Chat control system, chat control method and program

Abstract

To provide a chat control system, a chat control method, and a program that can reduce the load for generating prompts while allowing a user to obtain information from a plurality of LLMs in an integrated manner, with the feeling that the user is chatting naturally with other people. [Solution] The intermediary agent system 10 stores shared information generated based on at least one of a chat message received from a user terminal 14 and a response received from an LLM 18a in a shared information storage unit. The intermediary agent system 10 outputs a prompt including a reference instruction to the shared information storage unit to an LLM 18c based on the chat message received from the user terminal 14. The intermediary agent system 10 outputs a chat message based on a response from the LLM 18c to the prompt to the user terminal 14.

Description

[Technical Field]

[0001] The present invention relates to a chat control system, a chat control method, and a program. [Background technology]

[0002] In recent years, techniques utilizing large-scale language models (LLMs) have been attracting attention. As an example of such techniques, Patent Literature 1 describes a technique that uses a large-scale language model when generating an automated assistant response. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Special Publication No. 2024-521053 Summary of the Invention [Problem to be solved by the invention]

[0004] By repeatedly outputting prompts to the LLM and receiving responses from it, users can obtain information as if they were chatting naturally with a person. In recent years, LLMs with various features have appeared, and it would be convenient if users could obtain various information in a unified manner without being aware of the existence of these LLMs.

[0005] However, because each of the multiple LLMs is independent, if the second LLM needs some of the information obtained through communication with the terminal or the first LLM, the second LLM must identify the information it needs from among the information, generate a prompt including the identified information, and output the generated prompt to the second LLM, which places a burden on the system for generating the prompt. In particular, if the amount of information obtained through communication between the terminal and the LLM is enormous, the burden of generating the prompt can be significant.

[0006] The present invention has been made in consideration of the above-mentioned problems, and one of its objectives is to provide a chat control system, chat control method, and program that can reduce the load for generating prompts while allowing users to obtain information from multiple LLMs in a unified manner, as if they were chatting naturally with a person. [Means for solving the problem]

[0007] (1) A chat control system according to the present invention includes: first prompt generation means for generating a prompt to a first large-scale language model based on a chat message received from a terminal; first prompt output means for outputting the prompt to the first large-scale language model to the first large-scale language model; first chat message output means for outputting to the terminal a chat message based on a response from the first large-scale language model to the prompt to the first large-scale language model; shared information generation means for generating shared information based on at least one of the chat message received from the terminal or the response received from the first large-scale language model and storing the generated shared information in shared information storage means; second prompt generation means for generating a prompt to a second large-scale language model including a reference instruction to the shared information storage means; second prompt output means for outputting the prompt to the second large-scale language model to the second large-scale language model; and second chat message output means for outputting to the terminal a chat message based on the response from the second large-scale language model to the prompt to the second large-scale language model.

[0008] (2) The chat control system may be as described in (1) above, wherein the first prompt output means outputs a prompt to each of the plurality of first large-scale language models, the first chat message output means outputs a chat message to the terminal for each of the plurality of first large-scale language models based on a response from the first large-scale language model, the shared information generation means generates a plurality of pieces of shared information and stores the generated plurality of pieces of shared information in the shared information storage means, and the second prompt generation means generates a prompt to the second large-scale language model including a reference instruction to the shared information storage means that stores the plurality of pieces of shared information.

[0009] (3) The chat control system may further include a missing information generation request receiving means for receiving a request to generate missing information from the second large-scale language model, and the shared information generation means may generate the missing information when receiving the generation request, and store the generated missing information as shared information in the shared information storage means.

[0010] (4) The chat control system may further include a third chat message output means for outputting a chat message requesting the missing information to the terminal when the generation request is accepted, and the shared information generation means may generate the missing information based on a chat message received from the terminal in response to the chat message requesting the missing information.

[0011] (5) The chat control system may further include a third prompt output means for outputting a prompt to a large-scale language model different from the second large-scale language model when the generation request is received, instructing the generation of the missing information, and the shared information generation means may generate the missing information based on a response to the prompt from the large-scale language model to which the prompt to instruct the generation of the missing information is output.

[0012] (6) The chat control system may be one of those described in any one of (1) to (5) above, wherein the second chat message output means outputs to the terminal a chat message based on a response from the first large-scale language model to a prompt to the first large-scale language model and a response from the second large-scale language model to a prompt to the second large-scale language model.

[0013] (7) A chat control method according to the present invention includes the steps of generating a prompt to a first large-scale language model based on a chat message received from a terminal, outputting the prompt to the first large-scale language model to the first large-scale language model, outputting a chat message to the terminal based on a response from the first large-scale language model to the prompt to the first large-scale language model, generating shared information based on at least one of the chat message received from the terminal or the response received from the first large-scale language model and storing the generated shared information in a shared information storage means, generating a prompt to a second large-scale language model including a reference instruction to the shared information storage means, outputting the prompt to the second large-scale language model to the second large-scale language model, and outputting a chat message to the terminal based on the response from the second large-scale language model to the prompt to the second large-scale language model.

[0014] (8) A program according to the present invention is a program for causing a computer to execute the steps of: generating a prompt for a first large-scale language model based on a chat message received from a terminal; outputting the prompt for the first large-scale language model to the first large-scale language model; outputting a chat message based on a response from the first large-scale language model to the prompt for the first large-scale language model to the terminal; generating shared information based on at least one of the chat message received from the terminal or the response received from the first large-scale language model and storing the generated shared information in a shared information storage means; generating a prompt for a second large-scale language model including a reference instruction to the shared information storage means; outputting the prompt for the second large-scale language model to the second large-scale language model; and outputting a chat message based on the response from the second large-scale language model to the terminal. [Brief explanation of the drawings]

[0015] [Figure 1] 1 is a diagram illustrating an example of a configuration of a chat system according to an embodiment of the present invention. [Figure 2] FIG. 2 is a functional block diagram showing an example of functions of the intermediary agent system according to an embodiment of the present invention. [Figure 3] FIG. 10 is a diagram illustrating an example of shared information. [Figure 4] FIG. 10 is a diagram illustrating an example of shared information. [Figure 5] FIG. 10 is a diagram illustrating an example of shared information. [Figure 6A] FIG. 2 is a flowchart showing an example of the flow of processing performed in the intermediary agent system according to one embodiment of the present invention. [Figure 6B] FIG. 2 is a flowchart showing an example of the flow of processing performed in the intermediary agent system according to one embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0016] Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings.

[0017] Fig. 1 is a diagram showing an example of the configuration of a chat system 1 according to this embodiment. As shown in Fig. 1, the chat system 1 according to this embodiment includes a mediation agent system 10, a plurality of processing agent systems 12 (12a, 12b, 12c, ...), and a user terminal 14. The mediation agent system 10, the processing agent systems 12, and the user terminal 14 are connected to a computer network 16 such as the Internet. Fig. 1 shows three processing agent systems 12 included in the chat system 1 as an example, but the number of processing agent systems 12 included in the chat system 1 is not limited to three.

[0018] The facilitating agent system 10 may be configured, for example, by a computer such as a server computer. The facilitating agent system 10 may include a processor 10a such as a CPU, a storage unit 10b such as a memory or a hard disk drive, a communication unit 10c such as a network interface card, etc. The facilitating agent system 10 may include multiple computers or may include a single computer.

[0019] The processing agent system 12 may be configured with, for example, a server computer including a CPU, a GPU, a memory, a communication interface, etc. The processing agent system 12 may include multiple computers or may include a single computer. In this embodiment, a trained large-scale language model (LLM) 18 is implemented in the processing agent system 12 according to this embodiment. As shown in FIG. 1 , in this embodiment, for example, an LLM 18a, an LLM 18b, and an LLM 18c are implemented in the processing agent system 12a, the processing agent system 12b, and the processing agent system 12c, respectively. The LLM 18a, the LLM 18b, and the LLM 18c are independent of each other.

[0020] The user terminal 14 may be configured, for example, by a computer (for example, a personal computer, a smartphone, a tablet terminal, etc.) including a CPU, a memory, a communication interface, an input device, and a display.

[0021] In this embodiment, the roles of the processing agent systems 12 are different from each other. For example, the processing agent system 12a may store a product purchase history of a user who uses the user terminal 14. Also, for example, the processing agent system 12b may store bank account information of the user of the user terminal 14. Also, for example, the processing agent system 12c may store information indicating the usage status of a credit card owned by the user of the user terminal 14.

[0022] In this embodiment, chat messages are communicated between the facilitating agent system 10 and the user terminal 14. That is, in this embodiment, for the facilitating agent system 10, the user terminal 14 corresponds to a terminal that is a communication partner of chat messages.

[0023] Furthermore, in this embodiment, the intermediary agent system 10 determines, as necessary, an output destination (here, for example, an output destination of a chat message or a prompt) from among the LLM 18a, LLM 18b, LLM 18c, and the user terminal 14 based on a chat message received from the user terminal 14. The method for determining the output destination will be described later.

[0024] Then, when the user terminal 14 is determined as the output destination, the facilitating agent system 10 outputs to the user terminal 14 a chat message that is a reply to the chat message accepted from the user terminal 14. Furthermore, when one of the LLMs 18 is determined as the output destination, the facilitating agent system 10 outputs a prompt to the determined LLM 18. Then, the facilitating agent system 10 generates a chat message based on the response from the LLM 18 to the prompt, and outputs the generated chat message to the user terminal 14.

[0025] In this way, the intermediary agent system 10 according to this embodiment mediates the exchange of data between the user terminal 14 and each of the multiple LLMs 18.

[0026] Fig. 2 is a functional block diagram showing an example of functions implemented in the facilitating agent system 10 according to this embodiment. Note that the facilitating agent system 10 according to this embodiment does not need to implement all of the functions shown in Fig. 2, and functions other than those shown in Fig. 2 may also be implemented.

[0027] As shown in FIG. 2, the intermediary agent system 10 according to this embodiment functionally includes, for example, a chat message receiving unit 20, an output destination determination unit 22, a prompt generation unit 24, a prompt output unit 26, a response receiving unit 28, a chat message generation unit 30, a chat message output unit 32, a shared information generation unit 34, and a shared information storage unit 36. The intermediary agent system 10 according to this embodiment functions as a chat control system that controls chats held in the chat system 1. The chat message receiving unit 20, the prompt output unit 26, the response receiving unit 28, and the chat message output unit 32 are implemented primarily in the communication unit 10c. The output destination determination unit 22, the prompt generation unit 24, the chat message generation unit 30, and the shared information generation unit 34 are implemented primarily in the processor 10a. The shared information storage unit 36 ​​is implemented primarily in the storage unit 10b.

[0028] The above functions may be implemented by executing a program including instructions corresponding to the above functions, which is installed in the intermediary agent system 10, which is a computer, on the intermediary agent system 10. This program may also be supplied to the intermediary agent system 10 via a computer-readable information storage medium such as an optical disk, magnetic disk, magnetic tape, or magneto-optical disk, or via the Internet, for example.

[0029] In this embodiment, the chat message receiving unit 20 receives chat messages from the user terminal 14, for example.

[0030] In this embodiment, the output destination determination unit 22 determines the output destination based on, for example, a chat message received by the chat message receiving unit 20. Here, for example, a topic indicated in the chat message received by the chat message receiving unit 20 may be estimated using a known natural language processing technique, such as topic modeling technology or estimation technology using a trained machine learning model. Then, the output destination may be determined based on the estimated topic. For example, if the received message is related to a purchase, such as "Please tell me what products you recommend," the LLM 18a may be determined as the output destination. Also, if the received message is related to a user's bank account, such as "Please tell me my deposit balance," the LLM 18b may be determined as the output destination. Also, if the received message is related to a credit card, such as "Please increase my credit card limit," the LLM 18c may be determined as the output destination. Furthermore, if the received message is other than the above, the user terminal 14 may be determined as the output destination.

[0031] In this embodiment, for example, when any of the LLMs 18 is determined as an output destination, the prompt generation unit 24 generates a prompt for the determined LLM 18. The prompt generation unit 24 may generate a prompt for the determined LLM 18 based on a chat message received by the chat message receiving unit 20.

[0032] The prompt output unit 26 outputs the prompt generated by the prompt generation unit 24 to the LLM 18 determined by the output destination determination unit 22.

[0033] The answer receiving unit 28 receives an answer to a prompt received from the LLM 18 to which the prompt output unit 26 outputs a prompt.

[0034] In this embodiment, the chat message generating unit 30 generates chat messages to be output to the user terminal 14, for example.

[0035] In this embodiment, the chat message output unit 32 outputs the chat message generated by the chat message generation unit 30 to the user terminal 14, for example.

[0036] When the output destination determination unit 22 determines the user terminal 14 as the output destination, the chat message generation unit 30 may generate a predetermined chat message or a chat message determined according to predetermined logic based on the received chat message. Then, the chat message output unit 32 may output the chat message generated in this manner to the user terminal 14.

[0037] Furthermore, when the answer receiving unit 28 receives an answer from the LLM 18, the chat message generating unit 30 may generate a chat message based on the answer in response to the reception of the answer. Here, a chat message including the answer may be generated. Note that the chat message may be the same character string as the answer, or may be a different character string. For example, a chat message may be generated by processing the answer (for example, by translating the answer). Then, the chat message output unit 32 may output a chat message based on the answer to the user terminal 14.

[0038] In this embodiment, the shared information generation unit 34 generates shared information that can be referenced by any of the multiple LLMs 18 included in the chat system 1, based on at least one of a chat message received from the user terminal 14 or a response received from the LLM 18, and stores the generated shared information in the shared information storage unit 36.

[0039] Here, the shared information generation unit 34 may identify ungenerated shared information from among multiple predetermined types of shared information. Then, the shared information generation unit 34 may determine, for each piece of necessary information identified as ungenerated, whether it can be generated based on a chat message received from the user terminal 14, using a known natural language processing technique, such as name recognition technology, topic modeling technology, or estimation technology using a trained machine learning model. Alternatively, the shared information generation unit 34 may determine, for each piece of necessary information identified as ungenerated, whether it can be generated based on a response received from the LLM 18, using, for example, the above-described natural language processing technique. Then, the shared information generation unit 34 may generate the shared information determined to be generable. Then, the shared information generation unit 34 may output the generated shared information to the shared information storage unit 36.

[0040] Figures 3, 4, and 5 are diagrams showing examples of shared information. As shown in Figures 3, 4, and 5, the shared information includes a combination of an index and a value. Figure 3 shows three pieces of shared information, Figure 4 shows five pieces of shared information, and Figure 5 shows six pieces of shared information.

[0041] In this embodiment, for example, in an initial state, there is no necessary information stored in the shared information storage unit 36. Then, as shared information is generated, the generated shared information is accumulated in the shared information storage unit 36.

[0042] For example, assume that the predetermined types (indexes) of shared information are "gender," "age group," "address," "hobbies," and "number of people in household." In this case, assume that it is determined that shared information with the index "gender," shared information with the index "age group," and shared information with the index "address" can be generated based on chat messages received from user terminal 14. In this case, these three pieces of shared information are generated and stored in shared information storage unit 36, as shown in FIG. 3.

[0043] In this embodiment, the prompt generation unit 24 may generate a prompt including a reference instruction to the shared information storage unit 36. The prompt output unit 26 may output the prompt including the reference instruction to the shared information storage unit 36 ​​to the LLM 18.

[0044] For example, assume that the URL of the shared information storage unit 36 ​​is "http: / / aaa.bbb.ccc." In this case, a prompt may be generated that includes a character string such as "Please answer by referring to the information under http: / / aaa.bbb.ccc."

[0045] Furthermore, the LLM 18 that receives the prompt may generate a response by referring to the reference indicated in the prompt.

[0046] Here, the LLM 18 may generate a response including a request to generate missing information, which is information that is missing when generating the response. The LLM 18 may then output the response generated in this manner to the processing agent system 12. The response receiving unit 28 of the processing agent system 12 may then receive the request to generate the missing information (here, for example, the response including the request to generate the missing information). When the response receiving unit 28 receives the request to generate the missing information, the shared information generating unit 34 may then generate the missing information and store the generated missing information as shared information in the shared information storage unit 36.

[0047] Here, when the response receiving unit 28 receives a request to generate the missing information, the shared information generating unit 34 may output an instruction to the chat message generating unit 30 to generate a chat message requesting the missing information. Then, the chat message generating unit 30 may generate a chat message requesting the missing information. Then, the chat message output unit 32 may output the chat message generated in this manner to the user terminal 14. Then, the shared information generating unit 34 may generate the missing information based on a chat message received from the user terminal 14 in response to the chat message requesting the missing information.

[0048] For example, suppose the answer receiving unit 28 receives an answer including the character string "Please tell me your hobbies and the number of people in your household." In this case, the chat message generation unit 30 may generate a chat message including the character string "Please tell me your hobbies and the number of people in your household.", and the chat message output unit 32 may output this chat message to the user terminal 14. The chat message receiving unit 20 may then receive a chat message output from the user terminal 14 in response to the chat message. The shared information generation unit 34 may then determine, based on the chat message received by the chat message receiving unit 20, whether shared information with the index "Hobbies" and shared information with the index "Number of people in your household" can be generated. If these pieces of shared information are generateable, the shared information generation unit 34 may generate these pieces of shared information and store the generated shared information in the shared information storage unit 36. In this case, as shown in FIG. 4, these two pieces of shared information are generated and stored in the shared information storage unit 36.

[0049] The prompt generation unit 24 may then generate a prompt including a string such as, for example, "Your hobbies and household size have been registered as information under http: / / aaa.bbb.ccc." The prompt output unit 26 may then output this prompt to the LLM 18, which has output a response including the string, "Please tell us your hobbies and the number of people in your household."

[0050] Furthermore, for example, when the answer receiving unit 28 receives a request to generate missing information, the shared information generating unit 34 may output an instruction to generate a prompt to the prompt generating unit 24, instructing the generation of the missing information. Then, the prompt generating unit 24 may generate a prompt instructing the generation of the missing information. Then, the prompt output unit 26 may output this prompt to an LLM 18 different from the LLM 18 that output the request to generate the missing information. Then, the shared information generating unit 34 may generate the missing information based on an answer to the prompt from the LLM 18 to which the prompt instructing the generation of the missing information is output.

[0051] For example, suppose the answer receiving unit 28 receives an answer including the character string "Please tell me your product purchase history." from the LLM 18c. In this case, the prompt generation unit 24 may generate a prompt including the character string "Please tell me your product purchase history." The prompt output unit 26 may output this prompt to the LLM 18a. The answer receiving unit 28 may then receive an answer output from the LLM 18a in response to the prompt. The shared information generation unit 34 may then generate shared information with an index of "purchase history" based on the answer received by the answer receiving unit 28, and store the generated shared information in the shared information storage unit 36. In this case, as shown in FIG. 5, this shared information is generated and stored in the shared information storage unit 36.

[0052] The prompt generation unit 24 may then generate a prompt including a character string such as, for example, "The product purchase history has been registered as information under http: / / aaa.bbb.ccc." The prompt output unit 26 may then output this prompt to the LLM 18c.

[0053] An example of the flow of processing performed in the intermediary agent system 10 according to this embodiment will now be described with reference to the flow diagrams shown in Figures 6A and 6B. In this processing example, it is assumed that the type (index) of shared information to be generated is predetermined.

[0054] In this processing example, the chat message receiving unit 20 waits for reception of a chat message (S101).

[0055] When the reception of a chat message is detected in the process shown in S101, the shared information generation unit 34 determines whether or not each piece of ungenerated shared information can be generated based on the chat message whose reception was detected in the process shown in S101 (S102).

[0056] Then, the shared information generation unit 34 generates the shared information determined to be generateable in the process shown in S102 based on the chat message whose acceptance was detected in the process shown in S101 (S103), and outputs the generated shared information to the shared information storage unit 36 ​​(S104). Note that if the necessary information determined to be generateable in the process shown in S102 does not exist, the processes shown in S103 and S104 are not executed.

[0057] Then, the output destination determination unit 22 determines an output destination based on the chat message whose acceptance was detected in the process shown in S101 (S105).The output destination determination unit 22 then checks whether the output destination determined in the process shown in S105 is any of the LLMs 18 (S106).

[0058] If the determined output destination is not the LLM 18 (it is the user terminal 14) (S106: N), the chat message generation unit 30 generates a chat message (S107), and the chat message output unit 32 outputs the chat message generated in the process shown in S107 to the user terminal 14 (S108), and the process returns to the process shown in S101.

[0059] If the determined output destination is the LLM 18 (S106: Y), the prompt generation unit 24 generates a prompt for the LLM 18, which includes a reference instruction to the shared information storage unit 36 ​​(S109), and the prompt output unit 26 outputs the generated prompt to the LLM 18 (S110). In the process shown in S109, for example, a prompt may be generated that includes a reference instruction to the shared information storage unit 36 ​​and a character string indicated by the chat message whose acceptance was detected in the process shown in S101.

[0060] Then, the response receiving unit 28 waits for a response to the prompt from the LLM 18 (S111).

[0061] When the reception of a response is detected in the process shown in S111, the shared information generation unit 34 determines whether or not each piece of ungenerated shared information can be generated based on the response whose reception was detected in the process shown in S111 (S112).

[0062] Then, the shared information generation unit 34 generates the shared information determined to be generateable in the process shown in S112 based on the response whose reception was detected in the process shown in S111 (S113), and outputs the generated shared information to the shared information storage unit 36 ​​(S114). Note that if the necessary information determined to be generateable in the process shown in S112 does not exist, the processes shown in S113 and S114 are not executed.

[0063] Then, the chat message generation unit 30 generates a chat message based on the response whose acceptance was detected in the process shown in S111 (S115), and the chat message output unit 32 outputs the chat message generated in the process shown in S115 to the user terminal 14 (S116), and the process returns to the process shown in S101.

[0064] In this embodiment, for example, the prompt generation unit 24 may generate a prompt for the first LLM 18 (e.g., LLM 18a) based on a chat message received from the user terminal 14. Here, for example, a prompt may be generated that includes a character string indicated in the chat message received from the user terminal 14. Alternatively, a prompt may be generated that includes a reference instruction to the shared information storage unit 36 ​​and a character string indicated in the chat message received from the user terminal 14. Then, the prompt output unit 26 may output the prompt to the first LLM 18.

[0065] Then, the chat message output unit 32 may output a chat message based on a response from the first LLM 18 (e.g., LLM 18a) to the prompt to the first LLM 18 (e.g., LLM 18a) to the user terminal 14. Here, for example, a chat message including a character string indicated by the response from the LLM 18a may be output to the user terminal 14.

[0066] Here, the shared information generation unit 34 may generate shared information based on at least one of a chat message received from the user terminal 14 or a response received from the first LLM 18 (e.g., LLM 18a), and store the generated shared information in the shared information storage unit 36.

[0067] Then, the prompt generator 24 may generate a prompt for the second LLM 18 (e.g., LLM 18c) that includes a reference instruction to the shared information storage unit 36. Here, for example, a prompt may be generated that includes a reference instruction to the shared information storage unit 36 ​​and a character string indicated by a chat message received from the user terminal 14.

[0068] For example, suppose that the prompt generation unit 24 generates a prompt for a first LLM 18 (e.g., LLM 18a) based on a chat message received from the user terminal 14. In this case, the prompt generation unit 24 may generate a prompt for a second LLM 18 (e.g., LLM 18c) based on the chat message, the prompt including a reference instruction to the shared information storage unit 36. Alternatively, the prompt generation unit 24 may generate a prompt for the second LLM 18 (e.g., LLM 18c) based on a chat message different from the first chat message, the prompt including a reference instruction to the shared information storage unit 36. The prompt output unit 26 may then output the prompt to the second LLM 18 (e.g., LLM 18c).

[0069] Then, the chat message output unit 32 may output a chat message based on a response from the second LLM 18 (e.g., LLM 18c) to the prompt to the second LLM 18 (e.g., LLM 18c) to the user terminal 14. Here, for example, a chat message including a character string indicated by the response from the LLM 18c may be output to the user terminal 14.

[0070] In the above description, there may be a plurality of the above-described first LLMs 18. For example, the LLMs 18a and 18b may correspond to the above-described first LLMs 18.

[0071] The prompt output unit 26 may then output a prompt to each of the multiple first LLMs 18. For example, the prompt output unit 26 may output a prompt to the LLM 18a based on a first chat message received from the user terminal 14, and output a prompt to the LLM 18b based on a second chat message received from the user terminal 14.

[0072] Then, the chat message output unit 32 may output, for each of the multiple first LLMs 18, a chat message based on a response from that first LLM 18 to the user terminal 14. For example, the chat message output unit 32 may output, to the user terminal 14, a chat message based on a response from the LLM 18a in response to the first chat message. Then, the chat message output unit 32 may output, to the user terminal 14, a chat message based on a response from the LLM 18b in response to the second chat message.

[0073] The shared information generator 34 may then generate multiple pieces of shared information and store the generated pieces of shared information in the shared information storage unit 36. For example, the shared information generator 34 may generate shared information based on at least one of the first chat message or a response received from the LLM 18a and store the generated shared information in the shared information storage unit 36. The shared information generator 34 may then generate shared information based on at least one of the second chat message or a response received from the LLM 18b and store the generated shared information in the shared information storage unit 36.

[0074] The prompt generator 24 may then generate a prompt for the second LLM 18 (e.g., LLM 18c) that includes a reference instruction to the shared information storage unit 36 ​​that stores multiple pieces of shared information. The prompt output unit 26 may then output this prompt to the second LLM 18 (e.g., LLM 18c).

[0075] In addition, in this embodiment, the chat message output unit 32 may output to the user terminal 14 chat messages based on a response from a first LLM 18 (e.g., LLM 18a) to a prompt to the first LLM 18 (e.g., LLM 18a), and a response from a second LLM 18 (e.g., LLM 18c) to a prompt to the second LLM 18 (e.g., LLM 18c).

[0076] For example, in response to the chat message receiving unit 20 receiving a chat message from the user terminal 14, the prompt generating unit 24 may generate a prompt for the LLM 18a and a prompt for the LLM 18c. In this case, the prompt for the LLM 18a and the prompt for the LLM 18c may be the same. For example, a prompt including a reference instruction to the shared information storage unit 36 ​​and a character string indicated by the chat message may be generated as the prompt for the LLM 18a and the prompt for the LLM 18c. Note that the prompt for the LLM 18a and the prompt for the LLM 18c may be different.

[0077] The prompt output unit 26 may then output a prompt to the LLM 18a to the LLM 18a, and may output a prompt to the LLM 18c to the LLM 18c. The reply receiving unit 28 may then receive a reply from the LLM 18a and a reply from the LLM 18c. The chat message generation unit 30 may then generate a chat message based on these replies (for example, a chat message including a character string indicating the reply from the LLM 18a and a character string indicating the reply from the LLM 18c). The chat message output unit 32 may then output this chat message to the user terminal 14.

[0078] Furthermore, in this embodiment, the shared information generation unit 34 may generate log information including chat messages received by the chat message receiving unit 20 and store the log information as shared information in the shared information storage unit 36. Furthermore, the shared information generation unit 34 may generate log information including answers received by the answer receiving unit 28 and store the log information as shared information in the shared information storage unit 36. Furthermore, the shared information generation unit 34 may generate log information including chat messages generated by the chat message generation unit 30 and store the log information as shared information in the shared information storage unit 36.

[0079] Then, the prompt output unit 26 may output to the LLM 18 a prompt including a reference instruction to the shared information storage unit 36 ​​that stores the log information.

[0080] In this embodiment, a user of a user terminal 14 can communicate chat messages with the processing agent system 12, thereby obtaining various information from these LLMs 18 in a unified manner without being aware of the existence of multiple LLMs 18.

[0081] Furthermore, in this embodiment, shared information is generated based on at least one of a chat message received from the user terminal 14 and an answer received from the first LLM 18, and the generated shared information is stored in the shared information storage unit 36. Then, based on the chat message received from the user terminal 14, a prompt for the second LLM 18, including an instruction to refer to the shared information storage unit 36, is output to the second LLM 18. Then, the second LLM 18 can select shared information necessary for generating an answer from the information stored in the shared information storage unit 36, and generate an answer based on the selected shared information.

[0082] Therefore, in this embodiment, even if the second LLM 18 requires some of the information obtained through communication with the user terminal 14 or the first LLM 18, it is not necessary to identify the information required by the second LLM 18 from among this information, generate a prompt including the identified information, and output the generated prompt to the second LLM 18.

[0083] In this way, according to this embodiment, the user can obtain information from multiple LLMs 18 in a centralized manner, as if they were chatting naturally with a person, while reducing the load of generating prompts.

[0084] The present invention is not limited to the above-described embodiment. For example, the shared information storage unit 36 ​​may be provided in a system different from the facilitating agent system 10, rather than in the facilitating agent system 10.

[0085] Furthermore, the specific character strings and numerical values ​​described above and the specific character strings and numerical values ​​in the drawings are examples, and the present invention is not limited to these character strings and numerical values. [Explanation of symbols]

[0086] 1 Chat system, 10 Intermediary agent system, 10a Processor, 10b Memory unit, 10c Communication unit, 12, 12a, 12b, 12c Processing agent system, 14 User terminal, 16 Computer network, 18, 18a, 18b, 18c Large scale language model (LLM), 20 Chat message reception unit, 22 Output destination determination unit, 24 Prompt generation unit, 26 Prompt output unit, 28 Answer reception unit, 30 Chat message generation unit, 32 Chat message output unit, 34 Shared information generation unit, 36 Shared information storage unit.

Claims

1. a first prompt generation means for generating a prompt for a first large-scale language model based on a chat message received from a terminal; a first prompt output means for outputting a prompt to the first large-scale language model to the first large-scale language model; a first chat message output means for outputting to the terminal a chat message based on a response from the first large-scale language model in response to a prompt to the first large-scale language model; shared information generation means for generating shared information based on at least one of chat messages received from the terminals and responses received from the first large-scale language model, and storing the generated shared information in shared information storage means; second prompt generation means for generating a prompt to a second large-scale language model that includes a reference to the shared information storage means; second prompt output means for outputting a prompt to the second large-scale language model to the second large-scale language model; a second chat message output means for outputting to the terminal a chat message based on a response from the second large-scale language model in response to a prompt to the second large-scale language model; A chat control system including:

2. the first prompt output means outputs a prompt to each of the first large-scale language models; the first chat message output means outputs, for each of the plurality of first large-scale language models, a chat message based on a response from the first large-scale language model to the terminal; the shared information generating means generates a plurality of pieces of shared information and stores the generated plurality of pieces of shared information in the shared information storage means; the second prompt generation means generates a prompt for the second large-scale language model, the prompt including a reference to the shared information storage means that stores a plurality of pieces of shared information; The chat control system of claim 1 .

3. a missing information generation request receiving means for receiving a request to generate missing information from the second large-scale language model; the shared information generating means, when receiving the generation request, generates the missing information and stores the generated missing information in the shared information storage means as shared information; The chat control system of claim 1 .

4. a third chat message output means for outputting a chat message requesting the missing information to the terminal when the generation request is accepted; the shared information generating means generates the missing information based on a chat message received from the terminal in response to a chat message requesting the missing information. The chat control system according to claim 3 .

5. a third prompt output means for outputting a prompt to instruct a large-scale language model different from the second large-scale language model to generate the missing information when the generation request is received; the shared information generating means generates the missing information based on a response to a prompt output from a large-scale language model that instructs generation of the missing information. The chat control system according to claim 3 .

6. the second chat message output means outputs to the terminal a chat message based on a response from the first large-scale language model to a prompt to the first large-scale language model and a response from the second large-scale language model to a prompt to the second large-scale language model. The chat control system of claim 1 .

7. generating prompts to a first large-scale language model based on chat messages received from the terminal; outputting a prompt to the first large-scale language model to the first large-scale language model; outputting to the terminal a chat message based on a response from the first large-scale language model to a prompt to the first large-scale language model; generating shared information based on at least one of chat messages received from the terminal and responses received from the first large-scale language model, and storing the generated shared information in shared information storage means; generating a prompt to a second large scale language model that includes a reference to the shared information store; outputting a prompt to the second large-scale language model to the second large-scale language model; outputting to the terminal a chat message based on a response from the second large-scale language model to a prompt to the second large-scale language model; A chat control method including:

8. generating prompts to a first large-scale language model based on chat messages received from the terminal; outputting a prompt to the first large-scale language model to the first large-scale language model; outputting to the terminal a chat message based on a response from the first large-scale language model to a prompt to the first large-scale language model; generating shared information based on at least one of chat messages received from the terminal and responses received from the first large-scale language model, and storing the generated shared information in shared information storage means; generating a prompt to a second large scale language model that includes a reference to the shared information store; outputting a prompt to the second large-scale language model to the second large-scale language model; outputting to the terminal a chat message based on a response from the second large-scale language model to a prompt to the second large-scale language model; A program that causes a computer to execute the following.

Images