the generation AI 24 to generate a correct answer with a high probability can be achieved.

[0071] In the above-mentioned first embodiment and its modified example, the case where the user 18 inputs a question and receives a response in natural language has been described, but the present disclosure is not limited to this. For example, the generation AI 24 may generate an image according to given conditions. In this case, the user 18 can input the conditions as instructions. The generation AI support system may provide the generation AI 24 with an instruction with a context including learning information, thereby causing the generation AI 24 to generate a special image that cannot be generated only from publicly available information. [Explanation of symbols]

[0072] 10, 30 Local computer system (PC system) 12, 32 Learning Adapter 14, 34 Question Adapter 16 Input / Output Interface (I / F) 18 users 20 Communication Interface (I / F) 22 Vector Database (DB) 24 Generation AI 36 Local Database (DB)

Claims

1. A generative AI support system that supports a generative AI having a function of processing natural language, The generation AI has a function of receiving input in natural language and generating a response on a network accessible to an unspecified number of Internet users, A communication interface for accessing the generation AI and a vector DB present on the network; an input / output interface for accepting input from a user, the input including text information in a natural language, and for outputting responses to the user; a learning adapter for processing learning information included in the input; a query adapter for processing natural language directives contained in said input; a local DB capable of communicating with the learning adapter and the question adapter and having access to which is blocked from the network; The learning adapter includes: A process of generating learning vector data as vector data for processing text information included in the learning information as natural language; A process of generating a unique ID specific to the learning information; A process of registering a set of the learning information and the unique ID in the local DB; A process of registering the unique ID and the learning vector data in the vector DB, The query adapter: A process of generating question vector data as vector data for processing the command as a natural language; A process of querying the vector DB for one or more unique IDs having learning vector data similar to the query vector data; A process of querying the local DB for learning information having a unique ID returned from the vector DB; A process of issuing the command to the generation AI with a context including the learning information returned from the local DB; A process of outputting the response of the generation AI to the question to the user.

2. The generative AI support system of claim 1 , wherein the question adapter is further configured to include in the context a prompt limiting the information source for generating the response to the learning information contained in the context.

3. The instruction is a question, the response is a natural language response; The query adapter: including in the context a prompt along with the response indicating the source of information used to support the response; and after receiving the response and the information source, sending a prompt to the generation AI to query the consistency of the response and the information source; The generation AI support system according to claim 1 , wherein the process of outputting the generation AI's response to the user is executed on the condition that a response is obtained indicating that the response is consistent with the information source.

4. The generation AI support system of claim 3, wherein the question adapter is further configured to execute a process of outputting to the user a message indicating that an answer is not possible if an answer indicating that the reply is consistent with the information source is not obtained.

5. A generative AI support program for supporting a generative AI that processes natural language, The generation AI has a function of receiving input in natural language and generating a response on a network accessible to an unspecified number of Internet users, A process for accessing the generation AI and a vector DB present on the network via a communication interface; receiving input from a user via an input / output interface, the input including text information in a natural language; outputting a response to the user via the input / output interface; A process for accessing a local DB to which access from the network is blocked; A process of generating learning vector data as vector data for processing text information included in the learning information included in the input as natural language; A process of generating a unique ID specific to the learning information; A process of registering a set of the learning information and the unique ID in the local DB; A process of registering the unique ID and the learning vector data in the vector DB; generating question vector data as vector data for processing a command in natural language included in the input as natural language; A process of querying the vector DB for one or more unique IDs having learning vector data similar to the query vector data; A process of querying the local DB for learning information having a unique ID returned from the vector DB; A process of issuing the command to the generation AI with a context including the learning information returned from the local DB; A process of outputting a response of the generation AI to the question to the user; A generative AI support program including a program for causing a computer to execute the above.

6. The program for assisting a generation AI according to claim 5, further comprising a program for causing the computer to execute a process of including in the context a prompt to limit the information source for generating the response to the learning information contained in the context.

7. The instruction is a question, the response is a natural language response; including in the context a prompt with the response providing the source of information used to support the response; After receiving the response and the information source, sending a prompt to the generation AI to query the consistency of the response and the information source; The method further includes a program for causing a computer to execute the The program for assisting generation AI according to claim 5, wherein the process of outputting the generation AI's response to the user is executed on the condition that a response is obtained indicating that the response is consistent with the information source.

8. The program for generating AI assistance as described in claim 7, further comprising a program for causing a computer to execute a process of outputting to the user a message indicating that an answer is not possible if an answer is not obtained indicating that the reply is consistent with the information source.

Images