Proxy device for executing multiple tasks through large-scale language model
By using the agent device to handle user responses and task execution agents, multiple tasks can be automatically selected and executed, solving the problem of inaccurate responses to multitasking requests in existing LLMs and achieving more efficient multitasking processing.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-06
- Publication Date
- 2026-03-10
AI Technical Summary
Existing large-scale language models (LLMs) require creating separate prompts for each task when handling multiple task requests, resulting in inaccurate answers and a cumbersome process.
An agent device is designed, comprising a user response agent unit and a task execution agent unit, which automatically selects and executes multiple tasks, integrates LLM responses, and achieves multi-task processing through a recursive structure.
It improves the accuracy of LLM's response to a single request, simplifies user operations, can automatically handle multiple task requests, and is easy to extend.
Smart Images
Figure CN121636772A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a proxy device for performing various tasks using Large Language Models (LLMs), and more specifically, to a proxy device for performing composite tasks that combine multiple tasks that can be temporarily executed in an LLM. Background Technology
[0002] With the commercial use of ChatGPT (a registered trademark), LLM is being applied in practice, and various technologies for utilizing LLM have been proposed. For example, Patent Document 1 proposes the following structure: In order to achieve the learning of region-specific language models without large-scale computing resources, an SLM learning unit is added. This SLM learning unit includes: a first model learning unit, which uses a text dataset that is not limited to a specific region to learn a small-scale language model from a pre-learned large-scale language model that has been learned using a text dataset that is not limited to a specific region, so that the output text from the small-scale language model becomes the output text that reflects the knowledge of a specific region, and performs region-specific learning using a text dataset that is limited to a specific region; a dataset generation unit, which uses the text dataset that is limited to a specific region to generate the region-specific small-scale language model, and generates a text dataset that includes input text described in a representation that is not a specific representation and output text described in a specific representation, and reflects the knowledge of the specific region; and a second model learning unit, which uses the text dataset generated by the dataset generation unit to perform task-specific learning for adjusting the representation of the representation to generate a task-specific small-scale language model. Patent document 2 proposes the following: As a pre-training method for a language model, the purpose is to enable the model to model multiple task data simultaneously, improve the model's diversity, enhance the model's robustness in task learning, better learn knowledge related to the task and dataset, and improve the model's transfer ability in cases with zero or few samples. As a pre-training method for a language model, a pre-training language dataset including unsupervised and supervised language data is constructed. Based on the pre-training language dataset, a hierarchical multi-template multi-task language dataset is generated. Based on the hierarchical multi-template multi-task language dataset, the language model is pre-trained.
[0003] Patent Document 1: Japanese Patent No. 7442239
[0004] Patent Document 2: Japanese Patent Application Publication No. 2023-12493
[0005] Non-patent literature 1: "Attention Is All You Need," Ashish Vaswani, et al.<https: / / arxiv.org / abs / 1706.03762> Summary of the Invention
[0006] In short, the existing (pre-learned) LLMs (such as OpenAI's Chat GPT, a registered trademark) are structured as follows: given input of any prompt word written in natural language, they perform the task requested by that prompt word and output an answer. In this regard, in the current LLMs described above, each task is performed for each prompt. However, to perform multiple tasks for a single request and obtain a more accurate answer, it is necessary to create prompts for each task used to respond to that request and obtain an answer from the LLM for each prompt. Therefore, it would be convenient if there existed a structure that could automatically create prompts for performing multiple tasks from the prompt representing a request in order to obtain a more accurate answer, input each prompt into the LLM to perform each task, and integrate the LLM's answers from each task. In this case, to more accurately answer the prompts created from the prompt representing a request, further prompts are created so that tasks are performed for each prompt and an answer is obtained, with the expectation of obtaining a more accurate answer.
[0007] In view of the above, the main objective of the present invention is to provide a structure that utilizes the LLM described above, in order to obtain a more accurate answer to a request, can automatically create prompts for performing multiple tasks from the prompt representing the request, can input each prompt into the LLM to perform each task, and can integrate the LLM answers from each task.
[0008] According to one aspect of the present invention, the above-mentioned problem is accomplished by means of an apparatus that uses a pre-learned large-scale language model to output a response to a prompt input by a user representing their request, wherein...
[0009] A task execution agent unit, configured to execute individual tasks separately, and configured to, upon receiving a prompt, execute the configured individual task using a large language model and generate a response to the received prompt; and
[0010] The user response agent unit is configured to receive prompts from the user and use a large-scale user response language model to generate and output responses to the prompts from the user. Furthermore, it is configured to, upon receiving a prompt from the user, select a task execution agent unit from the task execution agent units to execute the task for generating responses to the prompts from the user, send a prompt requesting task execution to the selected task execution agent unit, receive a response from the selected task execution agent unit to the prompt requesting task execution, and generate and output a response to the user based on that response.
[0011] In the above structure, the device of the present invention is implemented through the operation of a computer system according to a program. The large-scale language model (LLM) can be any large-scale language model available in the field. The large-scale language models used in this invention (large-scale language model for a single task, large-scale language model for user responses) are pre-learned LLMs. Furthermore, the large-scale language model for user responses and the large-scale language model for a single task can be the same or prepared separately. The prompts input by the user into the device can be arbitrarily constructed based on natural language, and the responses to these prompts are also constructed using natural language generated by the LLM. As described above, the "task execution agent unit" is configured to execute individual tasks, specifically, individual tasks can be searching, summarizing, translating, writing, task management (dividing the questioner's topic into smaller tasks for management and selecting from high-priority tasks), etc.
[0012] In the above structure, if a user inputs a prompt representing any request into the device, the prompt is received by the user response agent unit. The user response agent unit selects a task execution agent unit that uses the user response application LLM to respond to the prompt from the user, and sends a prompt requesting the task execution to the selected task execution agent unit. Furthermore, the user response agent unit is configured to input a prompt to the user response application LLM based on the prompt from the user, causing it to respond. This response is used to execute a task execution agent unit that responds to the prompt from the user. The user response agent unit is configured to generate a prompt requesting the task execution agent unit specified in the response of the user response application LLM to the prompt to perform the task, and sends the generated prompt to the specified task execution agent unit.
[0013] In this way, the task execution agent unit that receives the prompt to execute the task responds to the prompt to execute the task, uses a single task to execute the task requested by the LLM, and sends the execution result as a reply to the source of the prompt to execute the task, namely the user response agent unit. The user response agent unit generates and outputs a reply to the user based on the reply.
[0014] According to this structure, if the user inputs any prompt, the user should select a task execution agent unit to automatically execute the task required to answer the prompt, and then have the task execution agent unit execute the task, generate and output an answer to the user from the prompt. Therefore, even if the user's request requires the execution of multiple tasks, the user does not need to create prompts for each of the multiple tasks that should be executed for the request; they only need to create a prompt corresponding to the user's request and input it into the device to expect a more accurate answer.
[0015] Furthermore, in the structure of the present invention described above, the task execution agent unit can be configured as follows: not only receiving prompts from the user response agent unit or other task execution agent units, executing tasks set for itself according to the received prompts, generating a response to the received prompts, and outputting it to the user response agent unit or other task execution agent unit that sent the received prompts, but also, in order to generate a response to the received prompts, being able to select other task execution agent units that execute tasks different from the tasks it executes, being able to send prompts requesting the execution of other tasks to the selected other task execution agent units, being able to receive responses from the selected other task execution agent units regarding the requests to execute other tasks, and being able to generate a response to the received prompts according to the responses and output it to the user response agent unit or other task execution agent unit that sent the received prompts.
[0016] That is, each task execution agent unit can be configured as follows: in order to respond to the prompt it receives, it requests the execution of a task different from its own from other task execution agents, and based on their responses, generates a response from the source that sent the prompt it received. In other words, the user should be able to form a "recursive" structure with agent units and multiple task execution agents, which repeatedly performs the following mechanism: forming a hierarchical structure, with requests from upstream agent units to downstream agent units and responses from downstream agent units to upstream agent units. According to this structure, it is expected that combining multiple tasks for a request will yield a more accurate response. Furthermore, when new tasks are needed, the structure can be easily added, thus facilitating functional expansion.
[0017] In this implementation, the user response agent unit can be composed of a preprocessing unit and a postprocessing unit. In the preprocessing unit, prompts from the user, responses from task execution agent units, information about each selectable task execution agent unit and the tasks they perform, and a prompt requesting that at least one task execution agent unit be output to perform a task when at least one task execution agent unit is used to generate a response to the user's prompt are input to the user response application LLM. Additionally, to generate a response to the user's prompt, the response from the task execution agent unit causes any task execution agent unit to perform a task, and this response is input to the user response application LLM upon receipt. Thus, the user response application LLM understands the user's request based on the input prompts, and understands the content of the task that can be executed by the device as a whole and the task execution agent unit that performs the task. When it determines that any task execution agent unit should perform the task set therein in order to accurately respond to the user's request, it outputs a prompt requesting the task execution agent unit to perform the task in the response. Furthermore, when the user has received a response from any task execution agent and that response is in response to the user's request, the user should output its own response based on that response.
[0018] Then, in the post-processing unit, a response to a prompt input during the preprocessing of the user-responsive LLM is received. If the response includes a prompt requesting at least one task execution agent unit to perform a task, a prompt requesting execution is output to the task execution agent unit requested to perform the task. Conversely, if the response from the user-responsive LLM does not include a prompt requesting a task execution agent unit to perform a task, i.e., if the response from any task execution agent unit includes a response to the user, a response for the user is generated based on that response. Thus, the task execution agent unit performs the task to obtain an accurate response to the user.
[0019] Furthermore, as an implementation method, each task execution agent unit can be configured as follows: a prompt specifying the content of the task it will execute is input into a single task LLM, and this single task LLM is set to execute the task it will perform. Thus, each task execution agent unit can configure its single task LLM in a way that accurately completes its respective task. Additionally, the single task LLM can be fine-tuned to ensure that it accurately performs the single task.
[0020] Furthermore, the task execution agent unit can also be composed of a preprocessing unit and a postprocessing unit. In the preprocessing unit, prompts from the user response agent unit or other task execution agent units, responses from other task execution agent units, prompts for each selectable other task execution agent unit and information about the tasks they perform, and prompts requesting that at least one other task execution agent unit output a prompt to perform other tasks when using it to generate a response to a received prompt are input into the single task LLM. Additionally, to generate a response to a received prompt, a response from another task execution agent unit causes any of the other task execution agent units to perform another task, and upon receiving this response, it is input into the single task LLM. Therefore, a single-task LLM uses the input prompts to understand the content of requests from user response agents or other task execution agents, and further understands the content of other executable tasks and other task execution agents that can perform those tasks. When it determines that any other task execution agent should execute a pre-defined task in order to accurately respond to a request from a user response agent or other task execution agent, it outputs a prompt requesting the other task execution agent to perform that task in its response. Furthermore, once a single-task LLM has received a response from any other task execution agent that responds to the request from the user response agent or other task execution agent, it outputs its own response based on that response.
[0021] Then, in the post-processing unit, a response to the prompt is received from the preprocessed input of the LLM for a single task. When the response includes a prompt requesting the execution of another task from at least one other task execution agent unit, or when such a request is pre-set, a prompt requesting the execution of another task is output to the other task execution agent unit that is requested to perform the other task. When the response from the LLM for a single task does not include a prompt requesting the execution of another task from another task execution agent unit and is not pre-set, that is, when based on the response from any other task execution agent unit, including requests based on responses from the user response agent unit or other task execution agent units, a response to the received prompt is generated based on the response from the LLM for a single task. This causes the other task execution agent unit to execute the task in order to obtain an accurate response to the requests from the user response agent unit or other task execution agent units.
[0022] Invention Effects
[0023] Thus, based on the above structure, using LLM, in order to obtain a more accurate answer to a request, prompts for performing multiple tasks can be automatically created from the prompts representing that request. Each prompt can be input into the LLM to execute each task, and the LLM responses from each task can be integrated. In this invention, agents (task execution agent units) are configured for individually responding to specific tasks, and each agent effectively executes a single task. Compared to the case of giving vague requests, LLM can achieve high performance for a single task. Furthermore, by combining multiple agents, an apparatus capable of performing multitasking using LLM can be constructed that maintains high performance.
[0024] Other objects and advantages of the present invention will become apparent from the following description of preferred embodiments of the invention. Attached Figure Description
[0025] Figure 1 This is a schematic diagram of a computer system that implements an agent device for performing tasks using an LLM based on this embodiment.
[0026] Figure 2 This is a conceptual diagram illustrating the relationship between the user response agent and the task execution agent in the agent device based on this embodiment.
[0027] Figure 3 It is a diagram that schematically represents the structure of each agent (user response agent and task execution agent) in the form of a block diagram.
[0028] Figure 4 It is a diagram that represents the processing in each agent (user response agent and task execution agent) in the form of a flowchart.
[0029] Symbol Explanation
[0030] 1-Computer device (terminal), 2-Display, 3-Input device, 10-Network system (LLM), 101-Proxy preprocessor, 102-LLM, 103-Proxy postprocessor. Detailed Implementation
[0031] Hereinafter, several preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the drawings, the same symbols denote the same parts.
[0032] The structure of a computer system
[0033] The agent device that performs tasks using the LLM based on this embodiment, through, as... Figure 1This is implemented as an example of a computer system that operates according to a procedure commonly used in this field. In the computer system, the user inputs requests to an agent device via computer device 1 and refers to the responses from the agent device. Computer device 1 is typically equipped with a CPU, storage devices, and input / output devices (I / O) interconnected via a bidirectional shared bus. It is also typically equipped with a display 2 and input devices 3 such as a keyboard and mouse. If a program is started, the user follows the program steps, uses the input devices 3 to give various instructions and inputs to computer device 1 according to the display on display 2, and can visually confirm responses from the agent device on display 2. Furthermore, computer device 1 is configured to communicate appropriately with network system 10, typically with computer device 12 equipped with an LLM, database, etc., present in network system 10, and utilizes the LLM, database, etc., in task processing.
[0034] Structure of the agent device
[0035] In the proxy device that is the object of this embodiment, if a user inputs a request, the LLM performs the task required to answer the input request, such as searching, summarizing, writing, translating, etc., and generates and outputs an answer to the user based on the execution result. In this regard, when multiple tasks need to be performed in order to accurately respond to a request from a user, the LLM is not good at handling multiple tasks simultaneously. Therefore, conventionally, the following cumbersome process is performed: prompting the user to perform a single task required to answer a request, and providing these prompts individually to the LLM, etc.
[0036] Therefore, this embodiment provides a structure that eliminates the need for the user to create prompts for performing individual tasks, as was the case previously. More specifically, the agent device in this embodiment comprises a "user response agent" and multiple "task execution agents" configured to mutually send and receive responses including prompts and task execution results (when simply referred to as "agents," this includes both "user response agents" and "task execution agents"). Specifically, firstly, the task execution agent is configured to perform the single task described above for responding to a request. Then, the user response agent is configured to: upon receiving a request from a user, automatically select a task execution agent to perform the task required to respond to the request, provide the selected task execution agent with a prompt indicating task execution, enable it to perform the task, receive the task execution result, and generate a response to the user based on the result. Therefore, if a user makes a request to the agent, the user response agent will provide prompts to the task execution agent, which is responsible for automatically responding to the request, indicating the various tasks. Upon receiving the prompts, the task execution agent executes the indicated single task and returns its response to the user response agent. Based on the response from the task execution agent, the user response agent generates and outputs a response to the user. According to this structure, even without the individual prompting of the single task required to generate a response to the user's request, multiple tasks for responding to LLM-based user requests are executed individually, potentially providing a more accurate response to the user.
[0037] Furthermore, in this embodiment described above, each task execution agent performing a single task can be configured as follows: receiving prompts for task execution instructions not only from the user response agent but also from other task execution agents; selecting other task execution agents that automatically execute other necessary tasks in response to the received instructions; providing prompts for other task execution instructions to the selected other task execution agents; executing the other tasks; receiving the results of the other task executions; and generating and outputting responses to the other task execution agents that provided the instructions based on the results. Then, the other task execution agents that receive the responses can be configured to generate and output responses to user response agents or task execution agents that further instruct themselves based on the responses.
[0038] Figure 2 This schematically illustrates the sending and receiving relationship of prompts p and responses r between user U and user response agent C0 and multiple task execution agents C1, C2, ... . For example, from... Figure 2In this embodiment, the proxy device is configured as follows: At the upstream of the flow where user response proxy C0 sends and receives prompts p and responses r to user U, prompts p indicating task execution are provided to multiple task execution proxies C1, C2, ... , indicating the execution of each individual task. Response r, including the result of task execution, is received from the task execution proxy at the destination of the prompt, and is sent back to user U based on the response r. Similarly, each task execution proxy Ci (i denotes the symbol for each task execution proxy) is configured as follows: on one hand, it receives instructions from its upstream user response proxy C0 or other task execution proxies Ci-1; on the other hand, it assigns command prompts p indicating the execution of each individual task to multiple other task execution proxies Ci+1, receives responses r, including the result of task execution, from the task execution proxies Ci+1 at the destination of the command prompt, and sends responses r to the upstream proxies based on the responses r. That is, in the apparatus of this embodiment, the structure for sending and receiving prompts p and responses r between agents can be "recursively" expanded, thereby making it easy to expand the functionality of performing various tasks for responding to requests.
[0039] Structure of each agent
[0040] like Figure 3 As schematically depicted, each agent Ci in the agent device of this embodiment can have essentially the same structure. Specifically, in each agent Ci, firstly, the preprocessing unit 101 receives a prompt pi from the user U or the upstream agent Ci-1, and sends a prompt p* generated based on the prompt pi, together with a prompt p+ representing information required by the LLM unit 102 to perform a task. The LLM unit 102 receives the prompts p* and p+, and uses LLM to generate its response r*. Then, the postprocessing unit 103 receives the response r* from the LLM unit 102, generates a response ri to the user U or the upstream agent Ci-1, and when the response r* from the LLM unit 102 contains an instruction to perform a task for the downstream agent Ci+1, or when the downstream agent Ci+1 is set to perform a task, it creates and outputs a prompt pi for performing the task. Additionally, when a prompt pi instructing the downstream agent Ci+1 to perform a task is sent, the preprocessing unit 101 receives the response ri+1 from the downstream agent Ci+1, generates a prompt input into the LLM 102 based on the response ri+1, and sends it to the LLM 102. Then, the LLM 102 generates a new response based on the prompt generated from the response ri+1, and the postprocessing unit 103 receives this response to generate a response ri for either the user U or the upstream agent Ci-1.
[0041] In the structures of the aforementioned agents, the preprocessing unit 101 and the postprocessing unit 103 are configured through programming-based prompting engineering. The LLM in the LLM unit 102 can utilize any pre-learned LLM; specifically, it can utilize Transformer models based on "Attention Is All You Need" (Non-Patent Document 1), such as OpenAI, Inc.'s GPT system, Anthoropic, Inc.'s Claude system, METAPlatforms, Inc.'s LLaMA and its derivative models. The LLM of each agent can be the same LLN, or it can be a specialized LLM for a single task.
[0042] Furthermore, when the user responds to agent C0, a prompt p+ indicating information required to execute the task assigned to the LLM from preprocessing unit 101 to LLM unit 102 is generated, along with prompts indicating the type of task execution agent and the content of the respective task to be executed, indicating that an instruction to execute a task can be sent from itself. On the other hand, in the case of each task execution agent, a prompt p+ indicating information required to execute the task assigned to the LLM from preprocessing unit 101 to LLM unit 102 is generated, along with prompts indicating the content of the task (executable processing) to be executed by itself, prompts indicating that an instruction to execute other tasks can be sent from itself, and the content of the other tasks to be executed. As described above, in each agent, by providing the prompt p+ indicating information required to execute a task to the LLM, the LLM of LLM unit 102 is set to execute a single task.
[0043] In the post-processing unit 103 of each agent, as described above, an LLM response r* from the LLM unit 102 is received, and a response ri is generated for user U or upstream agent Ci-1, or a prompt indicating task execution for downstream agent Ci+1 is generated based on the response r*. Then, regarding the response ri for user U or upstream agent Ci-1, the response r* can be directly used as the response ri for user U or upstream agent Ci-1, depending on the content of the task and the status of the response; alternatively, various processes that do not use LLM (e.g., keyword search within a database) can be performed based on the response r*. Regarding the prompt indicating task execution to downstream agent Ci+1, a prompt is generated and output in a manner that allows agent Ci+1 to perform the task, as specified or pre-set by the response r*. In this case, when generating the prompt for agent Ci+1, various processes that do not use LLM (e.g., keyword search within a database) can also be performed based on the response r*. That is, the processing in the post-processing unit varies depending on the task set to be performed by each agent.
[0044] Furthermore, in the above structure, the prompts p and responses r exchanged between agents are basically described in natural language, except for some symbolic descriptions.
[0045] The work of each agent
[0046] Each agent can be structured as basically as follows Figure 4 The process operates as shown in the flowchart. In this regard, in the user response agent C0, the upstream of the prompting process is the user, and the task performed here is to assign tasks to task execution agents that execute the individual tasks corresponding to the prompting requests from the user. On the other hand, in the task execution agents C1, C2, ..., the upstream of the prompting process is the user response agent C0 or other task execution agents, and the task performed here is to assign tasks to other task execution agents that execute other tasks when individual tasks and other tasks are required. The operation of the user response agent and the task execution agents will be explained below (the flowcharts referenced are all shown in...). Figure 4 middle.).
[0047] (1) User's response to the agent's work
[0048] refer to Figure 4 In the user response agent, if the preprocessing unit 101 receives a prompt from the user (step S1), it performs the generation of the prompt input into the LLM (step S2). Here, as shown in the following example, a prompt is generated that incorporates information about the task execution agent (e.g., the agent's name, what it can do, and what kind of input it expects) that includes the prompt from the user and the prompt that can send task execution instructions from itself.
[0049] [prompt example]
[0050] <system>:
[0051] You are an AI assistant and are chatting with a user. Please respond to the user's request.
[0052] You can invoke the following Agent as needed.
[0053] #List of Agents
[0054] • (Web Search Agent): Capable of searching the internet and retrieving information. (Input: search keywords)
[0055] • (Internal Search Agent): Searches the company's internal database and retrieves information. (Input: Search keywords)
[0056] ·(○○Agent):…
[0057] Note: When calling the Agent, please output " <call> {Agent name}, input passed to the Agent< / call> "
[0058] <user>:
[0059] "Investigate any defects in the car's threaded connections and provide a brief explanation."
[0060] [A tip from a user]
[0061] <assistant>:
[0062] [End of example]
[0063] As described above, if a prompt is generated for the LLM, the prompt is input into the LLM of the LMM unit 102 (step S3). If the LLM understands the user's request in the prompt generated by the preprocessing unit 101 and determines that a certain task execution agent should perform the task, then a response including the response example described below is generated (step S4).
[0064] [Example Answer]
[0065] <call> Internal company search agent, defective car parts< / call>
[0066] [End of example answer]
[0067] In this example response, the internal search agent, which performs the task, selects to search the company's internal database and performs the task of obtaining information. The internal search agent contains the information required for the task execution agent to perform the task (keyword: defective car with thread).
[0068] Thus, as in the above response example, when a task execution agent is specified to perform the task (step S5: Yes), the information required to perform the task and a prompt to perform the task are sent to the specified task execution agent Ci (step S6). Then, as explained later, if the preprocessing unit 101 receives a response from the agent Ci that specified the task (step S7), the preprocessing unit 101 generates a prompt for the LLM based on the response (step S8) and inputs the prompt to the LLM unit 102 (step S3). In response to this, if the LLM in the LLM unit 102 does not include the text for selecting a task execution agent (step S5: No), the response from the LLM can be provided as a response to the user (step S9).
[0069] [Example Answer] The search results revealed the following information.
[0070] Poor thread relationships in automobiles can lead to various problems.
[0071] [End of example answer]
[0072] Additionally, in the above structure, when the LLM responds to the prompt generated based on the answer from the task execution agent with a literal reply that causes a certain task execution agent to execute a task, steps S6 to S8 can be repeated.
[0073] (2) Task execution agent's work
[0074] As described above, the task execution agent references the prompt from the upstream agent and uses the LLM to execute a separately configured task. Regarding this, depending on the content of the task being executed, there may be cases where the entire LLM is executed, or cases where the LLM executes only a portion of the task while omitting the rest. Furthermore, a task execution agent can be configured to automatically cause other task execution agents to execute tasks during task execution, and to continue executing tasks based on their responses.
[0075] For example, in the case of the agent (search agent) that performs a specified search to obtain information in the above example of the user response agent, firstly, the LLM performs the task of selecting a database for the search, receives its response, and performs a keyword search in the database selected by the LLM without using LLM programming. Then, according to pre-set parameters, the files found through the search are sent to an agent (summarization agent) that performs a summarization task. Here, the LLM is used to generate summaries for each file, and the contents of the summarized files are returned to the search agent. Finally, the task of generating a response using the summaries is performed to provide the user response agent with the response.
[0076] Specifically, refer to again Figure 4 In the task execution agent, if the preprocessing unit 101 receives a prompt from another agent (step S1), it first performs the generation of the prompt input into the LLM (step S2). Here, as shown in the following example, a prompt is generated that incorporates the prompt from the user, information about the content of the task to be executed by the LLM, and information required to execute the task.
[0077] [prompt example]
[0078] <system>:
[0079] You are the robot selecting the database to search. Please select the database based on the user's keywords.
[0080] #DB Overview
[0081] DB name stores data
[0082] Thread data and thread defects
[0083] Coating data, coating defects
[0084] Personnel data, personnel information data
[0085] ...
[0086] <user>:
[0087] Defective Thread Car
[0088] [A notification from the upstream agent]
[0089] <assistant>
[0090] [End of example]
[0091] In the example above, the types of databases that can be used in LLM are suggested, and a database is generated requesting the selection of keywords suitable for the prompt from the upstream agent. Additionally, the prompt can also include information about the task execution agent that can send instructions for other task executions from itself, such as the agent's name, what it can do, and what kind of input it expects.
[0092] As described above, if a prompt is generated for the LLM, the prompt is input into the LLM of the LMM unit 102 (step S3). Then, the LLM, based on the prompt generated by the preprocessing unit 101, understands the content of the prompt from the upstream agent and the task it should perform, and generates a response including the response example described below (step S4).
[0093] [Example Answer]
[0094] Thread data
[0095] [End of example answer]
[0096] If a response is received from the LLM, the task is further executed in the post-processing unit 103 based on that response. In this case, in the example of the retrieval agent, a keyword search can be performed in the database described in the response, and the discovered documents can be collected. Furthermore, in the example of the search agent, it is set to have the summary agent summarize the searched documents after the search is performed. Therefore, even if no other task execution agent is specified in the LLM's response, it is determined to automatically send a task execution instruction to the summary agent (step S5: Yes). A prompt can be sent along with the information required to perform the task (in this example, the documents found through the search), indicating the task execution in the following example (step S6).
[0097] [prompt example]
[0098] <system>:
[0099] You are an information extraction and summarization robot. Please extract the parts from the following document that will help answer user questions.
[0100] #document
[0101] ~~ ...
[0102] ~~ ...
[0103] ~~ ...
[0104] <user>:
[0105] Please investigate the defective thread relationship of the car and provide a brief explanation.
[0106] [A tip from a user]
[0107] <assistant>
[0108] [End of example]
[0109] Upon receiving the above prompt, another agent (summary agent) performs another task (summary task). Then, if the preprocessing unit 101 receives a response from the other agent (step S7), the preprocessing unit 101 generates a prompt to be provided to the LLM again based on the response as follows (step S8).
[0110] [prompt example]
[0111] <system>:
[0112] Please refer to the following information to answer the user's question.
[0113] If the following information is insufficient, you can change the keywords and search again.
[0114] #File A
[0115] ~~ ...
[0116] ~~ ...
[0117] #File B
[0118] ~~ ...
[0119] ~~ ...
[0120] #File C
[0121] ~~ ...
[0122] ~~ ...
[0123] ~~ ...
[0124] <user>:
[0125] Please investigate the defective thread relationship of the car and provide a brief explanation.
[0126] [A tip from a user]
[0127] <assistant>
[0128] [End of example]
[0129] If the above prompt is input into the LLM of LLM unit 102 (step S3), the LLM outputs the answer as described below (step S4) and returns the output answer to the upstream agent (step S5: no, step S9).
[0130] [Example Answer] Poor thread relationships in automobiles include ~~ ...
[0131] [End of example answer]
[0132] (In the search agent, the response to LLM's prompts regarding results using the digest agent is set to not issue any instructions to the digest agent.)
[0133] LLM pre-tuning
[0134] In the agent device of this embodiment described above, the task to be performed is specified using data from user-provided prompts and their responses, and LLM learning is performed, thereby consistently obtaining more accurate answers with lower learning costs than before. According to experiments conducted by the inventors of this embodiment, good functional acquisition was achieved through two periods of learning using approximately 2000 case data. Furthermore, by mixing modes that do not specify tasks to be performed on a portion of the learning data, learning can be achieved while preserving normal dialogue functions.
[0135] Experimental Example
[0136] The effectiveness of this embodiment has been confirmed through the following experimental examples. Furthermore, it should be understood that the following experimental examples are illustrative of the effectiveness of this embodiment and do not limit the scope of the invention.
[0137] In the proxy device of this embodiment, as a prompt from the user, the message "Regarding the mechanism of ultrasound examination, please investigate the Domain and provide a brief explanation." is entered.
[0138] The prompt a user should see when executing a task in the proxy is as follows.
[0139] "Function calling{"name":"domain_searcher","arguments":{"key_words":"ultrasound inspection mechanism"}}"
[0140] The final response to the user is as follows.
[0141] Ultrasonic testing is a non-destructive testing method primarily used to detect defects or foreign objects existing inside materials or products. Its structure is as follows:
[0142] Ultrasonic wave transmission and reception:
[0143] Using a device called a probe, ultrasonic waves are sent (emitted) into the interior of the material. This probe typically uses a ceramic transducer.
[0144] If an ultrasonic wave incident inside a material encounters a foreign object or defect, it is reflected and the reflected wave is received again by the probe.
[0145] Signal processing:
[0146] The received reflected waves are converted into electrical signals and displayed as waveforms by display devices such as oscilloscopes.
[0147] By analyzing the waveform, the location or size of the internal defect can be determined.
[0148] Inspection methods: Use direct contact or automated flaw detection. Typically, the probe is pressed directly onto the surface of the test specimen for inspection.
[0149] The flaw detection sensitivity is set according to the JIS Z2345 STB-G V15-2 standard, and the size of the defect is expressed as the bottom surface ratio.
[0150] PAUT (Phased Array Ultrasonic Testing):
[0151] Multiple pulse generators / receivers are used to synthesize an ultrasonic beam, and the phase of each oscillator is electronically controlled to output a larger signal.
[0152] Therefore, it can detect defects with high precision, and is especially suitable for the inspection of discontinuous materials or high-attenuation materials.
[0153] Thus, ultrasonic testing, as a highly reliable non-destructive testing method, is widely used in the quality management of materials or products.
[0154] Thus, according to this embodiment, the following is shown: Using LLM, in order to obtain a more accurate answer to a request, prompts for performing multiple tasks can be automatically created from the prompts representing the request, each prompt can be input into the LLM to perform each task, and the LLM answers from each task can be integrated.
[0155] The above description relates to the embodiments of the present invention, but those skilled in the art can easily make numerous modifications and changes. The present invention is not limited to the embodiments illustrated above, and can obviously be applied to various devices without departing from the concept of the present invention.< / assistant> < / user> < / system> < / assistant> < / user> < / system> < / assistant> < / user> < / system> < / assistant> < / user> < / system>
Claims
1. An apparatus that outputs an answer to a cue representing a request of a user inputted by the user using a large language model that is pre-learned, the apparatus characterized by comprising: a task execution agent unit configured as a task execution agent unit set to execute a single task respectively, and configured to execute the set single task using a single task large language model if a cue is received, and generate an answer to the received cue; and a user coping agent unit configured as a user coping agent unit that receives a cue from the user, and generates and outputs an answer to the cue from the user using a user coping large language model, and configured to, if a cue from the user is received, be able to select a task execution agent unit of the task execution agent units that executes a task for generating an answer to the cue from the user, be able to send a cue requesting execution of the task to the selected task execution agent unit, and be able to receive an answer to the cue requesting execution of the task from the selected task execution agent unit, and be able to generate and output an answer to the user in accordance with the answer.
2. The apparatus according to claim 1, characterized in that the task execution agent unit is respectively configured to: receive a cue from the user coping agent unit or another task execution agent unit, execute a task set to itself in accordance with the received cue, generate an answer to the received cue, and output it to the user coping agent unit or the other task execution agent unit that is a source of transmission of the received cue, and in order to generate an answer to the received cue, be able to select another task execution agent unit that executes a task different from the task executed by itself, be able to send a cue requesting execution of the other task to the selected other task execution agent unit, be able to receive an answer to the cue requesting execution of the other task from the selected other task execution agent unit, be able to generate an answer to the received cue in accordance with the answer and output it to the user coping agent unit or other task execution agent unit that is a source of transmission of the received cue.
3. The apparatus according to claim 1, characterized in that the user coping agent unit includes: a preprocessing unit configured to input a cue from the user, an answer from the task execution agent unit, a cue that selects each of the task execution agent units and provides information of a task executed by them, and a cue that requests output of a cue requesting execution of a task to at least one of the task execution agent units in order to complete a task for generating an answer to a cue from the user using the at least one of the task execution agent units, into the user coping large language model; and a post-processing unit configured to receive a response to a prompt input in the pre-processing from the user-dealing large language model, and when the response from the user-dealing large language model includes a prompt requesting a task to be executed by at least one of the task-executing agent units, output the prompt requesting the task to be executed to the task-executing agent unit requested to execute the task, and when the response from the user-dealing large language model does not include a prompt requesting the task to be executed by the task-executing agent unit, generate a response to the received prompt based on the response from the user-dealing large language model.
4. The apparatus according to claim 1, wherein the task-executing agent units are each configured to input a prompt specifying a content of a task to be executed by the task-executing agent unit to the single-task large language model, and set the single-task large language model to execute the task to be executed by the task-executing agent unit.
5. The apparatus according to claim 2, wherein the task-executing agent units each include: a pre-processing unit configured to input, to the single-task large language model, a prompt from the user-dealing agent unit or another task-executing agent unit, a response from the other task-executing agent unit, a prompt from each of the other task-executing agent units and information on a task executed by the other task-executing agent unit, and a prompt requesting output of a prompt requesting a task to be executed by at least one of the other task-executing agent units to the at least one of the other task-executing agent units when the at least one of the other task-executing agent units is used to complete a task for generating a response to the received prompt; and a post-processing unit configured to receive a response to a prompt input in the pre-processing from the single-task large language model, and when the response from the single-task large language model includes a prompt requesting a task to be executed by at least one of the other task-executing agent units or is set in advance to request the task to be executed by the other task-executing agent unit, output the prompt requesting the task to be executed to the other task-executing agent unit requested to execute the task, and when the response from the single-task large language model does not include a prompt requesting the task to be executed by the other task-executing agent unit and is not set in advance to request the task to be executed by the other task-executing agent unit, generate a response to the received prompt based on the response from the single-task large language model.