Large model artificial intelligence assistant with configurable strategy

Abstract

The invention relates to the technical field of artificial intelligence, and provides a strategy-configurable large-model artificial intelligence assistant, which comprises a planning strategy configuration module used for configuring different types of planning strategies and corresponding configuration parameters; the historical message processing strategy configuration module is used for configuring different types of historical message processing strategies and corresponding configuration parameters; and the tool retrieval configuration module is used for configuring the number of tool retrieval and a relevancy threshold. According to the technical scheme, the problem that the final result is affected due to the fact that an artificial assistant cannot be flexibly configured in the related technology is solved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a strategy-configurable large-model artificial intelligence assistant. Background Art

[0002] AI assistants can be developed through some open source software frameworks and tools. These tools do not provide models themselves, but implement task processes by connecting to large third-party models. This results in the separation of workflows (specific tasks) and models, and lacks customized optimization for specific processes and tasks. From the perspective of usage threshold, users not only need to manually maintain the interaction with the model, but also need to handle multiple prompts during debugging and connect to multiple tools for adaptation. In addition, users also need to have a deep understanding of large models and their related tool chains.

[0003] AI assistants can also provide services to users in the form of APIs, and customize and optimize the use of common tasks and tools based on existing models, so that users can obtain more stable usage results through more intuitive and concise dialogue interactions. This greatly reduces the threshold and maintenance costs for users to use AI assistants. Users do not need to pay too much attention to process construction and details, but only need to focus on defining tasks and concentrate their energy on how to use AI assistants to better solve business problems.

[0004] In the operation assistant phase, the large model can only plan and arrange according to an inherent planning strategy and cannot be adjusted according to the actual scenario. For example, some scenarios pursue planning accuracy, while others pursue planning timeliness (low latency), and it is necessary to flexibly configure the artificial intelligence assistant according to the scenario.

[0005] In the operation assistant phase, the processing strategy of historical messages is fixed and cannot be adjusted according to the actual scenario. For example, in some scenarios, historical messages are not closely related, and in some scenarios, messages from a long time ago need to be traced back, and the AI ​​assistant needs to be flexibly configured according to the scenario.

[0006] The number of supported tools is limited, and the effect decreases when the number of tools increases. In fact, in most scenarios, the number of tools required to solve user problems is certain, but all tools are still provided to the big model for decision-making, resulting in a decrease in planning accuracy and affecting the final result. Summary of the invention

[0007] The present invention proposes a strategy-configurable large-model artificial intelligence assistant, which solves the problem in the related art that the artificial assistant cannot be flexibly configured, which affects the final result.

[0008] The technical solution of the present invention is as follows:

[0009] A strategy-configurable large-model AI assistant, including:

[0010] Planning strategy configuration module, used to configure different types of planning strategies and corresponding configuration parameters;

[0011] A historical message processing strategy configuration module is used to configure different types of historical message processing strategies and corresponding configuration parameters;

[0012] The tool retrieval configuration module is used to configure the quantity and relevance threshold of tool retrieval.

[0013] Optionally, the planning strategy configuration module includes a React strategy, a PS strategy, a TOT strategy and a GOT strategy.

[0014] Optionally, the historical message processing policy configuration module includes a truncation policy, a summary policy and a retrieval policy.

[0015] Optionally, the tool retrieval configuration module can retrieve the tool most relevant to the context according to configured parameters.

[0016] A method for using any of the strategy-configurable large-model artificial intelligence assistants described, comprising the following steps:

[0017] S1. Select and configure at least one planning strategy based on the specific usage scenario;

[0018] S2. Select and configure at least one historical message processing strategy based on the conversation history and context requirements;

[0019] S3. Configure the number and relevance thresholds of tool retrieval based on the conversation context and task requirements.

[0020] S4, interacting with the user through the artificial intelligence assistant and receiving the task instructions input by the user;

[0021] S5. Plan and arrange the task instructions using the configured planning strategy;

[0022] S6. Use the configured historical message processing strategy to process historical messages in multiple rounds of dialogue to generate context information suitable for large model processing;

[0023] S7, using the configured tool retrieval strategy to retrieve the tool most relevant to the current task from the tool set;

[0024] S8, driving the large model to execute tasks according to the planning results and the retrieved tools, and generating task execution results;

[0025] S9. Feedback the task execution result to the user.

[0026] Optionally, the planning strategy selected in step S1 may be at least one of a React strategy, a PS strategy, a TOT strategy or a GOT strategy, and the task planning and scheduling algorithm formula is:

[0027] Among them, P is the optimal planning result, ρ is the set of all possible plans, Ui is the utility weight of the task, Utility(p i ) is the planning step p i is the utility function, λ is the trade-off coefficient between complexity and utility, and Compelxity(p) is the complexity measure of planning p.

[0028] Optionally, the historical message processing strategy selected in step S2 may be at least one of a truncation strategy, a summary strategy or a retrieval strategy, and the historical message processing algorithm formula is:

[0029] Among them, M t is the message being processed at time t, m t is the current message, m t-i is the i-th message in the past, ω i is the weight of the corresponding message.

[0030] Optionally, the tool search strategy configured in step S3 includes setting the number of tool searches to not exceed a preset maximum value and setting a threshold for the relevance of the tool to the context. The tool search algorithm formula is:

[0031] T = argmax t∈τ (Relevance(t, C)-α·Cost(t)), where T is the selected tool, τ is the set of all tools, Relevance(t, C) is the relevance of tool t to the current context C, α is the trade-off coefficient between cost and relevance, and Cost(t) is the cost of using tool t.

[0032] Optionally, in step S5 to step S8, the artificial intelligence assistant uses the configured strategies and parameters to automatically complete the entire process from task reception to execution without user intervention.

[0033] Optionally, the task execution result fed back to the user in step S9 includes the task completion status, result details, and possible follow-up operation suggestions.

[0034] The working principle and beneficial effects of the present invention are:

[0035] The present invention mainly proposes a strategy-configurable large-model artificial intelligence assistant, which can be adapted according to the actual usage scenarios of the artificial intelligence assistant by configuring planning strategies and historical message processing strategies to obtain better results; by configuring tool retrieval strategies, the limitation on the number of tools supported by a single artificial intelligence assistant can be broken, and the most suitable tools can be provided to the model in each round of dialogue, thereby improving the accuracy of planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The preferred implementation modes will be described below in a clear and understandable manner with reference to the accompanying drawings to further illustrate the above-mentioned characteristics, technical features, advantages and implementation methods of the present invention.

[0037] Figure 1 Configure the module diagram for the AI ​​assistant;

[0038] Figure 2 Implementing module graphs for AI assistants;

[0039] Figure 3 A flowchart for an AI assistant;

[0040] Figure 4 This is a message processing flow chart.

[0041] In the figure: 1. Planning strategy configuration module, 2. Historical message processing strategy configuration module, 3. Tool retrieval configuration module. DETAILED DESCRIPTION

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings and other implementation methods can be obtained based on these drawings without creative work.

[0043] In order to simplify the drawings, only the parts related to the invention are schematically shown in each figure, and they do not represent the actual structure of the product. In addition, in order to simplify the drawings and facilitate understanding, in some figures, only one of the parts with the same structure or function is schematically shown, or only one of them is marked. In this article, "one" not only means "only one", but also means "more than one", and "several" includes "two" and "more than two".

[0044] In this document, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0045] In addition, in the description of the present application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0046] Reference Figure 1 to Figure 4 , is an embodiment of the present invention, and proposes

[0047] A strategy-configurable large-model AI assistant, including:

[0048] Planning strategy configuration module 1, used to configure different types of planning strategies and corresponding configuration parameters;

[0049] A historical message processing strategy configuration module 2 is used to configure different types of historical message processing strategies and corresponding configuration parameters;

[0050] The tool search configuration module 3 is used to configure the quantity and relevance threshold of tool searches.

[0051] Optionally, the planning strategy configuration module 1 includes a React strategy, a PS strategy, a TOT strategy and a GOT strategy.

[0052] Optionally, the historical message processing strategy configuration module 2 includes a truncation strategy, a summary strategy and a retrieval strategy.

[0053] Optionally, the tool retrieval configuration module 3 can retrieve the tool most relevant to the context according to the configured parameters.

[0054] A method for using any of the strategy-configurable large-model artificial intelligence assistants described, comprising the following steps:

[0055] S1. Select and configure at least one planning strategy based on the specific usage scenario;

[0056] S2. Select and configure at least one historical message processing strategy based on the conversation history and context requirements;

[0057] S3. Configure the number and relevance threshold of tool retrieval based on the conversation context and task requirements;

[0058] S4, interacting with the user through the artificial intelligence assistant and receiving the task instructions input by the user;

[0059] S5. Plan and arrange the task instructions using the configured planning strategy;

[0060] S6. Use the configured historical message processing strategy to process historical messages in multiple rounds of dialogue to generate context information suitable for large model processing;

[0061] S7, using the configured tool retrieval strategy to retrieve the tool most relevant to the current task from the tool set;

[0062] S8, driving the large model to execute tasks according to the planning results and the retrieved tools, and generating task execution results;

[0063] S9. Feedback the task execution result to the user.

[0064] Optionally, the planning strategy selected in step S1 may be at least one of a React strategy, a PS strategy, a TOT strategy or a GOT strategy, and the task planning and scheduling algorithm formula is:

[0065] Among them, P is the optimal planning result, ρ is the set of all possible plans, Ui is the utility weight of the task, Utility(p i ) is the planning step p i is the utility function, λ is the trade-off coefficient between complexity and utility, and Compelxity(p) is the complexity measure of planning p.

[0066] Optionally, the historical message processing strategy selected in step S2 may be at least one of a truncation strategy, a summary strategy or a retrieval strategy, and the historical message processing algorithm formula is:

[0067] Among them, M t is the message being processed at time t, m t is the current message, m t-i is the i-th message in the past, ω i is the weight of the corresponding message.

[0068] Optionally, the tool search strategy configured in step S3 includes setting the number of tool searches to not exceed a preset maximum value and setting a threshold for the relevance of the tool to the context. The tool search algorithm formula is:

[0069] T = argmax t∈τ(Relevance(t, C)-α·Cost(t)), where T is the selected tool, τ is the set of all tools, Relevance(t, C) is the relevance of tool t to the current context C, α is the trade-off coefficient between cost and relevance, and Cost(t) is the cost of using tool t.

[0070] Optionally, in step S5 to step S8, the artificial intelligence assistant uses the configured strategies and parameters to automatically complete the entire process from task reception to execution without user intervention.

[0071] Optionally, the task execution result fed back to the user in step S9 includes the task completion status, result details, and possible follow-up operation suggestions.

[0072] AI Assistant Configuration Module Figure 1 , including general configuration, planning strategy configuration, historical message processing strategy configuration, and tool retrieval configuration.

[0073] Conventional configuration module: a module not specific to the present invention, the conventional configuration of the artificial intelligence assistant, including large model configuration, instruction configuration, tool configuration, etc.

[0074] Planning strategy configuration module 1: Configure the planning strategy of the artificial intelligence assistant. The planning strategy includes the type of strategy and the configuration parameters of the strategy, such as React strategy, PS strategy, TOT strategy, GOT strategy, etc. Each strategy has different configuration parameters for more refined strategy control.

[0075] Historical message processing strategy configuration module 2: Configure the historical message processing strategy of the AI ​​assistant, including the type of strategy and the configuration parameters of the strategy, such as truncation strategy, summary strategy, retrieval strategy, etc. Each strategy has different configuration parameters for more refined strategy control. You can choose not to adopt any historical message processing strategy, in which case all user historical messages will be passed to the big model.

[0076] Tool search configuration module 3: Configure the tool search of the AI ​​assistant, including the number of tool searches and the relevance threshold. You can choose not to configure tool search, that is, to proceed in the traditional mode.

[0077] Example 1

[0078] like Figure 3 , taking the configuration of an "office smart assistant" as an example, the configuration process is explained.

[0079] Complete the general configuration of the AI ​​assistant, including configuring the AI ​​assistant's macro model, instructions, and tools.

[0080] In order to complete the configuration of the office intelligent assistant, we can select a large model with a suitable number of parameters for it and configure its instructions, such as "You are an office intelligent assistant, happy to help colleagues solve various office problems, and you can use various tools to assist you in completing tasks."

[0081] Finally, we add several tools to the assistant, such as conference room reservation, train reservation, air ticket reservation, reimbursement, etc. There are many tools, so this article will not list them one by one.

[0082] Complete the configuration of the planning strategy.

[0083] Although large language models have strong reasoning capabilities, they are somewhat unstable and uncertain when dealing with specific task planning and tool calling tasks. Since tool calling involves the combination of models and engineering, the planning strategy configuration for artificial intelligence assistants plays a big role.

[0084] The present invention proposes to provide a variety of planning strategies for users to choose from. For example, we configure a React strategy for the office intelligent assistant, which is more suitable for the current scenario, allowing the large model to decompose the task according to the context and complete the assigned tasks step by step.

[0085] Complete the configuration of the historical message processing strategy.

[0086] The length of text that a large model can process at one time is limited, and the conversations between AI assistants and humans are often multi-round. The messages that AI assistants reply to are not only related to the current conversation, but often involve historical messages. For example, the following conversation:

[0087] Human: Help me find currently available meeting rooms.

[0088] AI: The currently available conference rooms are: A01, B03, D08.

[0089] Human: Help me reserve the first meeting room.

[0090] If only the text of one round of conversation is input to the AI ​​assistant, the only instruction it gets is "Help me book the first meeting room." Obviously, the intention is unclear and needs to be given in combination with historical conversations. If the entire historical conversation is input to the AI ​​assistant, it may exceed the text processing length limit of the large model. In different scenarios, the most appropriate historical message processing strategy is different.

[0091] The present invention proposes to provide a variety of historical message processing strategies for users to choose from. For example, we configure a truncation strategy for the office intelligent assistant, and input the most recent five rounds of conversations into the large model each time. This strategy can not only control the length of the input message, but also adapt to the general demands of the office scene.

[0092] Complete the configuration of tool retrieval.

[0093] The description of the tool needs to be provided to the big model in as much detail as possible so that the AI ​​assistant can accurately select the tool to solve the problem. Similarly, the description of the tool is also text, which is limited by the length of text that the big model can process at one time. The number of tools involved in the interaction process between the AI ​​assistant and humans is limited, so the most suitable tool can be retrieved from the tool set configured by the AI ​​assistant and then provided to the big model, rather than providing all the tools at once.

[0094] The present invention proposes a method for tool retrieval configuration, which can configure multiple tools for an artificial intelligence assistant and configure corresponding parameters to perform retrieval when running the artificial intelligence assistant. For example, our office intelligent assistant configures the retrieval quantity to 2 and the threshold to 0.8. As mentioned above, although we have configured multiple tools for the office intelligent assistant, only three tools that are most relevant to the context are provided to the large model each time.

[0095] Example 2

[0096] After completing the configuration of the AI ​​assistant, we can use it. A typical message processing process is as follows: Figure 4 The following uses the example of using the "Office Smart Assistant" to reserve a meeting room to explain the operation process.

[0097] The following is the historical conversation between the user and the office smart assistant:

[0098] …

[0099] User: Please book me a meeting room for this afternoon.

[0100] Office Smart Assistant: OK, when does the meeting start and end?

[0101] User: From 3 to 5 this afternoon.

[0102] Office Smart Assistant: OK, are you booking an online or offline meeting room?

[0103] The user sends the message "A01 Meeting Room".

[0104] Since we have configured a historical message processing strategy for the office smart assistant, we will not send the "A01 conference room" directly to the big model, causing the model to be unable to process it. Instead, we will send the last five rounds of messages to the big model in full, which already contains all the key information.

[0105] Since we have configured tool retrieval for the office smart assistant, we will not provide all the tools to the big model at this time, which will make the text too long. Instead, we will first use the requests after multiple rounds of rewriting to perform tool retrieval, retrieve the two most relevant tools and provide them to the big model. Suppose the retrieved tools are "offline conference room reservation" and "online conference room reservation".

[0106] Since we have configured the React planning strategy for the office intelligent assistant, the large model performs planning, tool execution, feedback, and reflection according to the established strategy (this process is the execution process of the React planning strategy, which has no direct connection with the present invention and is not described in detail here), and finally reserves the A01 conference room for the user from 3 to 5 pm this afternoon.

[0107] The present invention can configure different planning strategies according to different usage scenarios, which can effectively improve the accuracy of artificial intelligence assistants in handling problems. It can configure different historical message processing strategies according to different usage scenarios, which can effectively improve the quality of historical message contexts provided to large models, and improve the accuracy of artificial intelligence assistants in handling problems in multi-round dialogue scenarios. Improve the accuracy of artificial intelligence assistants in handling problems in multi-round dialogue scenarios. Configurable tool retrieval strategies can increase the number of tools that artificial intelligence assistants can use and improve the effect of artificial assistant planning and arrangement.

[0108] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A strategy-configurable large-model artificial intelligence assistant, characterized in that: include: A planning strategy configuration module (1), used to configure different types of planning strategies and corresponding configuration parameters; A historical message processing strategy configuration module (2), used to configure different types of historical message processing strategies and corresponding configuration parameters; The tool search configuration module (3) is used to configure the quantity and relevance threshold of the tool searches.

2. A strategy-configurable large-model artificial intelligence assistant according to claim 1, characterized in that: The planning strategy configuration module (1) includes a React strategy, a PS strategy, a TOT strategy and a GOT strategy.

3. A strategy-configurable large-model artificial intelligence assistant according to claim 1 or 2, characterized in that: The historical message processing strategy configuration module (2) includes a truncation strategy, a summary strategy and a retrieval strategy.

4. A strategy-configurable large-model artificial intelligence assistant according to any one of claims 1 to 3, characterized in that: The tool search configuration module (3) can search for the tool most relevant to the context according to the configured parameters.

5. A method for using the strategy-configurable large-model artificial intelligence assistant as described in claims 1 to 4, characterized in that: The following steps are involved: S1. Select and configure at least one planning strategy based on the specific usage scenario; S2. Select and configure at least one historical message processing strategy based on the conversation history and context requirements; S3. Configure the number and relevance threshold of tool retrieval based on the conversation context and task requirements; S4, interacting with the user through the artificial intelligence assistant and receiving the task instructions input by the user; S5. Plan and arrange the task instructions using the configured planning strategy; S6. Use the configured historical message processing strategy to process historical messages in multiple rounds of dialogue to generate context information suitable for large model processing; S7, using the configured tool retrieval strategy to retrieve the tool most relevant to the current task from the tool set; S8, driving the large model to execute tasks according to the planning results and the retrieved tools, and generating task execution results; S9. Feedback the task execution result to the user.

6. A method according to claim 5, characterized in that The planning strategy selected in step S1 may be at least one of the React strategy, the PS strategy, the TOT strategy or the GOT strategy. The task planning and scheduling algorithm formula is: Among them, P is the optimal planning result, ρ is the set of all possible plans, Ui is the utility weight of the task, Utility(p i ) is the planning step p i is the utility function, λ is the trade-off coefficient between complexity and utility, and Compelxity(p) is the complexity measure of planning p.

7. A method according to claim 5 or 6, characterized in that: The historical message processing strategy selected in step S2 may be at least one of a truncation strategy, a summary strategy or a retrieval strategy. The historical message processing algorithm formula is: Among them, M t is the message being processed at time t, m t is the current message, m t-i is the i-th message in the past, ω i is the weight of the corresponding message.

8. A method according to any one of claims 5 to 7, characterized in that: The tool search strategy configured in step S3 includes setting the number of tool searches to not exceed a preset maximum value and setting a threshold for the relevance of the tool to the context. The tool search algorithm formula is: T = argmax t∈τ (Relevance(t, C)-α·Cost(t)), where T is the selected tool, τ is the set of all tools, Relevance(t, C) is the relevance of tool t to the current context C, α is the trade-off coefficient between cost and relevance, and Cost(t) is the cost of using tool t.

9. A method according to any one of claims 5 to 8, characterized in that: In step S5 to step S8, the artificial intelligence assistant uses the configured strategies and parameters to automatically complete the entire process from task reception to execution without user intervention.

10. A method according to any one of claims 5 to 9, characterized in that: The task execution result fed back to the user in step S9 includes the task completion status, result details and possible follow-up operation suggestions.

Images