Intelligent task planning and execution method and system for interactive command line tools

Abstract

The application provides an intelligent task planning and execution method and system for an interactive command line tool, comprising the following steps: S1, using a first large language model to generate a corresponding command sequence template for each operation of the target tool, and constructing an executable knowledge base; S2, calculating the semantic similarity between the task description and each command sequence template to obtain a candidate set, and using a second large language model to select the most relevant template as a retrieval context; S3, modeling the task execution process as a state graph search problem, selecting a candidate node from the reachable nodes of the current node based on an algorithm, and using the second large language model to select the optimal node according to the current state and the retrieval context; S4, executing the execution action corresponding to the optimal node, using the second large language model to parse the tool output, adjusting the heuristic value of the corresponding node and executing the corresponding rollback strategy; and S5, repeating steps S3 to S4 until the reasoning result shows that the task reaches the target node or the number of attempts exceeds the preset threshold.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to an intelligent task planning and execution method and system for interactive command-line tools. Background Technology

[0002] Large Language Models (LLMs) have demonstrated powerful capabilities in natural language understanding and generation, and have been widely applied in dialogue systems, code generation, and question answering. In recent years, researchers have begun exploring the use of LLMs as the core inference engine for intelligent agents, enabling them to interact with external tools and environments and automatically complete complex tasks. These LLM-based intelligent agent systems show great application potential in software development, data analysis, and system maintenance.

[0003] Interactive command-line tools are crucial interfaces for computer system management and operation, widely used in specialized fields such as database management, container orchestration, network configuration, and security testing. These tools typically feature complex command syntax, rich parameter configurations, and diverse output formats. Traditionally, using these tools requires professionals with deep domain knowledge and extensive operational experience. Automating the operation of interactive command-line tools using large language model agents can significantly lower the professional barrier, improve work efficiency, and reduce human error.

[0004] However, existing tools based on large language models face the following core technical challenges in operating intelligent agents:

[0005] First, there's the gap between semantic understanding and executable commands. Large language models primarily learn general language knowledge during pre-training, lacking accurate understanding of the precise usage of domain-specific tools. When the model attempts to generate tool commands, it's prone to "illusion": the generated commands appear grammatically correct but contain non-existent parameter names, incorrect command formats, or incompatible option combinations. This gap between semantic understanding and executable commands leads to a low success rate in automating tasks.

[0006] Second, there is the challenge of multi-step task planning. Complex tasks typically require the execution of multiple interdependent command steps, with the correct execution of subsequent commands depending on the successful completion of preceding commands. Existing large language model agents often employ linear decision frameworks such as ReAct, selecting the next action at each step based on current observations. This approach lacks an understanding of the overall task structure and an effective backtracking mechanism: when a step fails, the agent struggles to systematically explore alternative solutions, easily falling into a vicious cycle of repeatedly trying the same failed path.

[0007] Third, there is the difficulty in parsing and judging the status of tool output. Command-line tools output complex and varied information, including status displays, execution results, warning messages, and error messages. Correctly parsing this output and judging the task execution status accordingly is crucial for achieving closed-loop control. However, while large language models excel at high-level reasoning, they perform poorly in handling the subtle differences in tool output, struggling to accurately distinguish between different states such as "command syntax error," "command executed successfully but task not completed," and "task fully successful," leading to subsequent decision-making errors.

[0008] Fourth, there is the disconnect between knowledge retrieval and task execution. Retrieval-Augmented Generation (RAG) technology enhances model capabilities by retrieving relevant information from external knowledge bases, mitigating the illusion problem to some extent. However, traditional RAG methods typically retrieve descriptive text information, such as tool manuals or user instructions. The model still needs to convert the retrieved text information into executable commands, a conversion process that can still introduce errors. Furthermore, there is a lack of systematic solutions on how the retrieved knowledge can effectively guide task planning.

[0009] In summary, how to construct an intelligent task execution system that can directly retrieve executable knowledge, perform effective planning in complex state spaces, and adaptively adjust strategies based on tool feedback is a core problem that has not yet been solved in the field of artificial intelligence technology. Summary of the Invention

[0010] The purpose of this invention is to propose an intelligent task planning and execution method and system for interactive command-line tools, in order to solve the problems of semantic gap, planning difficulties, inaccurate state judgment, and disconnect between knowledge retrieval and execution in large language models in tool operation tasks.

[0011] To achieve the above objectives, the technical solution of the present invention is as follows:

[0012] A method for intelligent task planning and execution for interactive command-line tools includes:

[0013] S1. Construct an executable knowledge base: Utilize the first major language model to generate corresponding command sequence templates for each operation of the target tool based on its source code and metadata;

[0014] S2. Knowledge Retrieval and Context Construction: Calculate the semantic similarity between the task description of the current task and each command sequence template in the executable knowledge base, extract the top K command sequence templates with the highest similarity as a candidate set, and use the second largest language model to select the command sequence template most relevant to the current task from the candidate set as the retrieval context.

[0015] S3. Retrieval Enhancement Planning: Model the current task execution process as a state graph search problem, using the discrete states of the tool as nodes and state transition relationships as edges. Starting from the initial state, iteratively search subsequent nodes; for each current node, based on... The algorithm selects k candidate nodes from the reachable nodes of the current node, and uses the second largest language model to select the optimal node from the candidate nodes based on the current tool state, retrieval context and historical execution records;

[0016] S4. Closed-loop execution and feedback reasoning: Execute the action corresponding to the optimal node and obtain the tool output; use the second language model to perform structured parsing and state reasoning on the tool output, adjust the heuristic value of the corresponding node according to the reasoning result and execute the corresponding backoff strategy.

[0017] S5. Iterative Loop: Repeat steps S3 to S4 until the inference result shows that the task has reached the target node or the number of attempts exceeds the preset threshold.

[0018] Preferably, the construction of the executable knowledge base specifically includes:

[0019] Obtain the source code and metadata of all operation modules of the target tool, including command identifiers, required parameter sets, optional parameter sets, default values, applicable conditions, and expected output;

[0020] Based on the source code and metadata of the target tool, the first major language model is used to generate a corresponding original command sequence template for each operation of the target tool. Each original command sequence template corresponds to multiple command sequences, and placeholders are used to represent dynamic parameters.

[0021] The original command sequence template is converted into a high-dimensional vector representation through an embedding model. The obtained command sequence template is then stored in a vector database and a semantic index is established.

[0022] Preferably, the construction of the state diagram includes:

[0023] Define a set of nodes to represent the discrete states of the tool, including initial state, configuration state, execution state, success state, and failure state;

[0024] Define an edge set to represent the state transition relationship;

[0025] Edge weights are defined to represent the costs of different types of state transitions, and are predefined based on expert experience.

[0026] Preferably, for each current node, based on The algorithm selects k candidate nodes from the reachable nodes of the current node, as follows:

[0027] Based on the node type, each node in the state diagram Pre-set basic heuristic values ; and set initial dynamic heuristic values ​​for each node. : ,in Here, α is the confidence weight parameter, 0 < α ≤ 1. For nodes The search confidence level;

[0028] A closed-loop dynamic adjustment mechanism based on time steps and feedback coefficients is introduced to update the corresponding nodes according to the closed-loop execution and feedback inference results of the previous iteration. Heuristic values: ,in, Represents a node Updated heuristic values, Represents a node Heuristic value before update These are update coefficients based on execution feedback;

[0029] based on The algorithm's node evaluation function calculates the reachable nodes from the starting node. Total estimated cost to reach the target node :

[0030]

[0031]

[0032] in, From the starting node to the reachable node They have already paid the price. From the starting node to the current node The actual cost, To start from the current node To reach the node The cost of a single-step transition, i.e., the corresponding edge weight; For the updated reachable nodes The heuristic value;

[0033] For all reachable nodes Corresponding total estimated cost Sort the nodes in ascending order and select the top k reachable nodes corresponding to the total estimated cost as candidate nodes.

[0034] Preferably, the method for determining the retrieval confidence is as follows:

[0035] According to the node The state characteristics determine the required knowledge type and construct the query vector;

[0036] Retrieve nodes from the executable knowledge base The required knowledge type matches the knowledge fragment;

[0037] If the retrieval is successful, calculate the distance between the query vector and the vector of each retrieved knowledge fragment. and will As a semantic similarity metric, if the highest semantic similarity score exceeds a preset threshold, then... Assign the highest semantic similarity; otherwise, Set to the preset low confidence level.

[0038] Preferably, the step of using the second language model to perform structured parsing and state reasoning on the tool output specifically includes:

[0039] The second largest language model is used to identify the status of command execution results;

[0040] If the task is in a successful execution state, the second largest language model is used to evaluate the task progress and determine whether the current tool output indicates that the task is progressing toward the target state.

[0041] If the execution fails, the error message output by the second language model parsing tool will be used to classify the cause of failure as parameter error, status error, environment error, permission error, resource error, network error, or other error.

[0042] Preferably, the heuristic value of the corresponding node is adjusted based on the reasoning result, and the corresponding backoff strategy is executed, specifically including:

[0043] When the command executes successfully and the task progresses towards the target state, the update coefficient is set. , This represents the decay rate of the heuristic value upon successful execution, and Reduce the heuristic value of executed nodes to reinforce the success path;

[0044] When the command executes successfully but the task does not progress, set the update coefficient. Keep the current heuristic value of the executed node unchanged;

[0045] When a command fails, the nodes whose heuristic values ​​need to be updated are determined based on the failure category, and the update coefficients are set accordingly. , This indicates the percentage increase in the heuristic value when execution fails, and Increase the heuristic value of the corresponding node; and execute the corresponding rollback strategy according to the failure category, updating the current node to the rolled-back position.

[0046] An intelligent task planning and execution system for interactive command-line tools includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs any of the steps in the aforementioned intelligent task planning and execution method for interactive command-line tools.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. Eliminate the semantic-to-command conversion gap: An executable knowledge base is built offline using the first language model, and combined with the online optimization mechanism of the second language model, command sequence templates that have been double-verified (vector similarity + semantic relevance) are directly retrieved. This eliminates the need for the online model to perform semantic-to-command conversion from scratch, significantly improving the accuracy and usability of command generation.

[0049] 2. Implement structured multi-step task planning: utilizing... Heuristic search algorithms provide structured modeling of the task execution process, and combined with the semantic understanding capabilities of the second-largest language model, they support systematic path exploration and failure backtracking.

[0050] 3. Precise tool output parsing and status judgment: Through the multi-step reasoning capability of the second language model, the tool output is structured and parsed to accurately identify the execution status and failure type, providing a reliable basis for subsequent decision-making.

[0051] 4. Deep integration of knowledge retrieval and planning execution: Using retrieval confidence as a moderating factor in the heuristic function, the retrieved executable knowledge directly guides task planning, achieving the integration of knowledge and planning. Attached Figure Description

[0052] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0053] The following is in conjunction with the appendix Figure 1 The technical solution of the present invention will be described in detail below.

[0054] This invention proposes an intelligent task planning and execution method for interactive command-line tools, comprising:

[0055] S1. Construct an executable knowledge base: Utilize the first major language model to generate corresponding command sequence templates for each operation of the target tool based on its source code and metadata;

[0056] S2. Knowledge Retrieval and Context Construction: Calculate the semantic similarity between the task description of the current task and each command sequence template in the executable knowledge base (vectorize the task description to construct the retrieval query, calculate the cosine similarity between the retrieval query and the command sequence template), extract the top K command sequence templates with the highest similarity as a candidate set, and use the second language model to select the command sequence template most relevant to the current task from the candidate set as the retrieval context (the prompt words contain the task description and K candidate templates, and the second language model outputs the index of the most matching template).

[0057] S3. Retrieval Enhancement Planning: Model the current task execution process as a state graph search problem, using the discrete states of the tool as nodes and state transition relationships as edges. Starting from the initial state, iteratively search subsequent nodes; for each current node, based on... The algorithm selects k candidate nodes from the reachable nodes of the current node, and uses the second largest language model to select the optimal node from the candidate nodes based on the current tool state, retrieval context and historical execution records;

[0058] S4. Closed-loop execution and feedback reasoning: Execute the action corresponding to the optimal node and obtain the tool output; use the second language model to perform structured parsing and state reasoning on the tool output, adjust the heuristic value of the corresponding node according to the reasoning result and execute the corresponding backoff strategy.

[0059] S5. Iterative Loop: Repeat steps S3 to S4 until the inference result shows that the task has reached the target node or the number of attempts exceeds the preset threshold.

[0060] In this embodiment, the construction of the executable knowledge base specifically includes:

[0061] Obtain the source code and metadata of all operation modules of the target tool, including command identifiers, required parameter sets, optional parameter sets, default values, applicable conditions, and expected output;

[0062] Based on the source code and metadata of the target tool, the first major language model is used to generate corresponding original command sequence templates for each operation of the target tool. Each original command sequence template corresponds to multiple command sequences, and placeholders are used to represent dynamic parameters, which are replaced with actual values ​​at runtime according to the specific task context.

[0063] The original command sequence template is converted into a high-dimensional vector representation through an embedding model. The obtained command sequence template is stored in a vector database and a semantic index is established. The embedding model adopts a pre-trained language model based on the Transformer architecture.

[0064] In this embodiment, the construction of the state diagram includes:

[0065] Define a set of nodes to represent the discrete states of the tool, including initial state, configuration state, execution state, success state, and failure state;

[0066] Define an edge set to represent the state transition relationship;

[0067] Edge weights are defined to represent the costs of different types of state transitions, and are predefined based on expert experience.

[0068] In this embodiment, for each current node, based on The algorithm selects k candidate nodes from the reachable nodes of the current node, as follows:

[0069] Based on the node type, each node in the state diagram Pre-set basic heuristic values ; and set initial dynamic heuristic values ​​for each node. : ,in Here, α is the confidence weight parameter, 0 < α ≤ 1. For nodes The search confidence level;

[0070] A closed-loop dynamic adjustment mechanism based on time steps and feedback coefficients is introduced to update the corresponding nodes according to the closed-loop execution and feedback inference results of the previous iteration. Heuristic values: ,in, Represents a node Updated heuristic values, Represents a node Heuristic value before update These are update coefficients based on execution feedback;

[0071] based on The algorithm's node evaluation function calculates the reachable nodes from the starting node. Total estimated cost to reach the target node :

[0072]

[0073]

[0074] in, From the starting node to the reachable node They have already paid the price. From the starting node to the current node The actual cost, To start from the current node To reach the node The cost of a single-step transition, i.e., the corresponding edge weight; For the updated reachable nodes The heuristic value;

[0075] For all reachable nodes Corresponding total estimated cost Sort the nodes in ascending order and select the top k reachable nodes corresponding to the total estimated cost as candidate nodes.

[0076] In this embodiment, the method for determining the retrieval confidence level is as follows:

[0077] According to the node The state characteristics determine the required knowledge type and construct the query vector;

[0078] Retrieve nodes from the executable knowledge base The required knowledge type matches the knowledge fragment;

[0079] If the retrieval is successful, calculate the distance between the query vector and the vector of each retrieved knowledge fragment. and will As a semantic similarity metric, if the highest semantic similarity score exceeds a preset threshold, then... Assign the highest semantic similarity; otherwise, Set to the preset low confidence level.

[0080] In this embodiment, the step of using the second largest language model to perform structured parsing and state reasoning on the tool output specifically includes:

[0081] The second largest language model is used to identify the status of command execution results;

[0082] If the task is in a successful execution state, the second largest language model is used to evaluate the task progress and determine whether the current tool output indicates that the task is progressing toward the target state.

[0083] If the execution fails, the error message output by the second language model parsing tool will be used to classify the cause of failure as parameter error, status error, environment error, permission error, resource error, network error, or other error.

[0084] In this embodiment, adjusting the heuristic value of the corresponding node based on the reasoning result and executing the corresponding backoff strategy specifically includes:

[0085] When the command executes successfully and the task progresses towards the target state, the update coefficient is set. , This represents the decay rate of the heuristic value upon successful execution, and Reduce the heuristic value of executed nodes to reinforce the success path (e.g., );

[0086] When the command executes successfully but the task does not progress, set the update coefficient. Keep the current heuristic value of the executed node unchanged;

[0087] When a command fails, the nodes whose heuristic values ​​need to be updated are determined based on the failure category, and the update coefficients are set accordingly. , This indicates the percentage increase in the heuristic value when execution fails, and Increase the heuristic value of the corresponding node; and execute the corresponding rollback strategy according to the failure category, updating the current node to the rolled-back position.

[0088] The following table provides exemplary adjustment and rollback schemes:

[0089]

[0090] An intelligent task planning and execution system for interactive command-line tools includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs any of the steps in the aforementioned intelligent task planning and execution method for interactive command-line tools.

[0091] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for intelligent task planning and execution for interactive command-line tools, characterized in that, include: S1. Construct an executable knowledge base: Utilize the first major language model to generate corresponding command sequence templates for each operation of the target tool based on its source code and metadata; S2. Knowledge Retrieval and Context Construction: Calculate the semantic similarity between the task description of the current task and each command sequence template in the executable knowledge base, extract the top K command sequence templates with the highest similarity as a candidate set, and use the second largest language model to select the command sequence template most relevant to the current task from the candidate set as the retrieval context. S3. Retrieval Enhancement Planning: Model the current task execution process as a state graph search problem, using the discrete states of the tool as nodes and state transition relationships as edges. Starting from the initial state, iteratively search subsequent nodes; for each current node, based on... The algorithm selects k candidate nodes from the reachable nodes of the current node, and uses the second largest language model to select the optimal node from the candidate nodes based on the current tool state, retrieval context and historical execution records; S4. Closed-loop execution and feedback reasoning: Execute the action corresponding to the optimal node and obtain the tool output; use the second language model to perform structured parsing and state reasoning on the tool output, adjust the heuristic value of the corresponding node according to the reasoning result and execute the corresponding backoff strategy. S5. Iteration loop: Repeat steps S3 to S4 until the inference result shows that the task has reached the target node or the number of attempts exceeds the preset threshold. For each current node, based on The algorithm selects k candidate nodes from the reachable nodes of the current node, as follows: Based on the node type, each node in the state diagram Pre-set basic heuristic values ; and set initial dynamic heuristic values ​​for each node. : ,in Here, α is the confidence weight parameter, 0 < α ≤ 1. For nodes The search confidence level; A closed-loop dynamic adjustment mechanism based on time steps and feedback coefficients is introduced to update the corresponding nodes according to the closed-loop execution and feedback inference results of the previous iteration. Heuristic values: ,in, Represents a node Updated heuristic values, Represents a node Heuristic value before update These are update coefficients based on execution feedback; based on The algorithm's node evaluation function calculates the reachable nodes from the starting node. Total estimated cost to reach the target node : in, From the starting node to the reachable node They have already paid the price. From the starting node to the current node The actual cost, To start from the current node To reach the node The cost of a single-step transition, i.e., the corresponding edge weight; For the updated reachable nodes The heuristic value; For all reachable nodes Corresponding total estimated cost Sort the nodes in ascending order and select the top k reachable nodes corresponding to the total estimated cost as candidate nodes.

2. The intelligent task planning and execution method for interactive command-line tools according to claim 1, characterized in that, The construction of the executable knowledge base specifically includes: Obtain the source code and metadata of all operation modules of the target tool, including command identifiers, required parameter sets, optional parameter sets, default values, applicable conditions, and expected output; Based on the source code and metadata of the target tool, the first major language model is used to generate a corresponding original command sequence template for each operation of the target tool. Each original command sequence template corresponds to multiple command sequences, and placeholders are used to represent dynamic parameters. The original command sequence template is converted into a high-dimensional vector representation through an embedding model. The obtained command sequence template is then stored in a vector database and a semantic index is established.

3. The intelligent task planning and execution method for interactive command-line tools according to claim 1, characterized in that, The construction of the state diagram includes: Define a set of nodes to represent the discrete states of the tool, including initial state, configuration state, execution state, success state, and failure state; Define an edge set to represent the state transition relationship; Edge weights are defined to represent the costs of different types of state transitions, and are predefined based on expert experience.

4. The intelligent task planning and execution method for interactive command-line tools according to claim 1, characterized in that, The method for determining the retrieval confidence level is as follows: According to the node The state characteristics determine the required knowledge type and construct the query vector; Retrieve nodes from the executable knowledge base The required knowledge type matches the knowledge fragment; If the retrieval is successful, calculate the distance between the query vector and the vector of each retrieved knowledge fragment. and will As a semantic similarity metric, if the highest semantic similarity score exceeds a preset threshold, then... Assign the highest semantic similarity; otherwise, Set to the preset low confidence level.

5. The intelligent task planning and execution method for interactive command-line tools according to claim 1, characterized in that, The method of using the second largest language model to perform structured parsing and state reasoning on the tool output specifically includes: The second largest language model is used to identify the status of command execution results; If the task is in a successful execution state, the second largest language model is used to evaluate the task progress and determine whether the current tool output indicates that the task is progressing toward the target state. If the execution fails, the error message output by the second language model parsing tool will be used to classify the cause of failure as parameter error, status error, environment error, permission error, resource error, network error, or other error.

6. The intelligent task planning and execution method for interactive command-line tools according to claim 5, characterized in that, The step of adjusting the heuristic values ​​of the corresponding nodes based on the reasoning results and executing the corresponding backoff strategy specifically includes: When the command executes successfully and the task progresses towards the target state, the update coefficient is set. , This represents the decay rate of the heuristic value upon successful execution, and Reduce the heuristic value of executed nodes to reinforce the success path; When the command executes successfully but the task does not progress, set the update coefficient. Keep the current heuristic value of the executed node unchanged; When a command fails, the nodes whose heuristic values ​​need to be updated are determined based on the failure category, and the update coefficients are set accordingly. , This indicates the percentage increase in the heuristic value when execution fails, and Increase the heuristic value of the corresponding node; and execute the corresponding rollback strategy according to the failure category, updating the current node to the rolled-back position.

7. An intelligent task planning and execution system for interactive command-line tools, characterized in that, It includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the intelligent task planning and execution method for interactive command-line tools as described in any one of claims 1-6.

Images