Program, information processing device, method, and system

Abstract

In an interactive session between a user and a generative AI model, the generative AI model makes uncertain inferences. [Solution] The program executes the steps of identifying dynamic information to be updated from the history of a first dialogue session, generating and storing text information for carrying over the content of the dynamic information to a second dialogue session, determining an application mode of the text information at the start of the second dialogue session, and controlling at least one of response generation and response output of a generative AI model in response to an input in the second dialogue session. The controlling step includes at least one of a first control that suppresses the output of the response content when the response content includes speculative information, and a second control that maintains consistency of the generation result by using the text information as an essential precondition.

Description

[Technical Field]

[0001] The present disclosure relates to a program, an information processing device, a method, and a system. [Background technology]

[0002] In recent years, technologies have been developed that enable personalized dialogue for each user across sessions in systems using dialogue-based generative AI models such as ChatGPT (registered trademark). For example, Patent Document 1 discloses a technology that automatically extracts and summarizes information such as a user's personal information or context from a dialogue history as a "personalization notepad," permanently stores the information in a database outside the generative AI model, and loads this information into memory at the start of the next dialogue, thereby enabling consistent dialogue. With this technology, if the generative AI model detects an inconsistency in the user's personal information during a dialogue, the previous information can be overwritten and corrected based on the user's own correction instructions or the AI's automatic judgment. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] US Patent Application Publication No. 2025 / 0200361 Summary of the Invention [Problem to be solved by the invention]

[0004] The technology disclosed in Patent Document 1 allows a generative AI model to learn naturally without explicit user instructions, enabling a conversation that reflects past conversation history when starting a new chat. However, on the other hand, there is a problem that if a user wants to have a conversation that does not take into account past conversation history and personal information, or if inappropriate personal information is unintentionally stored, it can be inconvenient and make the conversation less efficient.

[0005] The objective of the present disclosure is to suppress uncertain inferences made by a generative AI model during an interactive session between a user and the generative AI model. [Means for solving the problem]

[0006] In order to solve the above-mentioned problem, one embodiment of a program of the present disclosure is a program to be executed by a computer having a processor, and causes the processor to execute the following steps: identify, from the history of a first dialogue session between a user and a generative AI model, dynamic information that is updated as the dialogue in the first dialogue session progresses; generate and store text information for carrying over the content of the dynamic information to a second dialogue session based on the identified dynamic information; read the stored text information at the start of the second dialogue session and determine a manner in which the text information will be applied in the second dialogue session; accept input from the user in the second dialogue session; and control at least one of the generation and output of a response content to the accepted input by the generative AI model in accordance with the determined manner. The control step includes at least one of a first control that suppresses the output of the response content when information contained in the response content is speculative information that is not supported by at least one of the dialogue history and the text information, and a second control that maintains consistency of the generation results of the generative AI model under the same preconditions by making the text information a required precondition in the second dialogue session. [Effects of the Invention]

[0007] According to the present disclosure, it is possible to suppress uncertain inferences by a generative AI model during an interactive session between a user and the generative AI model. [Brief explanation of the drawings]

[0008] [Figure 1] 1 is a block diagram illustrating an example of an overall configuration of a system according to an embodiment of the present disclosure. [Figure 2]FIG. 2 is a block diagram illustrating an example hardware configuration of a user terminal according to an embodiment of the present disclosure. [Figure 3] FIG. 2 is a block diagram showing functional units realized by a control unit of the user terminal described above. [Figure 4] FIG. 2 is a block diagram illustrating an example hardware configuration of a server according to an embodiment of the present disclosure. [Figure 5] FIG. 2 is a block diagram showing functional units realized by a control unit of the server described above. [Figure 6] FIG. 2 is a diagram illustrating an example of a data structure of a dialogue history database. [Figure 7] FIG. 2 is a diagram illustrating an example of a data structure of a text information database. [Figure 8] FIG. 2 is a diagram illustrating an example of a data structure of an attribute information database. [Figure 9] 10 is a flowchart illustrating an example of the operation of the server described above. [Figure 10] 10 is a flowchart illustrating another example of the operation of the server. [Figure 11] FIG. 10 is a schematic diagram illustrating an example of a memory confirmation screen according to an embodiment of the present disclosure. [Figure 12] FIG. 10 is a schematic diagram showing an example of a hallucination suppression screen according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0009] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. In all drawings describing the embodiments, common components are designated by the same reference numerals, and repeated explanations will be omitted. Note that the following embodiments do not unduly limit the content of the present disclosure described in the claims. Furthermore, not all components shown in the embodiments are necessarily essential components of the present disclosure. Furthermore, each drawing is a schematic diagram and is not necessarily a precise illustration.

[0010] In the following description, a "processor" refers to one or more processors. A processor may be expressed as, for example, processing circuitry. The at least one processor is typically a microprocessor such as a CPU (Central Processing Unit), but may also be another type of processor such as a GPU (Graphics Processing Unit). The at least one processor may be single-core or multi-core. The at least one processor may also be a general-purpose processor or a special-purpose processor.

[0011] Furthermore, the at least one processor may be a processor in the broad sense, such as a hardware circuit (for example, a field-programmable gate array (FPGA) or an application-specific integrated circuit (ASIC)) that performs part or all of the processing.

[0012] In the following explanation, information that produces an output for an input may be described using expressions such as "xxx table," but this information may be data of any structure, or may be a learning model such as a neural network that produces an output for an input. Therefore, an "xxx table" may be referred to as "xxx information."

[0013] Furthermore, in the following description, the configuration of each table is an example, and one table may be divided into two or more tables, or all or part of two or more tables may be one table.

[0014] The program may be pre-installed in the information processing device described below, or may be stored on a recording medium (e.g., non-transitory) that can be read by the information processing device and then installed in the information processing device. The program may also be transmitted from a program distribution server to the information processing device and installed. In the following description, two or more programs may be realized as one program, or one program may be realized as two or more programs.

[0015] In the following description, identification information for various objects is used, but the identification information may be information that indicates a specific object, and the specific data is not limited to those in the embodiments. The identification information may be an identification number or an identifier including alphabets or codes.

[0016] [Embodiment] <1. Overview and definition of terms> The system of this embodiment allows a dialogue between a user and a generative AI model to persist the dialogue context and rules (dynamic information) as text information even after the session ends, and by loading this information in the next session, it is possible for the model to behave as if it were a partner with a memory.

[0017] The definitions and specific examples of major terms used in this specification are as follows.

[0018] "Dynamic information" refers to information that is updated (including generated and changed) as the dialogue between the user and the generative AI model progresses. Specifically, dynamic information includes volatile information (context) that may disappear when the dialogue session ends, temporary agreements formed through the dialogue, or the progress of work. Specific examples include dialogue restrictions in the first dialogue session (e.g., rules such as "no casual expressions" and "use honorific language"), the progress of the first dialogue session (e.g., status such as "requirements definition phase completed" and "design phase next"), tasks to be performed in the second dialogue session and beyond (e.g., schedule such as "review the deliverables next"), and interests expressed by the user during the dialogue (e.g., context such as "security requirements are emphasized").

[0019] Here, the "first dialogue session" refers to a past session in which a dialogue took place between a user and a generative AI model, and from which dynamic information is identified and extracted from the dialogue history. The "second dialogue session" refers to a session that is newly started after the first dialogue session, in which text information inherited from the first dialogue session is read and applied as a prerequisite for generating a response from the generative AI model.

[0020] "Text information" refers to information in a format that is generated to carry over dynamic information identified in a first interactive session to a second interactive session and is stored in a permanent memory area (such as storage). Text information includes, for example, classification information, operation definition information, status record information, etc.

[0021] Here, "classification information" refers to information that organizes the items determined in the first dialogue session that will be required in the second dialogue session and thereafter into categories corresponding to the dialogue constraints in the first dialogue session, the progress of the first dialogue session, and the tasks to be performed in the second dialogue session and thereafter. In other words, when classification information is included in the text information, at least one of constraints, progress, or tasks is identified as dynamic information. Furthermore, "operation definition information" refers to invariant rules that include invariant rules that apply across multiple dialogue sessions. Furthermore, "status record information" refers to information that records at least one of the user's work status, the subject of the user's interest, or the progress of the first dialogue session at a specific point in time in the dynamic information.

[0022] "Attribute information" refers to fixed or static information about a user that is stored in the system in advance regardless of the progress of the dialogue. Specifically, attribute information is a user profile, basic settings, or prerequisite knowledge that does not change over time. Specific examples include a user's work history, organization, basic profile, language preference, and job role.

[0023] "Reference information" refers to a group of information identified in the second dialogue session as information that the generative AI model should refer to in order to generate a response. Specifically, reference information is information including text information and attribute information stored in the system, and is distinct from the vast history of past dialogues.

[0024] The term "first control" refers to a control that suppresses the output of a response content when the response content to be generated by the generative AI model contains information (speculative information) that is not supported by at least one of the dialogue history and text information from the first dialogue session. Here, "speculative information" refers to information containing a token (a component of the response content) generated by the generative AI model when the confidence level of the token falls below a predetermined threshold. Alternatively, it refers to information containing a token (a predetermined keyword) indicating insufficient information when the response content of the generative AI model contains the token. Furthermore, "suppressing the output of the response content" refers to restricting, blocking, or replacing the transmission or display of all or part of the response content to a user terminal with alternative content to prevent the user from perceiving the response content generated by the generative AI model as a definite fact. Specific examples of the first control include a process of outputting "unknown" information ("I don't know") instead of the response content, indicating that the information contained in the response content is unknown, and a process of outputting a confirmation request to the user to confirm the speculative information. Other examples include a process of presenting multiple options as speculative information and at least one of perspectives that serve as a basis for the user to judge the validity of each of the multiple options.

[0025] "Second control" refers to control that maintains consistency by providing text information to the generative AI model as an essential precondition (context) in the second dialogue session, so that it always responds using the same judgment criteria under the same preconditions. A specific example of second control is a process that forcibly and preferentially loads operation definition information and status record information into the generative AI model as system prompts, and restricts the generative AI model from making arbitrary interpretations or character changes.

[0026] <2. System Configuration> Figure 1 is a block diagram showing an example of the overall configuration of a system 1 according to this embodiment. The system 1 is a server-client type system including a user terminal 10, a server 20, and a generation AI system 30. The user terminal 10, the server 20, and the generation AI system 30 are connected to each other via a network 80 such as the Internet or a LAN (Local Area Network) using a protocol such as TCP / IP so as to be able to communicate with each other.

[0027] The user terminal 10 is a client device (computer, information processing device) operated by a user, and is realized by a PC (Personal Computer), a tablet terminal, a smartphone, or the like. The user may be anyone who wishes to use the interactive service provided by the system 1, such as a company employee, a manager, or an individual user. Note that while FIG. 1 shows an example in which the system 1 includes one user terminal 10, the system 1 may include multiple user terminals 10.

[0028] The server 20 is an information processing device that forms the core of the system 1 and is an example of a computer and information processing device according to one aspect of the present disclosure. The server 20 is typically a cloud server composed of one or more computers, and executes various processes according to this embodiment. Specifically, for example, the server 20 manages an interaction session between a user and a generative AI model, identifies dynamic information from the history of the interaction in the interaction session (hereinafter abbreviated as "dialogue history"), generates and stores text information, and controls responses using the generative AI model. The server 20 executes processes in cooperation with the generative AI system 30 in response to requests from the user terminal 10.

[0029] The generative AI system 30 includes one or more generative AI models and provides the use of the one or more generative AI models. In the present disclosure, a case will be described in which the generative AI system 30 provides the use of large-scale language models (LLMs), which are one type of generative AI model. A large-scale language model is a natural language model designed to perform multiple tasks in natural language processing. A large-scale language model is an example of a trained model, trained using a large number of parameters (e.g., billions to hundreds of billions) and high-level computational resources. A large-scale language model is a computer program or algorithm designed to perform natural language processing tasks. For example, natural language processing involves processes such as morphological analysis, syntactic analysis, information extraction, and sentence generation, allowing a computer to analyze language used by humans (i.e., natural language) and perform predetermined processing. When a prompt is input, the large-scale language model generates an output based on the text data, image data, etc. of the prompt. The prompt can be specified in natural language.

[0030] Examples of large-scale language models include the GPT (registered trademark) series (Generative Pre-Trained Transformer) developed by OPEN AI, StableLM developed by Stability AI, Llama2 developed by Meta, Palm2 (registered trademark) and LamDA2 (registered trademark) developed by Google, etc. Note that the generative AI model is not limited to a large-scale language model, and may be, for example, a multimodal generative AI model capable of recognizing image data and audio data.

[0031] A "prompt" refers to all input data input to a generative AI model, including instruction information that instructs the generative AI model to process information. A prompt may include, for example, text data, character strings, still image data, video data, audio data, etc. A prompt may also include gestures, reference data, etc. A user instructs a generative AI model to process information by inputting a prompt to the generative AI model. A prompt may include, for example, information for causing the generative AI model to output a desired output result.

[0032] The prompt is input, for example, by an input operation from the user via the user terminal 10. The prompt may also be input by uploading a document file, still image data, video data, audio data, or the like.

[0033] The prompt may include information for specifying an output format. Examples of information for specifying an output format include function definition information in a function call, a JSON schema, an XML format specification, and other structured data format specifications. This information allows output from the generative AI model to be obtained in a predetermined structured format. For example, function definition information may include a function name, parameter names, parameter types, parameter descriptions, etc. The generative AI model generates output results in the specified format based on this information specifying the output format. This function is particularly useful when structuring dynamic information as text information (such as state record information).

[0034] As mentioned above, prompts contain instruction information. Instruction information includes instruction sentences (text data) as well as still data, video data, audio data, etc. that indicate the content of the instruction. Reference data is data that the generative AI model uses as reference (learns) when performing information processing.

[0035] The prompt may include reference data as described above. The instruction information and reference data may be included in the prompt in any form. For example, the instruction information and reference data may be included in the prompt while being included in text data (instructions) input in response to a user operation. As another example, the instruction information may be included in the prompt as text data (instructions) input in response to a user operation, and the reference data may be included in the prompt as a file separate from the text data. Furthermore, the instruction information and reference data for the same instruction content do not necessarily need to be input to the generative AI model as a single prompt. For example, a prompt including instruction information may be input to the generative AI model first, and then a prompt including reference data may be input to the generative AI model.

[0036] In this embodiment, the generative AI system 30 is described as an information processing device separate from the server 20, but this is not limited to this. The functions of the generative AI system 30 may be implemented inside the server 20 (for example, as part of a program executed by the control unit 201, or using hardware resources such as a GPU provided in the server 20). In this case, the generative AI system 30 constitutes part of the server 20. In other words, the system 1 is configured without the generative AI system 30.

[0037] <3. User terminal configuration> Fig. 2 is a block diagram showing an example of the hardware configuration of the user terminal 10 according to this embodiment. As shown in Fig. 2, the user terminal 10 includes a control unit 101, a storage unit 102, a communication unit 103, an input unit 104, and an output unit 105. The blocks included in the user terminal 10 are electrically connected by, for example, a bus or the like.

[0038] The control unit 101 executes various programs stored in the storage unit 102 to perform various processes. The control unit 101 is, for example, a processor such as a CPU. The storage unit 102 includes a main storage device and an auxiliary storage device, and stores various programs and various information. The communication unit 103 performs processes such as modulation and demodulation for communication between the user terminal 10 and an external device (e.g., server 20).

[0039] The input unit 104 accepts instructions or information input by the user. The input unit 104 may be realized, for example, by a touch-sensitive device in which instructions, etc. are input by touching the operation surface. If the user terminal 10 is a PC or the like, the input unit 104 may be realized by a reader, keyboard, mouse, etc. The input unit 104 converts instructions, etc. input by the user into electrical signals and outputs them to the control unit 101. Note that the input unit 104 may include, for example, a receiving port that accepts electrical signals input from an external input device. The input unit 104 may also include a microphone that accepts audio input.

[0040] The output unit 105 presents information to the user. The output unit 105 is realized by, for example, a display. The display displays various information according to the control of the control unit 101. The display is realized by, for example, an LCD (Liquid Crystal Display) or an organic EL (Electro-Luminescence) display. The output unit 105 may include, for example, an output port that outputs an electrical signal to an external output device. The output unit 105 may include a speaker that outputs sound. In other words, presentation includes display on the display unit and output to a speaker or other output device.

[0041] Fig. 3 is a block diagram showing functional units realized by the control unit 101. As shown in Fig. 3, the control unit 101 includes, as functional units, an operation reception unit 131, a transmission / reception unit 132, and a presentation control unit 133. Specifically, the control unit 101 reads an application program stored in the storage unit 102 and executes instructions included in the application program to realize each functional unit.

[0042] The operation receiving unit 131 performs processing for receiving instructions or information input from the input unit 104. Specifically, for example, the operation receiving unit 131 receives, via the input unit 104, operations such as input of a chat message by a user, an instruction to store generated text information (classified information), or selection of priority information (selection of an option) when information conflicts.

[0043] The transmitting / receiving unit 132 transmits and receives data to and from the server 20 via the communication unit 103. Specifically, for example, the transmitting / receiving unit 132 transmits an input message or instruction to the server 20, and receives a response from the generation AI model, generated text information, confirmation options, or the like from the server 20.

[0044] The presentation control unit 133 controls the output unit 105 to present various information to the user. Specifically, for example, the presentation control unit 133 causes the display to display a response message from the generation AI model, a confirmation screen asking whether or not to memorize (see FIG. 11), or a hallucination suppression screen (see FIG. 12) including "multiple options that are candidates for inferred information" and "perspectives that serve as judgment criteria" generated by the first control.

[0045] <4. Server configuration> Fig. 4 is a block diagram showing an example of the hardware configuration of the server 20 according to this embodiment. As shown in Fig. 4, the server 20 includes a control unit 201, a storage unit 202, a communication unit 203, and an input / output IF 204. The blocks included in the server 20 are electrically connected by, for example, a bus or the like.

[0046] The control unit 201 executes various processes by executing various programs stored in the storage unit 202. The control unit 201 is, for example, a processor such as a CPU.

[0047] The storage unit 202 includes a main storage device and an auxiliary storage device, and stores various programs and various information. In this embodiment, the storage unit 202 stores, for example, a dialogue history database 221, a text information database 222, and an attribute information database 223. Details of these databases will be described later.

[0048] The communication unit 203 performs processes such as modulation and demodulation for the server 20 to communicate with external devices (e.g., the user terminal 10, the generation AI system 30). The input / output IF 204 functions as an interface for an input device (not shown) for receiving input operations from the manager / operator of the server 20, and an output device (not shown) for outputting information to the manager / operator.

[0049] 5 is a block diagram showing functional units realized by the control unit 201. The control unit 201 includes, as functional units, an identification unit 211, a generation unit 212, a determination unit 213, a reception unit 214, and a dialogue control unit 215. Specifically, the control unit 201 realizes each functional unit by reading an application program stored in the storage unit 202 and executing instructions included in the application program.

[0050] The identification unit 211 identifies dynamic information that is updated as the dialogue in the first dialogue session progresses from the dialogue history of the first dialogue session (the log recorded in the dialogue history database 221). As an example, the identification unit 211 uses natural language processing or the like to extract and identify, from the dialogue in the first dialogue session, constraints on the dialogue in the first dialogue session, the progress of the first dialogue session, or tasks to be performed in the second dialogue session or later.

[0051] Based on the identified dynamic information, the generation unit 212 generates text information for carrying over the content of the dynamic information to the second dialogue session. As one example, the generation unit 212 extracts operation definition information and status record information from the dynamic information, and then generates text information as structured data including the operation definition information and the status record information. As another example, the generation unit 212 generates classification information to be presented to the user as text information. The generated text information is stored in the text information database 222.

[0052] At the start of the second dialogue session, the determination unit 213 reads the text information and attribute information stored in the corresponding database in the storage unit 202, and determines a manner in which the text information and attribute information are to be applied in the second dialogue session. As an example of the "applying manner," the determination unit 213 identifies the read information as reference information and performs initialization settings so that only the reference information is input to the generative AI model. As another example of the "applying manner," when there is a difference between the content of a new instruction from the user and the read text information, the determination unit 213 performs settings to prioritize the content of the new instruction. Here, the "new instruction from the user" refers to an instruction input by the user at the start of the second dialogue session.

[0053] The receiving unit 214 receives input from the user in an interaction session with the generative AI model (including both the first interaction session and the second interaction session) via the communication unit 203.

[0054] The dialogue control unit 215 controls at least one of the generation and output of a response content by the generative AI model according to the determined mode. Specifically, the dialogue control unit 215 inputs reference information (text information, etc.) as context to the generative AI system 30 and causes the generative AI model to generate a response. Then, the dialogue control unit 215 executes at least one of a first control that suppresses the output of the generated response content if the generated response content includes inferred information, and a second control that maintains consistency of the generation results of the generative AI model by using the loaded text information as an essential precondition.

[0055] <5. Data Structure> The data structures of various databases stored in the server 20 will be described with reference to Figures 6 to 8. The databases shown in Figures 6 to 8 refer to relational databases, which are used to manage data sets called tables in tabular format, structurally defined by rows and columns, in association with one another. In a database, a table is called a table, a column in a table is called a column, and a row in a table is called a record. In a relational database, relationships between tables can be set and associated.

[0056] Typically, a column that serves as a primary key for uniquely identifying a record is set in the tables that make up each database, but setting a primary key in a column is not essential. The control unit 201 can cause the processor to add, delete, or update records in specific tables stored in the storage unit 202 according to various programs.

[0057] 6 to 8 are merely examples and do not exclude data that is not listed. Furthermore, even if data is listed in the same database, it may be stored in separate storage areas in the storage unit 202.

[0058] 6 is a diagram showing an example of the data structure of the dialogue history database 221. The dialogue history database 221 manages various information related to the dialogue history of the first dialogue session between the user and the generative AI model. The dialogue history database 221 includes columns for, for example, a log ID, a session ID, a user ID, a timestamp, a speaker, and a message content.

[0059] The "log ID" column stores an identifier (log ID) for uniquely identifying the dialogue history of each first dialogue session. The "session ID" column stores an identifier (session ID) for uniquely identifying each first dialogue session. The "user ID" column stores an identifier (user ID) for uniquely identifying the user who engaged in dialogue in each first dialogue session. The "timestamp" column stores the date and time when a message was sent or generated in each first dialogue session. The "speaker" column stores information indicating whether the sender of the message in each first dialogue session is the user or the generative AI model. The "message content" column stores the body of the message (text data) exchanged during the dialogue in each first dialogue session. The identification unit 211, for example, analyzes the content accumulated in the "message content" column to identify dynamic information.

[0060] 7 is a diagram showing an example of the data structure of the text information database 222. The text information database 222 functions as a summary information table, structuring dynamic information that should be carried over across multiple interaction sessions and managing it in text format. In other words, the text information database 222 manages various information related to text information generated based on dynamic information. The text information database 222 includes columns for, for example, file ID, user ID, source session ID, file type, file contents, and update date and time.

[0061] The "File ID" column stores an identifier (file ID) for uniquely identifying each piece of text information. The "User ID" column stores an identifier (user ID) for uniquely identifying the user who conducted the dialogue in the first dialogue session that was the source of each piece of text information. The "Source Session ID" column stores an identifier (source session ID) for uniquely identifying the first dialogue session that was the source of each piece of text information. The "File Type" column stores information for uniquely identifying each piece of text information as either operation definition information, status record information, or a transfer package. The "File Content" column stores each piece of text information itself. For example, the "File Content" column stores text information such as a set of rules written in Markdown format or state transition data written in JSON format. The "Update Date and Time" column stores the date and time when each piece of text information was last created or updated.

[0062] 8 is a diagram showing an example of the data structure of the attribute information database 223. The attribute information database 223 functions as a user fixed information table and manages fixed data that does not change frequently even as the dialogue progresses in a dialogue session, i.e., user attribute information. The attribute information database 223 includes columns for, for example, user ID, name, work history, affiliated organization, used language, and basic profile.

[0063] The "User ID" column stores an identifier (user ID) for uniquely identifying a user who is the target of each attribute information. The "Name" column stores the name of a user who is the target of each attribute information. The "Work Experience" column stores the past work history (which may include skill sets) of a user who is the target of each attribute information. The "Organization" column stores information about the current organization to which a user who is the target of each attribute information belongs (e.g., basic information such as organization name and location). The "Language Used" column stores the main language (e.g., Japanese, English) used in an interactive session by a user who is the target of each attribute information. The "Basic Profile" column stores other basic information (e.g., age, gender, registered place of residence, family composition) of a user who is the target of each attribute information. The determination unit 213 reads this attribute information together with text information and uses it as reference information.

[0064] <6. Details of text information> In this embodiment, text information with a specific data structure is used to persist the dynamic information identified from the dialogue history of the first dialogue session and to reliably reproduce it in the second dialogue session and subsequent dialogue sessions. Specifically, the text information used includes operation definition information, status record information, and transfer packages.

[0065] Operation definition information is information that defines immutable rules, prohibitions, and allowances that apply across multiple sessions. In other words, operation definition information is information that persists the dynamic information that is established through dialogue. Specifically, operation definition information includes, for example, a definition section, a top-level rule section, a prohibition rule section, an allowance rule section, and a behavior principle section.

[0066] The definition section includes a statement declaring that the operational definition information is a rule to be used by System 1 and is not fictionalized text. The top-level rules section includes statements defining priorities, such as giving top priority to explicit user instructions and not providing general conversation assistance. The prohibited rules section includes statements enumerating prohibitions on the use of casual expressions, making arbitrary assumptions or assertions about the user's intentions, and describing implemented features that have not yet been implemented. The allowable rules section includes statements enumerating permitted behaviors, such as summarizing objective facts based on the dialogue history of the first dialogue session and generating deliverables (e.g., program code, documentation) with clear descriptions of the areas and reasons for corrections. The behavior principles section includes statements specifying that when information is insufficient, "unknown" should be stated or that no guesswork should be made. These sections establish the "personality" and "behavior" of the generative AI model and form the basis of the first control (suppressing guesswork) and the second control (maintaining consistency).

[0067] The state record information is a state snapshot that unambiguously indicates at least one of the following: the user's work status at a specific point in time, the user's interests, or the progress of the first dialogue session. The machine-readable JSON format, rather than natural language, is used to prevent inconsistent interpretation. Specifically, the state record information includes keys such as current_context, workstreams, constraints, and depi_rule. The current_context element has an object value that defines the top theme (e.g., design and review of the business support AI model), the "mode" (e.g., ideation, detailed design, or code review) indicating the functional role of the dialogue, and the "stage" (e.g., requirements definition phase, implementation phase, or testing phase) indicating the overall progress of the project. The workstreams element has an array value that manages ongoing work items, and each element includes a name, goal, next action, and deliverable. The constraints element has an object value that defines the current constraints, such as the "tone" (e.g., logical and concise, or thorough explanations for beginners) and prerequisites that specify the response style of the generative AI model. depi_rule has as its value an object that defines rules such as "dependency relationships" that indicate order constraints between tasks (e.g., task B cannot be started without completing task A), and "required output sets" that indicate combinations of information that must be included when outputting a response (e.g., supporting literature and risk assessments must be included when outputting). These keys ensure that the "stage at which" the first dialogue session ended is accurately carried over, preventing the loss of context.

[0068] A transfer package is information that contains meta-information for continuing an interactive session by referencing operation definition information and status record information as the original. Specifically, a transfer package does not include operation definition information or status record information, but includes application declarations that specify the original as the reference source, prohibit guessing and completion, and prohibit the creation of semantic duplication. In addition, a transfer package also includes a session continuation procedure that, when a new session starts, presents the operation definition information, status record information, and this package in that order to treat them as a continuous dialogue between the same subject.

[0069] <7. Example of operation> An example of the operation of the server 20 will be described below with reference to Fig. 9. Fig. 9 is a flowchart showing an example of the operation of the server 20 in the information perpetuation phase.

[0070] First, in step S101, the identification unit 211 of the server 20 identifies dynamic information from the dialogue history of the first dialogue session. Specifically, for example, the identification unit 211 reads log data (dialogue history) of the first dialogue session in which the user engaged in dialogue from the dialogue history database 221, triggered by a session termination instruction or an explicit storage instruction from the user. Then, the identification unit 211 analyzes text data included in the read log data and extracts constraints, progress, tasks, and the like that have newly emerged during the dialogue process. At this time, the identification unit 211 may, for example, send a prompt to the generative AI system 30 (generative AI model) instructing it to summarize, and use a response from the generative AI model for extraction. Also, for example, the identification unit 211 may perform keyword extraction using natural language processing.

[0071] Next, in step S102, the generation unit 212 of the server 20 generates text information for the second dialogue session based on the identified dynamic information. Specifically, for example, the generation unit 212 formats the dynamic information, including the extracted constraints, progress status, tasks, etc., in accordance with the data structures of the operation definition information and the status record information. At this time, the generation unit 212 structures the data according to a predetermined JSON schema or Markdown notation so that the generation AI model can easily interpret it.

[0072] Next, in step S103, the generation unit 212 of the server 20 determines whether there is a contradiction between the content of the generated text information and information already stored in the database (existing information). Specifically, for example, the generation unit 212 compares each item of the generated text information with the attribute information stored in the attribute information database 223 and with past text information already stored in the text information database 222 (text information stored in past dialogue sessions). For example, if the setting for "language used" is "Java" in the attribute information but "Python" in the generated text information, the generation unit 212 determines that there is a contradiction.

[0073] If the server 20 detects that there is a contradiction (Yes in step S103), the process proceeds to step S104. On the other hand, if the server 20 determines that there is no contradiction (No in step S103), the process proceeds to step S105.

[0074] In step S104, the generation unit 212 of the server 20 generates a plurality of options for resolving the contradiction and presents them to the user. Specifically, for example, the generation unit 212 generates a plurality of option data including actions such as "keep the existing information," "overwrite with new information," or "store both as separate information," and transmits the generated data to the user terminal 10. Then, the generation unit 212 controls the presentation control unit 133 to display an option presentation screen (not shown) for when information conflicts occur on the display of the user terminal 10 based on the received plurality of option data. When the user selects one of the options displayed on the option presentation screen, the operation reception unit 131 receives the selection operation, and the transmission / reception unit 132 returns the selection result to the server 20. The reception unit 214 of the server 20 receives the selection result and determines the content to be stored.

[0075] In step S105, the generation unit 212 of the server 20 inquires of the user whether or not to memorize the generated (or confirmed) text information. Specifically, the generation unit 212 transmits a preview (classification information, etc.) of the generated text information to the user terminal 10. Then, the generation unit 212 controls the presentation control unit 133 to display a memorization confirmation screen 1100 as shown in FIG. 11 on the display of the user terminal 10.

[0076] 11 is an example of a text information memorization confirmation screen 1100. The memorization confirmation screen 1100 is provided with a summary display area 1101 that displays a summary of the identified dynamic information (e.g., an extracted bulleted list along with a message saying "I will memorize the following rules and tasks"), a memorize button 1102, and a discard button 1103. The user checks the displayed content and presses the memorize button 1102 if they want to carry it over to the second dialogue session. On the other hand, if they do not want to carry it over, they press the discard button 1103.

[0077] Finally, in step S106, the generation unit 212 of the server 20 accepts a storage instruction from the user and stores the generated text information in the text information database 222. Specifically, for example, the generation unit 212 associates the confirmed text data with the user ID, source session ID, file type, and update date and time, and stores (INSERT) the data as a new record in each column of the text information database 222, or updates (UPDATE) an existing record.

[0078] Another example of the operation of the server 20 will be described below with reference to Fig. 10. Fig. 10 is a flowchart showing an example of the operation of the server 20 in the information handover and response control phase.

[0079] First, in step S201, when the determination unit 213 of the server 20 detects the start of a second dialogue session (new chat), it reads necessary information from each database. Specifically, for example, the determination unit 213 reads the latest text information (e.g., operation definition information, status record information) associated with the user ID of the user who will conduct the second dialogue session from the text information database 222. In addition, the determination unit 213 reads attribute information associated with the user ID of the user from the attribute information database 223.

[0080] Next, in step S202, the determination unit 213 of the server 20 identifies the read text information and attribute information as reference information, and then determines how to apply the reference information to the second dialogue session. Specifically, for example, the determination unit 213 sets the generative AI model to input only structured text information, rather than the entire vast amount of log data (dialogue history) in the dialogue history database 221. This makes it possible to accurately inherit only important context while suppressing token consumption in the context window.

[0081] Next, in step S203, if there is a start instruction (initial prompt) from the user, the determination unit 213 of the server 20 compares the content of the instruction with the read text information. If there is a difference between the two (for example, if the stored rule specifies "answer in English" but the current instruction specifies "answer in Japanese"), the determination unit 213 determines that the instruction from the user should be given priority.

[0082] Next, in step S204, the dialogue control unit 215 of the server 20 generates an initialization prompt for initializing the generative AI model based on the determined aspect and inputs it to the generative AI model. Specifically, for example, the dialogue control unit 215 inputs an initialization prompt with the following configuration to the API of the generative AI model. That is, the dialogue control unit 215 includes the contents of the operation definition information in the initialization prompt as a system message, along with an instruction such as "Please strictly observe the rules described in the following operation definition information." Furthermore, the dialogue control unit 215 includes the contents of the state record information in the initialization prompt as a user message or context, along with an explanation such as "The current state record information is as follows." As a result, the generative AI model enters a standby state with the second control (consistency maintenance) in effect.

[0083] Next, in step S205, the receiving unit 214 of the server 20 receives the user's input (question or instruction) transmitted from the user terminal 10.

[0084] Next, in step S206, the dialogue control unit 215 of the server 20 requests the generative AI model to generate a response content. Specifically, for example, the dialogue control unit 215 combines the reference information with text data indicating the input content from the user, generates a prompt with a control instruction (instruction) added to the end, such as "Please respond based on the above information without making any guesses," and sends the generated prompt to the generative AI system 30. The generative AI system 30 inputs the prompt received from the server 20 into the generative AI model and requests the generation of a response content. At this time, the dialogue control unit 215 may add an instruction to the prompt, such as "If information is insufficient, do not make a guess, but output 'insufficient information'," as internal processing for the first control.

[0085] Next, in step S207, the dialogue control unit 215 of the server 20 determines whether the response content from the generative AI model includes inferred information. Specifically, for example, the dialogue control unit 215 analyzes the response content to determine whether the response content from the generative AI model received from the generative AI system 30 includes the keyword "insufficient information," or whether the confidence level of the generated token is below a predetermined threshold.

[0086] If it is determined that inference information is included (Yes in step S207), the server 20 proceeds to step S208. On the other hand, if it is determined that inference information is not included (No in step S207), the server 20 proceeds to step S209.

[0087] In step S208, as a first control, the dialogue control unit 215 of the server 20 suppresses output of a definitive answer and instead presents multiple options to the user. Specifically, for example, the dialogue control unit 215 generates an additional prompt (prompt chain) with the following content: "Information necessary to answer the user's question is insufficient. Instead of providing a definitive answer, please provide multiple possible options to fill in the missing information, the prerequisites for adopting each option, the basis for generating each option, and a perspective for judging the validity of each option." The dialogue control unit 215 then transmits the additional prompt (prompt chain) to the generative AI system 30. The generative AI system 30 inputs the additional prompt received from the server 20 into the generative AI model to generate multiple options, the prerequisites for each option, the basis for generating each option, and the perspective for judging each option. The dialogue control unit 215 formats the multiple options, the prerequisites for each option, the basis for generating each option, and the perspective for judging each option received from the generative AI system 30, and transmits them to the user terminal 10. Then, the dialogue control unit 215 controls the presentation control unit 133 to display a hallucination suppression screen 1200 as shown in FIG.

[0088] In step S209, the dialogue control unit 215 of the server 20 presents the generated response content (definitive answer) to the user as is. Specifically, for example, the dialogue control unit 215 transmits the generated response content as is to the user terminal 10, and controls the presentation control unit 133 to display it on the display. In this way, only evidence-based and reliable answers are provided to the user.

[0089] <8. Details of the first control: Presenting options and perspectives> A characteristic control of this embodiment is the first control, which prevents the generative AI model from filling in uncertain information with guesswork and providing answers. The example screen in Figure 12 is a hallucination suppression screen 1200 that shows the user interface for executing this first control.

[0090] For example, consider a scenario in which a user asks, "What are the risks of this project?" In this case, if the reference information contains only an overview of the project and lacks information to identify the specific risks, a typical generative AI model will speculate and list general risks (the risk of hallucination). In response, instead of causing the generative AI model in the generative AI system 30 to generate a single, definitive answer, the dialogue control unit 215 controls the presentation control unit 133 to display a hallucination suppression screen 1200 on the display of the user terminal 10. The hallucination suppression screen 1200 includes a message display area 1201, an option display area 1202, a generation basis display area 1203, a judgment viewpoint display area 1204, and a send button 1205.

[0091] In the message display area 1201, a message from the generative AI model is displayed, such as, "In order to accurately understand the intent of your question, please select which of the following perspectives of risk you are considering."

[0092] The option display area 1202 displays, for example, a first option, a second option, and a third option (e.g., all in the form of radio buttons). For example, "risk of schedule delay" is displayed as the first option, and a condition such as "if collaboration with external vendors is included" is added as its prerequisite. Furthermore, "risk of budget overrun" is displayed as the second option, and a condition such as "if additional development occurs" is added as its prerequisite. Furthermore, "risk of technical debt" is displayed as the third option, and a condition such as "if a legacy system is used" is added as its prerequisite. Note that the first to third options may be accompanied by a description of the prerequisites under which the option is valid.

[0093] The generation basis display area 1203 displays the generation basis indicating why the generative AI model selected multiple options as candidates for inferred information. Specifically, for example, information such as "Basis: Based on the description of 'workstreams' in the operation definition information" is displayed.

[0094] The judgment viewpoint display area 1204 displays a judgment viewpoint (e.g., "Please check the current resource situation") that serves as a criterion for the user to judge the validity of multiple options. Furthermore, the judgment viewpoint display area 1204 also displays a generation basis indicating why the generative AI model presented that judgment viewpoint. For example, information such as "Rationale for generating viewpoint: The current phase is "detailed design," and the fact that resource allocation is undetermined could be a risk factor" is added as the generation basis.

[0095] <9.Summary> According to the present disclosure, dynamic information formed through the dialogue in the first dialogue session is persisted as text information in an external file, etc., and applied as a required precondition at the start of the second dialogue session, thereby enabling a consistent dialogue environment to be reproduced across multiple sessions. Furthermore, the first control that suppresses the output of inferred information and the second control that fixes the preconditions make it possible to suppress unfounded inferences (hallucination) and pandering to context by the generative AI model, thereby improving the reliability and reproducibility of the dialogue.

[0096] More specifically, according to this embodiment, text information is generated and stored based on dynamic information (e.g., constraints, progress, tasks) identified from the history of the first dialogue session. At the start of the second dialogue session, the text information is read and identified as reference information. Then, only the identified reference information is input into the generative AI model, and initialization is performed with limited reference to dialogue history other than the reference information. This allows for accurate inheritance of only important context (rules, states) established in previous dialogues with the user, even when the dialogue session changes, thereby enabling efficient and highly accurate dialogue that eliminates noise caused by irrelevant dialogue history. Furthermore, by introducing the first and second controls, the generative AI model can be prevented from arbitrarily filling in uncertain information or changing its judgment due to fluctuations from session to session, thereby providing stable response quality that can withstand business use.

[0097] [Modification] <1. First Modification: Text Information Generation Format> The generation unit 212 may store the generated text information in the text information database 222 in a format that is easy for the user to confirm, rather than storing the generated text information as is. Specifically, for example, the generation unit 212 analyzes dynamic information using natural language processing techniques (e.g., morphological analysis, syntactic analysis, and summarization using a large-scale language model) to generate classified information that is categorized and organized into the categories of "constraints (rules)," "progress," and "tasks." The generation unit 212 then controls the presentation control unit 133 to display the generated classified information on the display of the user terminal 10, requesting user confirmation. This allows the user to intuitively understand what the generative AI model is trying to memorize.

[0098] Furthermore, as in the above-described embodiment, it is desirable that the text information be generated and managed by physically or logically separating it into operation definition information containing immutable rules and status record information containing variable states. This allows rule changes (modification of operation definition information) and status updates (updating status record information) to be performed independently, improving the efficiency of data management.

[0099] <2. Second Variation: Determining Priority in Case of Information Conflict> At the start of the second dialogue session, the determination unit 213 may read out the text information stored in the text information database 222 and compare it with the content of a new instruction (prompt: hereinafter, "new instruction") input by the user. For this comparison, for example, a similarity determination or difference extraction algorithm between the two is used. If there is a difference between the two (for example, the text information is "polite language" but the content of the new instruction is "informal speech"), the determination unit 213 determines to give priority to and apply the new instruction, which is the user's most recent intention.

[0100] Furthermore, when storing text information or determining the mode, if the generation unit 212 detects a contradiction between the identified dynamic information and already stored information (attribute information or past text information), the generation unit 212 performs a process of leaving the user to make a decision. Specifically, for example, the generation unit 212 controls the presentation control unit 133 to display a screen (not shown) on the display of the user terminal 10 that displays a comparison of the contradictory items. Then, the generation unit 212 controls the presentation control unit 133 to display a warning message such as "An information contradiction has been detected" on the display. For example, the screen displays existing information (already stored information, e.g., used language = Java) and new information (dynamic information, e.g., used language = Python) side by side. Furthermore, for example, the screen displays multiple options to allow the user to select which information to prioritize when storing, such as "keep existing information," "overwrite with new information," and "store both as separate information (version management)." The user selects one of these multiple options using, for example, a radio button or the like. This will ensure that unintended overwriting or loss of information due to incorrect assumptions made by AI is prevented.

[0101] <3. Third Modification: Variation of Response Control> First, the dialogue control unit 215 may monitor the reliability score for the output of the generative AI model. Specifically, for example, the dialogue control unit 215 obtains, as the reliability score, the logarithmic probability calculated when the generative AI model generates each token that constitutes the response content in the second dialogue session from the generative AI model. The logarithmic probability is a negative value that indicates how likely a token is to be selected contextually.

[0102] Then, the dialogue control unit 215 normalizes the acquired logarithmic probability and converts it into a certainty factor such as a percentage (0 to 100%), and then judges the inferred information using the following criteria (a) or (b). (a) The confidence levels of all tokens included in the response are averaged, and if the average confidence level obtained is below a predetermined threshold (e.g., 80%), the entire response is determined to be uncertain. (a) Using natural language processing such as morphological analysis, important keywords (e.g., proper nouns, numbers, dates) in the response content are identified, and only the individual confidence scores of tokens corresponding to those keywords are extracted. If this individual confidence score falls below a predetermined threshold (e.g., 95%), even if the entire sentence is fluent, it is determined that unfounded speculation (hallucination) is included in important factual relationships.

[0103] If the generated response content is deemed to contain speculative information based on the above-mentioned determination (A) or (B), the dialogue control unit 215 discards the generated response content without presenting it to the user, and instead outputs unclear information (an error message) such as "An answer cannot be generated using only the provided information (reference information)." Alternatively, the dialogue control unit 215 gives the generative AI model an additional instruction (prompt) to "generate a counter-question to clarify the intent of the question," and generates and outputs to the user a confirmation request such as, for example, "Does 'project risk' refer to Project X or to a company-wide risk?" This prevents the spread of incorrect information and eliminates misunderstandings in the dialogue.

[0104] The method for determining inferred information is not limited to the above-mentioned (A) or (B). For example, if the search score in Retrieval-Augmented Generation (RAG) is equal to or lower than a threshold, or if the determination result by a fact checking tool does not support the accuracy of the information (e.g., if the determination result is "false" or "unable to determine"), the dialogue control unit 215 may determine that the information is inferred.

[0105] Furthermore, when the dialogue control unit 215 detects insufficient information, it may not simply reject the answer but may actively generate information to support the user's decision-making. Specifically, for example, the dialogue control unit 215 uses a technique such as Chain of Thought prompting to instruct the generative AI model, "Since a unique answer cannot be determined from the current context (reference information), comprehensively list all possible possibilities and create a scenario for each." Upon receiving this instruction, the generative AI model generates multiple probable options (scenarios) rather than a single, definitive response.

[0106] The dialogue control unit 215 then causes the generative AI model to analyze the axes along which the generated multiple options branch off, and generates viewpoints that will serve as criteria for the user to select an option. For example, in the case of a question about a schedule, the dialogue control unit 215 causes the generative AI model to generate viewpoints that are in a trade-off relationship, such as "should we prioritize cost?" or "should we prioritize quality?" Next, the dialogue control unit 215 controls the presentation control unit 133 to display the generated viewpoints in the judgment viewpoint display area 1204 of the hallucination suppression screen 1200.

[0107] Furthermore, the dialogue control unit 215 may perform processing to provide logical justification for the generated multiple options and viewpoints to enhance the user's sense of acceptance. For multiple options, the dialogue control unit 215 causes the generative AI model to generate preconditions necessary for each option to be the correct answer (for each option to be adopted). For example, for "first option: release next week," the dialogue control unit 215 causes the generative AI model to generate an IF condition such as "precondition: no additional testing is required and resources are secured." Next, the dialogue control unit 215 controls the presentation control unit 133 to display the generated IF condition together with the first option on the display of the user terminal 10.

[0108] For a viewpoint, the dialogue control unit 215 causes the generative AI model to generate a basis for generating the viewpoint (generative basis). Specifically, for example, the dialogue control unit 215 applies a citation function in RAG to cause the generative AI model to clearly indicate which description in the reference information was used to derive each option and viewpoint. For example, the dialogue control unit 215 causes the generative AI model to generate an annotation such as "Basis: Based on the constraint 'budget freeze' described in the constraints field of the operation definition information." This allows the user to confirm that the result is not randomly generated by the generative AI model, but is a logical inference result based on the saved context, allowing the user to select the next action with confidence.

[0109] 〔supplementary explanation〕 In the above-described embodiment, the units and means are implemented by a processor, but this is not limiting. The units and means may be any hardware known to execute the functions.

[0110] In the above-described embodiment, the server 20 is provided with each function, but the present invention is not limited to this. Some or all of the functions of the server 20 according to this embodiment may be provided in the user terminal 10 instead of the server 20.

[0111] Although several embodiments of the present disclosure have been described above, these embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. For example, a configuration or process in one embodiment may be combined with a configuration or process in another embodiment, or a modified version of one embodiment may be applied to another embodiment. These embodiments and their modifications are intended to be included within the scope and spirit of the invention, as well as within the scope of the invention and its equivalents as defined in the claims.

[0112] [Note] The matters described in the above embodiments will be supplemented below.

[0113] <Appendix 1> A program to be executed by a computer having a processor, the program causing the processor to execute the following steps: identify, from the history of the dialogue in a first dialogue session between a user and a generative AI model, dynamic information that is updated as the dialogue in the first dialogue session progresses; generate and store text information for carrying over the content of the dynamic information to a second dialogue session based on the identified dynamic information; read the stored text information at the start of the second dialogue session and determine a manner in which the text information will be applied in the second dialogue session; accept input from the user in the second dialogue session; and control at least one of the generation and output of a response content to the accepted input by the generative AI model in accordance with the determined manner, wherein the control step includes at least one of a first control that suppresses the output of the response content if information included in the response content is speculative information that is not supported by at least one of the history and the text information, and a second control that maintains consistency of the generation results of the generative AI model under the same preconditions by making the text information a required precondition in the second dialogue session.

[0114] <Appendix 2> In the controlling step, the first control determines that the information contained in the response content is inferred information if the confidence of the token generated by the generation AI model is below a predetermined threshold, or if the response content contains a predetermined keyword indicating a lack of information. (Appendix 1) A program described in.

[0115] <Appendix 3> A program described in (Appendix 1) or (Appendix 2), wherein in the identifying step, the dynamic information identified is at least one of constraints on the dialogue in the first dialogue session, progress of the first dialogue session, or tasks to be performed in the second dialogue session or later.

[0116] <Appendix 4> The storing step includes a step of generating classification information as text information, which organizes the items determined in the first dialogue session that will be required in the second dialogue session and thereafter into categories corresponding to constraints, progress, and tasks, and presenting the classification information to the user; and a step of storing the classification information when an instruction to store the classification information is received from the user. (Appendix 3) The program described in

[0117] <Appendix 5> The program according to any one of (Appendix 1) to (Appendix 4), wherein the determining step includes a step of comparing the content of an instruction input by the user at the start of the second dialogue session with stored text information, and a step of determining that the content of the instruction should take priority over the text information if there is a difference between the content of the input instruction and the stored text information.

[0118] <Appendix 6> The program described in any one of (Appendix 1) to (Appendix 5), wherein the determining step includes a step of further reading pre-stored user attribute information and identifying the read text information and the attribute information as reference information to be referenced in the second dialogue session, and a step of inputting only the identified reference information into the generative AI model and initializing the second dialogue session under a restriction that prevents the generative AI model from referring to any history other than the reference information.

[0119] <Appendix 7> A program described in any one of (Appendix 1) to (Appendix 6), wherein in the storing step, the program extracts, from the dynamic information, operational definition information including invariant rules that are applied across multiple dialogue sessions, and status record information that records at least one of the user's work status at a specific point in time in the dynamic information, the subject of the user's interest, or the progress of the first dialogue session, thereby generating the text information including the operational definition information and the status record information.

[0120] <Appendix 8> A program described in any of (Appendix 1) to (Appendix 7), wherein in the controlling step, the first control includes either a process of not outputting the response content, but outputting unknown information indicating that the information contained in the response content is unknown, or a process of outputting a confirmation request to the user to confirm the information contained in the response content.

[0121] <Appendix 9> In the controlling step, the first control includes a process of inputting at least a portion of the identified reference information into the generative AI model, instead of having the generative AI model generate a single response content that asserts the inferred information as a definite fact, causing the generative AI model to generate multiple options that are candidates for the inferred information and at least one of a viewpoint that serves as a standard for the user to judge the validity of each of the multiple options, and presenting at least one of the generated multiple options and viewpoints to the user. (Appendix 6)

[0122] <Appendix 10> In the controlling step, the first control includes, if the generative AI model is to generate a plurality of options, having the generative AI model further generate at least one of a prerequisite for adopting the option and a basis for generating the option, corresponding to each of the plurality of options, and further presenting at least one of the generated prerequisites and generation basis to the user; and, if the generative AI model is to generate a viewpoint, having the generative AI model further generate a basis for generating the viewpoint and further presenting the generated generation basis to the user. (Appendix 9) A program described in.

[0123] <Appendix 11> The program described in (Appendix 6) in which, if a contradiction is detected in the storing step or the determining step between the content of the dynamic information identified from the history and the attribute information that has already been stored or the text information that has been stored in a past dialogue session, the program presents the user with multiple options to allow the user to select which information to prioritize when storing the information.

[0124] <Appendix 12> An information processing device including a processor, the processor executing all steps in the program according to any one of (Supplementary Note 1) to (Supplementary Note 11).

[0125] <Appendix 13> A method executed by a computer having a processor, wherein the processor executes all steps in the program described in any one of (Appendix 1) to (Appendix 11).

[0126] <Appendix 14> A system comprising one or more processors that execute all steps in the program described in any one of (Appendix 1) to (Appendix 11). [Explanation of symbols]

[0127] 1. System 10...User terminal 20...Server 30...Generative AI system 80…Network 101...Control unit 102...Storage section 103…Communications Department 104...input section 105...Output section 131...Operation reception section 132...Transmitter / receiver 133...Presentation control unit 201...Control unit 202...Storage section 203…Communications Department 204...Input / output interface 211…Specific section 212...Generation section 213...Decision Section 214…Reception 215...Dialogue control unit 221...Dialogue history database 222...Text information database 223...Attribute information database 1100...Memory confirmation screen 1101…Summary display area 1102...Memory button 1103...Discard button 1200...Hallucination suppression screen 1201...Message display area 1202...Options display area 1203...Generation basis display area 1204: Judgment viewpoint display area 1205...Send button

Claims

1. A program to be executed by a computer having a processor, the processor, Identifying dynamic information that is updated as the dialogue progresses in a first dialogue session from a dialogue history of the first dialogue session between the user and the generative AI model; generating and storing text information based on the identified dynamic information for carrying over the content of the dynamic information to a second dialogue session; At the start of the second interactive session, reading the stored text information and determining how to apply the text information in the second interactive session; accepting input from the user in the second interactive session; Controlling at least one of generating and outputting a response content to the received input by the generative AI model according to the determined aspect; Execute The controlling step includes at least one of a first control that suppresses output of the response content when the information included in the response content is speculative information that is not supported by at least one of the history and the text information, and a second control that maintains consistency in the generation results of the generative AI model under the same preconditions by making the text information an essential precondition in the second dialogue session.

2. 2. The program according to claim 1, wherein in the controlling step, the first control determines that the information contained in the response content is the inferred information if the confidence of the token generated by the generative AI model is below a predetermined threshold or if the response content includes a predetermined keyword indicating a lack of information.

3. 2. The program according to claim 1, wherein in the identifying step, the dynamic information is identified to be at least one of constraints on the interaction in the first interaction session, a progress status of the first interaction session, or a task to be performed in the second interaction session or later.

4. The storing step includes: generating, as the text information, classification information in which items that will be required in the second dialogue session and thereafter among the contents determined in the first dialogue session are organized into categories corresponding to the constraints, the progress status, and the tasks, and presenting the classification information to the user; storing the classification information when an instruction to store the classification information is received from the user; The program according to claim 3 , comprising:

5. The determining step includes: comparing the content of an instruction input by the user at the start of the second interaction session with the stored text information; a step of determining, when there is a difference between the content of the input instruction and the stored text information, that the content of the instruction is to be prioritized over the text information; The program of claim 1 , comprising:

6. The determining step includes: further reading pre-stored attribute information of the user, and specifying the read text information and the attribute information as reference information to be referred to in the second interaction session; inputting only the identified reference information into the generative AI model, and initializing the second dialogue session under a restriction that prevents the generative AI model from referencing the history other than the reference information; The program of claim 1 , comprising:

7. 2. The program according to claim 1, wherein in the storing step, operation definition information including an invariant rule that is applied across multiple dialogue sessions and status record information that records at least one of the user's work status at a specific point in time of the dynamic information, an object in which the user is interested, or a progress status of the first dialogue session are extracted from the dynamic information, thereby generating the text information including the operation definition information and the status record information.

8. 2. The program of claim 1, wherein in the controlling step, the first control includes either a process of not outputting the response content and outputting unknown information indicating that the information contained in the response content is unknown, or a process of outputting a confirmation request to the user to confirm the information contained in the response content.

9. The program of claim 6, wherein in the controlling step, instead of having the generative AI model generate a single response content that asserts the inferred information as a definite fact, the first control inputs at least a portion of the identified reference information into the generative AI model, thereby causing the generative AI model to generate a plurality of options that are candidates for the inferred information and at least one of a viewpoint that serves as a basis for the user to judge the validity of each of the plurality of options, and presents the generated plurality of options and at least one of the viewpoints to the user.

10. In the controlling step, the first control In the case where the generative AI model is made to generate the plurality of options, the method further includes a process of making the generative AI model generate at least one of a prerequisite for adopting the option and a basis for generating the option, corresponding to each of the plurality of options, and further presenting at least one of the generated prerequisite and the basis for generating the option to the user; The program of claim 9, further comprising a process for causing the generative AI model to generate a basis for generating the viewpoint, and further presenting the generated basis for generating the viewpoint to the user, when the generative AI model is caused to generate the viewpoint.

11. 7. The program according to claim 6, wherein, if a contradiction is detected in the storing step or the determining step between the content of the dynamic information identified from the history and the attribute information that has already been stored or the text information that has been stored in a past dialogue session, the program presents the user with a plurality of options to allow the user to select which information to prioritize when storing the information.

12. 12. An information processing apparatus comprising a processor, the processor executing all steps of the program according to claim 1.

13. A method executed by a computer having a processor, the processor executing all the steps of the program of any one of claims 1 to 11.

14. A system comprising one or more processors that execute all steps in the program according to any one of claims 1 to 11.

Images