Method for calling skills based on skill map

Abstract

This invention discloses a skill invocation method based on a skill graph, comprising: constructing a skill association graph, defining skill nodes, skill step nodes, tool nodes, tool input parameter nodes, tool output parameter nodes, intent nodes, question nodes, and edge relationships between nodes; receiving and parsing user questions, and jointly recalling candidate nodes based on multiple node types; generating candidate skill paths based on the candidate nodes and a hierarchical graph using heuristic search, and performing constraint verification and scoring on the candidate skill paths to determine the target skill path, selecting the target skill path to execute the skill invocation; replanning local paths when the target skill path fails to execute; and dynamically updating the skill association graph based on the execution results. This invention unifies the complex relationships between skills, tools, intents, and questions into a single model, achieving intelligent and precise processing across the entire link from user questions to skill invocation.

Description

Technical Field

[0001] This invention belongs to the technical field of skill invocation in business systems, and specifically relates to a skill invocation method based on skill graphs. Background Technology

[0002] With the rapid development of large language models, intelligent agent systems, and automated tool invocation technologies, more and more business systems are beginning to open up their business capabilities to models in the form of skills. However, with the continuous growth in the number of skills, the management and invocation of skills face serious challenges.

[0003] Currently, skill retrieval primarily employs two approaches: First, a retrieval-augmented generation (RAG)-based approach, which retrieves relevant skills from a skill library based on the semantic similarity to the user's question and recommends them to the model. Second, all skills are directly input into the model, allowing the model to automatically determine and select the required skills. Both approaches have significant drawbacks: RAG retrieval relies solely on semantic similarity for matching, making it difficult to accurately understand the deep correspondence between user intent and skill capabilities, resulting in insufficient skill retrieval accuracy and frequent mismatches between retrieved skills and user needs. Conversely, directly inputting all skills into the model leads to substantial token consumption, increasing computational costs, and the model is inefficient when faced with a massive number of skills, easily losing crucial information due to context length limitations.

[0004] Furthermore, existing solutions lack a structured representation of the complex relationships between skills, the skill invocation chain is lengthy and invisible, it is difficult to perform effective local repairs and dynamic adjustments when skill execution fails, and it is also impossible to continuously optimize the skill matching strategy based on historical execution experience. Summary of the Invention

[0005] This invention provides a skill invocation method based on skill graphs to solve the aforementioned technical problems, specifically adopting the following technical solution: A skill invocation method based on a skill graph includes the following steps: Construct a skill association graph, defining skill nodes, skill step nodes, tool nodes, tool input parameter nodes, tool output parameter nodes, intent nodes, problem nodes, and the edge relationships between each node; Receive and analyze user questions, and jointly recall candidate nodes based on multiple node types; Based on the candidate nodes, a heuristic search is performed using a hierarchical graph to generate candidate skill paths. The candidate skill paths are then constrained, validated, and scored to determine the target skill path. Finally, the target skill path is selected to execute the skill call. When the target skill path fails to execute, a partial path replanning is performed; The skill association graph is dynamically updated based on the execution results.

[0006] Furthermore, a skill association graph is constructed, including: extracting skill information, skill step information, tool information, and parameter information from skill files and importing them into corresponding nodes; establishing edge relationships between skill nodes and intent nodes, skill nodes and tool nodes, and tool nodes and parameter nodes; and vectorizing the node descriptions.

[0007] Furthermore, constructing a skill association graph also includes: identifying and classifying the intent of skill descriptions, and merging new intents with existing intents or creating new intent nodes based on intent merging scores; The intent merging score is calculated based on vector similarity, text similarity, keyword similarity, and rule consistency.

[0008] Furthermore, it receives and parses user questions, including: Analyze user questions to obtain a question summary, user intent, target action, business object, known parameters, missing parameters, time and scope conditions, permissions and compliance conditions, expected output, and prohibited operations; Candidate nodes are jointly recalled from multiple node types based on recall scores; The recall score is calculated based on vector similarity, keyword matching degree, text semantic similarity, business object matching degree, task action matching degree, and business rule matching degree.

[0009] Furthermore, when the combined matching score between the historical problem and the current problem reaches a high confidence threshold, the historical problem is associated with a verified skill path, the corresponding skills and tools are in a valid state, the parameters of the current problem are compatible with the historical problem, and there is no hard failure flag in the historical path, the corresponding historical skill path is taken as the preferred candidate path.

[0010] Furthermore, heuristic searches based on hierarchical graphs include: The skill association graph is divided into the problem and intent layer, the skill ability layer, the execution step layer, and the tool and parameter layer in sequence. After determining the candidate skill combinations by searching for candidate nodes in the problem and intent layer and the skill capability layer, the system then enters the execution step layer and the tool and parameter layer to search for skill steps, tools and parameters for the candidate skill combinations in order to generate candidate skill paths.

[0011] Furthermore, based on the constraints, multiple skill nodes in the skill capability layer are aggregated into skill clusters to form a skill cluster layer; first, the corresponding skill cluster is searched in the skill cluster layer, and then the multiple skill nodes contained in the selected skill cluster are retrieved and sorted.

[0012] Furthermore, the recalled candidate nodes are used to form a set of recall anchor points; The necessary set of completion nodes is determined based on the set of recall anchor points, and then the actual set of nodes allowed to be searched is determined. While satisfying the requirements of skill integrity, dependency, and parameter passing, access to the necessary completion node set outside the recall anchor set is allowed.

[0013] Furthermore, when there is a stable and verified intermediate sub-path between two non-adjacent level nodes, the system establishes a cross-level jump edge and performs a jump edge search.

[0014] Furthermore, heuristic search employs a search function: f(n) = g(n) + h(n), Where g(n) represents the cumulative cost from the initial state to the current state, and h(n) represents the expected remaining cost from the current state to completing the task objective. h(n) is calculated based on the similarity between the current node and the user's task objective, the number of necessary constraints that have not yet been met, the expected failure risk, the number of uncovered task objectives, and the expected remaining search depth.

[0015] Furthermore, constraint validation is performed on the candidate skill paths, including: The feasibility of candidate paths is determined based on a set of hard constraints, which include: skill prerequisites, step order, required tool parameters, calling permissions, input and output types of tools before and after, completeness of necessary steps, nodes that are prohibited from being called, and the current business status. Candidate skill paths that violate at least one non-relaxable hard constraint are either deleted or enter a path repair process. The repair cost is calculated for candidate skill paths that are subject to repair based on a set of repairable constraints. The repairable constraints include: missing parameters, precondition constraints, and format constraints.

[0016] Furthermore, candidate skill paths are scored based on fundamental utility and path penalty. Basic utilities include: recall node coverage, intent matching, skill matching, step completeness, tool and parameter suitability, constraint satisfaction, historical reliability, and output coverage. Path penalties include: repair cost, path length penalty, irrelevance penalty, execution risk, expected cost, expected delay, and failure penalty.

[0017] Furthermore, identifying the target skill path includes: Within the allowed score difference range, select the path with the higher score to construct a high-quality candidate path set; The Temperature parameter is dynamically determined based on task risk, candidate path differences, and path historical stability. The selection strategy is determined based on the Temperature parameter: when Temperature is 0, the path with the highest score is selected; when Temperature is greater than 0, the target path is selected from the set of high-quality candidate paths according to the probability distribution.

[0018] Furthermore, local path replanning includes: marking failed paths as failures, generating a hard prohibition set and a soft penalty set, and calculating failure penalties for nodes and edges in the soft penalty set; The failed paths are split up, the prefixes of the successfully executed paths are retained, and parameter repair or equivalent tool replacement is performed according to the failure type. Candidate repair paths are generated based on the objective function of the repair path. The objective function is calculated based on the task objective coverage, output reusability, reliability, constraint satisfaction, failure risk, structural differences between the repair path and the original remaining path, cost of adding calls, and cost of rolling back executed steps.

[0019] Furthermore, the equivalent tool replacement is based on the alternative tool adaptation score to select the alternative tool. The alternative tool adaptation score is based on capability similarity, input parameter compatibility, output result compatibility, historical reliability of the alternative tool, permission adaptation, current context failure risk, and tool switching cost calculation.

[0020] Furthermore, dynamic updates include: when a user confirms the recommended path or the path is executed successfully, the user's question is created as a question node and added to the skill association graph; based on the differences between the user's confirmed path and the recommended path, as well as the path execution statistics, the question skill matching model and path strategy parameters are updated.

[0021] The advantages of the skill invocation method based on skill graphs in this invention are: This invention constructs a structured skill association graph, unifying the complex relationships between skills, tools, intentions, and questions. It combines multi-type node joint recall, hierarchical heuristic search, constraint verification and comprehensive scoring, failure-aware local replanning, and dynamic graph updates to achieve intelligent end-to-end processing from user questions to skill invocation. Compared to traditional RAG systems, which suffer from insufficient skill recall accuracy or excessive token consumption due to full skill input, this invention significantly improves skill matching accuracy, simplifies the skill invocation chain, reduces recovery costs after execution failures, and continuously improves system performance through continuous learning and optimization. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the skill invocation method based on skill graphs according to this application. Detailed Implementation

[0024] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0025] In the description of this application, it should be noted that, unless otherwise specified and limited, the terms "installation", "connection" and "linkage" should be interpreted broadly, and can refer to mechanical or electrical connections, or internal connections between two components, or direct connections. "Up", "down", "left", "right", etc., are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may change.

[0026] This application provides a skill invocation method based on a skill graph, comprising the following steps: Step 1: Construct a skill graph, defining skill nodes, skill step nodes, tool nodes, tool input parameter nodes, tool output parameter nodes, intent nodes, question nodes, and the edge relationships between each node; Step 2: Receive and parse user questions, and jointly recall candidate nodes based on multiple node types; Step 3: Based on the candidate nodes, perform heuristic search to generate candidate skill paths, perform constraint verification and scoring on the candidate skill paths, determine the target skill path, and select the target skill path to execute the skill call; Step 4: Perform local replanning when the target skill path fails. Step 5: Dynamically update the skill association graph based on the execution results.

[0027] Based on the steps described above, this application constructs a structured skill association graph, unifying the complex relationships between skills, tools, intentions, and questions through modeling. It combines multi-type node joint recall, hierarchical heuristic search, constraint verification and comprehensive scoring, failure-aware local replanning, and dynamic graph updates to achieve intelligent end-to-end processing from user questions to skill invocation. Compared to the problems of insufficient accuracy in traditional RAG skill recall or excessive token consumption due to full skill input, this invention significantly improves the accuracy of skill matching, simplifies the skill invocation chain, reduces the repair cost after execution failure, and continuously improves system performance through continuous learning and optimization. The specific implementation methods of this application are as follows: For step 1, a skill graph is constructed, defining skill nodes, skill step nodes, tool nodes, tool input parameter nodes, tool output parameter nodes, intent nodes, question nodes, and the edge relationships between these nodes. In other words, the system uniformly models skills, skill steps, tools, tool parameters, user intents, and user questions, forming a graph structure that includes skill nodes, skill step nodes, tool nodes, tool input parameter nodes, tool output parameter nodes, intent nodes, and question nodes. Specifically: Skill nodes are used to describe the tasks or business capabilities that the system can perform; Skill step nodes are used to describe skill steps, and multiple steps are connected by edges; Tool nodes represent executable capabilities, including executable files, functions, API interfaces, external services, etc. The tool input parameter node is used to describe information such as parameter name, type, meaning, and whether it is required; The tool output parameter node is used to describe the result name, type, meaning, and data structure; Intent nodes represent the task objectives achieved by the skill and are used for semantic association with executable skills; The issue node is used to record the issue text, associating the user's issue with their intent.

[0028] Step 1 establishes a structured association between skills, tools, intentions, and problems by defining the resolution relationships between skill nodes and intent nodes, the prerequisite and successor relationships between skill nodes, the invocation relationships between skill nodes and skill step nodes and tool nodes, the inclusion relationships between skill nodes and skill step nodes, the Next relationships between skill step nodes, the association relationships between problem nodes and intent nodes, the input-output relationships between tool nodes and tool input and output parameter nodes, and the Next relationships between tool input and output parameter nodes. Specifically: Skill nodes resolve intent nodes, indicating that skills can resolve user intents and are used to query related skills from intents; The preceding skill node of a skill node indicates that one skill must be executed before another skill, and is used to describe the preceding dependency constraints between skills; The successor skill node of a skill node represents the subsequent skill that can be followed after a skill is executed, and is used to describe the execution continuity relationship between skills; A skill node calling a tool node indicates that a skill depends on a tool to complete its ability. A skill step node calling a tool node indicates that a skill depends on a certain tool to complete the realization of its ability; Skill nodes contain skill step nodes, indicating that process-type skills consist of multiple steps, and are used to expand the internal structure of the skill. The Next skill step node of a skill step node indicates the execution order between skill steps and is an important constraint for process orchestration. The issue node is associated with the intent node, indicating that the user's issue is related to one or more intents, which is used to accumulate historical issues and intent matching experience; The tool node is an input parameter node, indicating that the tool requires specified input parameters to execute; The tool node outputs the tool's parameters, indicating the specified output results that will be generated after the tool completes execution. The Next tool output node of the tool input parameter node indicates that the tool input parameters will affect or generate corresponding output results after being processed by the tool.

[0029] This solution transforms scattered skill file information into structured graph data, making the dependencies, execution order, and semantic relationships between skills explicit. This provides a unified data foundation for subsequent intelligent retrieval and path search, solving the problem of management difficulties as the number of skills increases.

[0030] Furthermore, a skill association graph is constructed, including: extracting skill information, skill step information, tool information, and parameter information from skill files and importing them into corresponding nodes; establishing edge relationships between skill nodes and intent nodes, skill nodes and tool nodes, and tool nodes and parameter nodes; and vectorizing the node descriptions. Specifically: A list of skills for which a skill association graph needs to be constructed is identified, and the file format of these skill lists must meet the requirements of the skill template for standardized information extraction. All skill metadata must include a name and description: process-type skills require steps as second-level headings, and within each step, the names of the required tools or interfaces, input parameter names and descriptions, and output results and descriptions must be provided; tool-type skills metadata must additionally include the tool name, with function points as second-level headings, and each function point must include a function description, input parameter names and descriptions, and output results and descriptions. Based on the structured skill files, multi-dimensional skill information is extracted for each skill to generate a structured skill profile. This structured skill profile serves as the source of points, edges, and attributes for subsequent graph construction.

[0031] Based on the skill information above, extract the skill name and skill description, and import them into the skill node using the skill name as the unique primary key. If a skill file mentions a reference (one skill file references / references another skill file), then add pre-dependency and post-dependency edges to the two skill files respectively.

[0032] For the skill step information mentioned above, extract the step title and step description, use the skill name plus the skill step number as the unique primary key, import them into the skill step node, and add edge relationships between the skill step nodes.

[0033] Based on the tool information above, extract the tool name and description, using the tool name as the unique primary key, and import them into the tool node. For each input parameter, extract the parameter name, parameter type, parameter meaning, and whether it is required, using the tool name plus the parameter name as the unique primary key, and import them into the tool input parameter type node. Simultaneously, establish input edge type relationships with the corresponding tool node.

[0034] For the parameter information mentioned above, for each output parameter, the system records a list of information including parameter name, parameter type, and parameter meaning. The output parameter nodes are set up in the format of tool name plus parameter name, and an output edge type relationship is established with the corresponding tool node.

[0035] When the output parameter of one tool is the input of the next tool, the two parameters are associated through the Next edge relationship to extract tool association information.

[0036] In other words, the system extracts skill names and descriptions from skill files and imports them into skill nodes; extracts skill names, step titles, and step descriptions and imports them into skill step nodes; extracts tool names and tool descriptions and imports them into tool nodes; extracts parameter names, parameter types, parameter meanings, and whether they are required and imports them into tool input parameter nodes and tool output parameter nodes respectively; extracts intent names and intent descriptions and imports them into intent nodes; and extracts the original question and question description and imports them into question nodes. It establishes edge relationships between skill nodes and intent nodes, skill nodes and tool nodes, and tool nodes and parameter nodes. Furthermore, it vectorizes the descriptions of skill nodes, skill step nodes, tool nodes, question nodes, and intent nodes to provide data support for subsequent vector similarity calculations. Through standardized information extraction and vectorization, the system ensures strict consistency between the graph data and skill files, providing a high-quality data foundation for subsequent semantic similarity calculations and vector retrieval, thus improving the computational efficiency and accuracy of the retrieval phase.

[0037] Furthermore, constructing a skill association graph also includes: identifying and classifying the intents of skill descriptions, and merging new intents with existing intents or creating new intent nodes based on the intent merging score (MergeScore). The intent merging score is based on vector similarity (Sim). vector Text similarity Sim text Keyword similarity Sim keyword and rule consistency Consistency The calculation, specifically the formula: , Here, b1, b2, b3, and b4 represent the weighting coefficients of each parameter. When the merged score exceeds a preset merge threshold, the new intent is merged into an existing intent node, and the corresponding skill is associated with the existing intent. If no existing intent meets the merge conditions, a new intent node is created. This multi-dimensional similarity calculation and rule consistency verification achieve automatic merging and classification of intents, avoiding redundancy and fragmentation of intent nodes, and improving the quality of intent nodes and the accuracy of intent-skill association.

[0038] For step 2, receive and analyze user questions, and jointly recall candidate nodes based on multiple node types.

[0039] The system receives user questions (q) and parses them to obtain: a question summary, user intent, target action, business object, known parameters, missing parameters, time and scope conditions, permission and compliance conditions, expected output, and prohibited operations. Then, it recalls candidate nodes from historical question nodes, intent nodes, skill nodes, skill step nodes, and tool nodes. Based on the recall score, candidate nodes are jointly recalled from multiple node types; the recall score is calculated based on vector similarity, keyword matching, text semantic similarity, business object matching, task action matching, and business rule matching.

[0040] The basic recall score for user question (q) and candidate nodes (n) is represented as follows:

[0041] Among them, Sim vector Sim represents vector similarity. keyword Indicates the degree of keyword matching, Sim text Representing text semantic similarity, ObjectMatch represents the degree of matching of business objects, ActionMatch represents the degree of matching of task actions, and RuleScore represents the degree of matching of business rules; w v w k w t w o w a w r These are the weight coefficients for the corresponding parameters. Based on the recall score, Top-K results are retained according to node type to avoid a large number of candidates for one node type crowding out the candidate space for other node types.

[0042] In this way, by using multi-dimensional similarity calculation and a categorized Top-K retention mechanism, the semantic information of user questions, business object information, and rule information can be fully utilized to avoid single-type nodes from excessively crowding out the candidate space and to improve the comprehensiveness and accuracy of candidate node recall.

[0043] Furthermore, when the following conditions are met: the overall matching score between the historical problem and the current problem reaches a high confidence threshold; the historical problem is associated with a skill path that has been confirmed by the user or verified by actual execution; the corresponding skill and tool are currently in a valid state; the parameters, business objects, and permission conditions of the current problem are compatible with the historical problem; and there is no hard failure flag in the historical path, the corresponding historical skill path will be taken as the preferred candidate path.

[0044] When key parameters, permission conditions, or business objects change, even if the problem text has a high degree of similarity, the historical path is not directly reused. Instead, the historical path is used as a candidate skeleton for graph search and the constraints are re-validated.

[0045] In this way, by reusing verified historical successful paths, the overhead of repeated searches and calculations is reduced, and the efficiency and reliability of path generation are improved, which is especially suitable for rapid response to similar problems. At the same time, parameter compatibility checks prevent incompatible historical paths from being reused incorrectly, ensuring the feasibility of reusable paths.

[0046] For step 3, based on the candidate nodes, a heuristic search is performed using a layered graph to generate candidate skill paths. These candidate skill paths are then constrained, validated, and scored to determine the target skill path. Finally, the target skill path is selected for skill invocation. Specifically: When performing heuristic search based on hierarchical graphs, the skill association graph is divided into four layers: problem and intent layer, skill and ability layer, execution step layer, and tool and parameter layer. Among them: The first layer, the problem and intent layer, includes: problem nodes, intent nodes, business objects, and task actions; The second layer of skill capabilities includes: skill nodes, skill cluster nodes, skill prerequisite relationships, and skill post-requirement relationships; The third execution step layer includes: skill step nodes, step sequence relationships, necessary steps, and branch steps; The fourth layer, the tools and parameters layer, includes: tool nodes, tool input parameter nodes, tool output parameter nodes, and parameter passing relationships.

[0047] The system first identifies candidate skill combinations in the high-level graph (first and second layers), and then develops skill steps, tools, and parameters for high-quality candidate combinations, thereby reducing indiscriminate searches in the complete graph.

[0048] After identifying candidate skill combinations by searching for candidate nodes in the problem and intent layer and the skill capability layer, the algorithm then proceeds to the execution step layer and the tool and parameter layer to search for skill steps, tools, and parameters for these combinations, thereby generating candidate skill paths. This solution employs a hierarchical search strategy, first identifying candidate skill combinations in a high-level graph, and then expanding the skill steps, tools, and parameters for high-quality candidate combinations. This reduces indiscriminate searching within the complete graph, lowers search complexity, and improves path generation efficiency.

[0049] In one specific implementation, step 3 aggregates multiple skill nodes from the skill capability layer into skill clusters based on constraints, forming a skill cluster layer. First, the corresponding skill cluster is searched within the skill cluster layer, and then the multiple skill nodes contained within that selected skill cluster are retrieved and sorted. Specifically: The system aggregates multiple skills into skill clusters based on skill intent similarity, business object consistency, tool set similarity, historical combination frequency, density of pre- and post-requirements, and recall performance in conjunction with user questions. A skill cluster is represented as follows: , The skill cluster scoring method obtained based on the constraints is as follows:

[0050] Wherein, IntentSim represents skill intent similarity; ObjectConsistency represents business object consistency; ToolJaccard represents tool set similarity; CoExecution represents the frequency of two skills appearing together in historical success paths, i.e., historical combination frequency; DependencyDensity represents the tightness of the dependency relationship between skills, i.e., the density of pre- and post-relationships; RecallCoOccurrence represents the frequency of two skills being jointly recalled by similar questions, i.e., the situation of user questions being jointly recalled; u1, u2, u3, u4, u5, and u6 are the weight coefficients corresponding to each parameter.

[0051] The system first searches at the skill cluster level, and then expands to specific skills. By aggregating and expanding skill clusters in layers, the number of search branches is further reduced, the search space at the skill capability level is reduced, and the search efficiency in large-scale skill scenarios is improved.

[0052] In one specific implementation, step 3 involves constructing a recall anchor set from the recalled candidate nodes. This recall anchor set is used to limit the scope of the graph search. Based on the recall anchor set, a necessary completion node set is determined, which in turn determines the actual set of nodes allowed for searching, and these are searched preferentially. Access to the necessary completion node set outside the recall anchor set is permitted, provided that skill integrity, dependency, and parameter passing requirements are met. In this scheme: The system will recall the problematic node A. question Intent Node A intent Skill Node A skill Skill Step Node A step and tool node A tool This constitutes the recall anchor set: , To ensure the completeness of the skill path, the system determines the following set of necessary completion nodes: , Among them, by using the dependency completion node set C dependency Skill step completion node set C skillstep Tool completion node set C tool and parameter completion node set C parameter This constitutes the set of nodes C that needs to be completed. required When it is found that the current set of recall anchors is insufficient to form a complete executable path, from C... requiredSelect necessary nodes to fill in. For example, in the graph, if the node information is incomplete due to the prerequisite relationship of node skills, select necessary nodes to fill in.

[0053] The actual set of nodes allowed for searching is:

[0054] This solution limits the search scope by recalling the anchor point set, while ensuring the completeness of the skill path by supplementing necessary nodes. It achieves a balance between search scope and path completeness by ensuring search efficiency and avoiding missing key nodes.

[0055] As a specific implementation, in step 3, when there is a stable and verified intermediate sub-path between two non-adjacent level nodes, the system establishes a cross-level jump edge and performs a jump edge search, thereby improving the path search capability.

[0056] Jump edges are represented as:

[0057] Where: u represents the jump start node; v represents the jump target node; πuv represents the compressed intermediate sub-path; Σuv represents the sub-path constraint summary, recording necessary inputs, outputs, prerequisites, permissions, and necessary steps; Reliability represents historical reliability; Cost represents the expected execution cost; FailureFeature represents historical failure characteristics; Version represents the applicable tool version; TTL represents the cache expiration time. Jump edges automatically become invalid when the tool version changes, the failure rate exceeds a threshold, or the constraint summary expires.

[0058] The constraints of a jump edge include: required prerequisite skills, required input parameters, output parameters that can be generated, required permissions, necessary steps, business prohibition conditions, historical failure types, and the current tool version.

[0059] Only intermediate subpaths that meet the following conditions can generate jump edges: , Where SRπ represents the historical success rate of the subpath, STπ represents the structural stability of the subpath, Nπ represents the number of historical validations, and ConstraintConflict... πThis indicates whether there are hard constraint conflicts. Only sub-paths that have been verified a sufficient number of times and do not have hard constraint conflicts can be compressed into jump edges to prevent accidentally successful paths from becoming fixed. After candidate paths are generated, the system also needs to expand the jump edges to perform a complete verification. This scheme compresses verified stable sub-paths through cross-layer jump edges, reducing the search hierarchy and the number of node expansions, improving path search efficiency, and ensuring the reliability of jump paths.

[0060] As a specific implementation method, in step 3, when the system expands the candidate nodes using heuristic search, for the search state n, a search function is used, expressed as: f(n) = g(n) + h(n), Where g(n) represents the cumulative cost from the initial state to the current state, and h(n) represents the expected remaining cost from the current state to completing the task objective. h(n) is calculated based on the similarity between the current node and the user's task objective, the number of necessary constraints that have not yet been met, the expected failure risk, the number of uncovered task objectives, and the expected remaining search depth.

[0061] Furthermore, the heuristic function h(n) is expressed as: , Among them, Sim goal (n,q) represents the similarity between the current node and the user's task goal, UnsatisfiedConstraint(n) represents the number of unsatisfied constraints, ExpectedFailure(n) represents the expected failure risk of the current path, RemainingGoal(n) represents the number of uncovered task goals, ExpectedDepth(n) represents the expected remaining search depth, and a1, a2, a3, a4, and a5 represent the weight coefficients corresponding to the above parameters, respectively. The system prioritizes expanding search states with smaller f(n). It simultaneously considers target distance, unsatisfied constraints, failure risk, remaining goals, expected depth, parameter gaps, and version conflicts. Compared with searches using only semantic similarity, it can prioritize paths with more complete parameters, better version compatibility, and lower execution risk. This scheme improves search quality and path reliability by using a heuristic function that comprehensively considers multiple dimensions such as target distance, unsatisfied constraints, failure risk, remaining goals, and expected depth, prioritizing paths with more complete parameters, better version compatibility, and lower execution risk.

[0062] As a preferred implementation, in step 3, when performing constraint verification on candidate skill paths, the feasibility of the candidate paths is determined based on a set of hard constraints. Hard constraints include: skill prerequisites, step order, required tool parameters, calling permissions, input / output types of tools before and after, completeness of necessary steps, prohibited nodes, and current business status, expressed as:

[0063] Wherein: Hdependency indicates whether the prerequisite dependencies of the skill are met; Horder indicates whether the order of steps is correct; Hparameter indicates whether the required parameters of the tool can be obtained; Hpermission indicates whether the calling permissions are met; Htype indicates whether the input and output types of the tool before and after are compatible; Hmandatory indicates whether the necessary steps are complete; Hforbidden indicates whether it contains nodes that are prohibited from being called; Hstate indicates whether the current business state allows execution.

[0064] For the k-th hard constraint: , The feasibility of a candidate path is expressed as: , When Feasible=0, it means that the candidate path violates at least one non-relaxable hard constraint and should be deleted or enter the path repair process.

[0065] Based on the set of repairable constraints, the repair cost is calculated for candidate skill paths entering the path for repair. Deletion or repair is selected based on the cost score. Repairable constraints include: missing parameters, precondition constraints, and format constraints. The set of repairable constraints is represented as:

[0066] Where: RmissingParam represents missing parameter constraints, Rprerequisite represents precondition constraints, and Rformat represents format constraints.

[0067] The cost of path repair is expressed as: , Where, r j (P) represents the degree of violation of the j-th type of repairable constraint; ρ j This indicates the corresponding repair cost. For example, missing a parameter that can be queried by the user has a low repair cost; requiring an additional data format conversion step has a medium repair cost; and requiring new external system permissions has a high repair cost. The specific cost can be adjusted according to the situation.

[0068] The above solution uses layered verification of hard constraints and repairable constraints to eliminate or repair infeasible paths in advance, avoiding sending paths that violate key constraints to the execution stage, thereby improving path feasibility and execution success rate.

[0069] Based on the above scheme, the system scores candidate skill paths based on basic utility and path penalty.

[0070] The aforementioned basic utilities include: recall node coverage, intent matching, skill matching, step completeness, tool and parameter suitability, constraint satisfaction, historical reliability, and output coverage. The scoring formula is expressed as follows:

[0071] Wherein: Coverage (P) represents the degree of coverage of high-scoring recall nodes; IntentMatch (P) represents the degree of matching between the path and the user's intent; SkillMatch (P) represents the degree of coverage of the skills in the path with the task objective; StepCompleteness (P) represents the degree of completeness of necessary steps; ToolFitness (P) represents the degree of fit between tool capabilities and parameters; ConstraintSatisfaction (P) represents the degree of satisfaction of repairable constraints and soft constraints; HistoryReliability (P) represents the reliability of the path's historical execution; OutputCoverage (P) represents the degree of coverage of the path with the user's expected output; w1, w2, w3, w4, w5, w6, w7, and w8 are the weight coefficients of the corresponding parameters.

[0072] Path penalties include: RepairCost(P), LengthPenalty(P), IrrelevancePenalty(P), ExecutionRisk(P), ExpectedCost(P), ExpectedLatency(P), and FailurePenalty(P). The scoring formula is as follows: , Where λ1, λ2, λ3, λ4, λ5, λ6, and λ7 are the weighting coefficients of the corresponding parameters.

[0073] The final path score, derived from Feasible (P), BaseUtility (P), and Penalty (P), is expressed as follows:

[0074] The above scheme comprehensively evaluates the quality and cost of candidate paths through a multi-dimensional assessment of basic utility and path penalty. It avoids prioritizing candidate paths with high semantic matching but high execution risk or lengthy paths. In other words, paths that violate hard constraints will not continue to enter the normal candidate set due to their high semantic matching, thereby improving the rationality of path selection and execution efficiency.

[0075] The above scoring scheme also includes: 1. Skill step integrity verification, Skill i The necessary steps are set as follows The actual set of steps contained in the candidate path is Then the completeness of the steps is expressed as: , When the completeness of a step is lower than a preset threshold, the system needs to supplement the missing steps, mark the path as a path to be repaired, or delete the candidate path. 2. Tool parameter completeness verification: For each tool in the candidate path, the system determines whether its required inputs can be provided by information such as user questions, user context, outputs of preceding tools, default graph parameters, external data sources, and user interaction supplements. The completeness of tool parameters is expressed as: , Where: Nrequired represents the total number of required parameters for the tool; NavailableRequired represents the number of required parameters currently available. For required parameters that cannot be obtained, the system determines whether they belong to: repairable parameters that can be supplemented by the user, parameters that can be generated by other skills, or essential parameters that cannot be repaired. When they are repairable parameters that can be supplemented by the user or parameters that can be generated by other skills, the path is marked as a path to be repaired; when they are essential parameters that cannot be repaired, the candidate path is deleted.

[0076] 3. Historical Reliability: Existing search results are scored based on historical reliability. The historical reliability of a path is expressed as follows: , Among them: SR p Indicates the historical execution success rate of the path; UR p Indicates user confirmation rate; Reuse p Indicates path reuse stability; CR p Indicates the user correction rate; ER p This represents the path anomaly occurrence rate; c1, c2, c3, c4, and c5 are the weights of the corresponding parameters. When the historical reliability is lower than a preset threshold, the path is marked as a path to be repaired or the candidate path is deleted.

[0077] As a preferred implementation, in step 3, when determining the target skill path: first, select the path with the higher score within the allowed score difference range to construct a high-quality candidate path set; then, dynamically determine the Temperature parameter based on task risk, candidate path differences, and path historical stability; finally, determine the selection strategy based on the Temperature parameter, i.e., when Temperature is 0, select the path with the highest score; when Temperature is greater than 0, select the target path from the high-quality candidate path set according to a probability distribution. Specifically: The system constructs a high-quality candidate path set Phigh based on the comprehensive path score PathScore. Let the highest path score be Scoremax, then: , in, The allowed score difference. The system will also select the top-N paths from all candidate paths to form a high-quality candidate set.

[0078] When Temperature equals 0, the system selects the path with the highest score: ; When Temperature is greater than 0, the system selects the target path from the set of high-quality candidate paths according to probability: , Where: T represents the Temperature parameter; λ represents the score scaling factor. When T is small, the system tends to choose the highest-scoring path; when T is large, the system explores multiple high-quality paths to some extent.

[0079] In other words, there may be multiple similar paths to the target skill, and their scores need to be ranked. However, the highest-scoring path is not always selected. Instead, path selection is made from a high-quality candidate path set using a Temperature-controlled approach. By dynamically adjusting the Temperature parameter, a flexible balance is achieved between deterministic selection and high-scoring path exploration. This ensures the stability of high-risk tasks while retaining the ability to explore multi-path scenarios, thus improving the adaptability and robustness of path selection.

[0080] In summary, the overall solution in step 3 first identifies candidate skill combinations in the high-level graph through a hierarchical search strategy, and then develops skill steps, tools, and parameters for high-quality candidate combinations, thereby reducing indiscriminate searching in the complete graph and lowering the complexity of the search space. By using a heuristic function that comprehensively considers multiple dimensions such as target distance, unmet constraints, failure risk, remaining targets, and expected depth, paths with more complete parameters, better version compatibility, and lower execution risk are prioritized. Through hierarchical verification of hard constraints and repairable constraints and multi-dimensional comprehensive scoring, infeasible paths are eliminated in advance or the cost of repair is quantified, avoiding sending paths that violate key constraints or have excessively high risks to the execution stage, thus improving the efficiency, feasibility, and success rate of path generation.

[0081] Regarding step 4, if the target skill path fails, a local path replanning is performed. Specifically: In step 4, during local path replanning: first, failed paths are marked as failures, generating a hard prohibition set and a soft penalty set. Failure penalties are calculated for nodes and edges in the soft penalty set, which affect the selection of subsequent candidate paths. Then, failed paths are split, retaining the prefixes of successfully executed paths, and performing parameter repair or equivalent tool replacement based on the failure type. Finally, candidate repair paths are generated based on the objective function of the repair path. The objective function is calculated based on task objective coverage, output reusability, reliability, constraint satisfaction, failure risk, structural differences between the repair path and the original remaining paths, additional call costs, and the cost of rolling back executed steps.

[0082] As a specific solution, once the target skill path is determined, the system executes the skills, skill steps, and tools in the order specified in the path. The system records: the actual skill executed, the actual execution steps, the input parameters for each step, the output results for each step, the tool call status, the call time, the call cost, exception information, and user interruption or modification information.

[0083] When a failure event occurs during the execution of the target skill path, the system categorizes the failure event by cause. The failure types include: Parameter failure: missing required parameter; incorrect parameter type; incorrect parameter format; parameter value out of range; conflicting parameters.

[0084] Permission failure: User lacks permission; system token expired; tool authorization scope insufficient; current business object inaccessible.

[0085] Tool status class failure: Tool unavailable; Interface timed out; Service rate limiting; Tool version expired; External system exception.

[0086] Data failure: previous output is empty; data structure does not meet expectations; input and output data types are incompatible; data has expired; data integrity check failed.

[0087] Process constraint failures: missing prerequisite skills; incorrect skill execution order; necessary steps not executed; approval status not met; current business status does not allow execution.

[0088] Unrecoverable failure: Irreversible operation has been partially completed; target resource has been deleted; business status has encountered a conflict that cannot be automatically resolved; the system is not allowed to automatically replace tools or steps.

[0089] The system generates a failure flag based on the failure event:

[0090] Here, ParameterPattern represents the parameter pattern that caused the failure; ContextCluster represents the context category in which the failure occurred; ConstraintState represents the constraint state when the failure occurred; and TTL represents the validity period of the failure flag.

[0091] For permanent failures, explicit permission restrictions, or business rule restrictions, add the relevant nodes or edges to the hard-prohibited set B. hard For nodes or edges that time out, are rate-limited, or are temporarily unavailable, add them to the soft penalty set B. soft .

[0092] In the above steps, when performing path splitting, the original path is: , Where: P done Indicates a path prefix that has been successfully executed; P failed Indicates the node or subpath where the failure occurred; P remain This indicates the remaining paths that have not yet been executed. The system reserves P. done The execution results are used to replan the subsequent path starting from the most recently successfully executed checkpoint.

[0093] The above scheme avoids the high cost of re-executing the entire path by preserving the path prefixes of successful executions and only performing local repairs on the failed parts. It prevents the repeated selection of known failed nodes through failure marking and penalty mechanisms, and minimizes additional costs and structural changes while resuming execution through multi-objective optimized repair path selection, thus significantly reducing the repair cost after execution failure.

[0094] As a specific solution, if the failure is due to missing parameters, incorrect parameter format, or incorrect parameter type, the system performs one or more of the following operations: re-extracts parameters from the user's question; obtains parameters from historical sessions or context; maps parameters from the output of preceding tools; invokes the parameter conversion step; or requests additional parameters from the user. The original tool is then re-executed after the issue is resolved.

[0095] As a specific solution, if the failure is due to a temporary timeout, rate limiting, or a brief service anomaly, and the number of retries has not reached the threshold, the system will use a preset backoff strategy to re-invoke the original tool. If the same failure type occurs consecutively beyond the threshold, the system will stop retrying the original tool and begin searching for alternative tools.

[0096] As a preferred implementation, if the original tool becomes unusable, the system searches for an alternative tool with the same or similar capabilities. This tool replacement is based on a Substitute Score, which is calculated based on capability similarity, input parameter compatibility, output result compatibility, historical reliability of the alternative tool, permission compatibility, current context failure risk, and tool switching cost.

[0097] The alternative tool adaptation is represented as follows: , Wherein: CapabilitySim represents capability similarity; InputCompatibility represents the compatibility of input parameters; OutputCompatibility represents the compatibility of output results; Reliability represents the historical reliability of the replacement tool, which is accumulated based on the scores of successful historical replacements; PermissionMatch represents the degree of permission adaptation, indicating the user's permissions for the tool; the values ​​of CapabilitySim, InputCompatibility, and OutputCompatibility are calculated based on the similarity of semantic vectors and field rules.

[0098] The above solution uses multi-dimensional adaptation and evaluation of alternative tools to quickly find functionally similar and compatible alternatives when a tool fails, ensuring the success rate of path repair and execution continuity, and reducing task interruptions caused by the failure of a single tool.

[0099] For nodes and edges in the soft failure set, the system calculates a failure penalty. When generating new candidate repair paths, the system comprehensively evaluates the failure penalty value as a path risk factor. If a candidate path contains the same or similar nodes, tools, or parameter conditions as historical failures, the failure risk score of that path is increased, thus reducing its priority in path ranking; for paths with historically stable execution and low failure risk, a higher selection priority is maintained.

[0100] When the repair path requires changing the data generated by already completed steps, the system determines whether it is necessary to: re-execute some of the preceding steps; execute the compensation steps; roll back the original execution results; or request user confirmation.

[0101] Based on the scheme in step 4 above, the system's failure repair process can be summarized as follows: Obtain the failed tools, steps, and error information; classify the causes of failure; determine whether the problem can be resolved through parameter repair or retrying the original tool; identify the most recently successfully executed checkpoint; generate a hard prohibition set and a soft penalty set; obtain the valid outputs generated by completed steps; determine the remaining task objectives; search for alternative tools, steps, or skills; generate multiple candidate repair paths; calculate the candidate repair path scores; perform hard constraint verification on the target repair path; connect the repair path with the prefix of the already executed path; continue executing the repair path; and record the final repair result.

[0102] The technical effects of step 4 in this application are as follows: by retaining the path prefixes that have been successfully executed and only performing local repairs on the failed parts, the high cost of replanning or re-executing the entire path is avoided; the failure marking and penalty mechanism prevents the repeated selection of known failed paths; execution is quickly restored through parameter repair or equivalent tool replacement, and the repair path selection takes into account target coverage, output reuse, reliability, constraint satisfaction, failure risk, structural differences, additional costs and rollback costs, achieving the optimal balance between repair costs and repair effects, and significantly reducing the repair costs after execution failure.

[0103] For step 5, the skill association graph is dynamically updated based on the execution results.

[0104] In this step, the system does not directly write all user issues into the graph. Instead, user issues are created as issue nodes only when one or more of the following conditions are met: the user explicitly confirms that the recommended path is reasonable, the user completes the task using the path, the path is actually executed successfully, the user confirms the final path after modifying the path, or the business personnel manually verify the information. In each issue node, the following information is recorded: original issue, issue summary, issue vector, associated intent, original recommended path, user's final confirmed path, actual execution result, user's modifications, path score, and graph version.

[0105] When the system recommends the Porigin path, but the user ultimately confirms the Puser path, the feedback difference is: , Feedback on discrepancies includes: the number of newly added skills, the number of deleted skills, the number of replaced skills, the number of times the skill order was adjusted, the number of added steps, the number of deleted steps, the number of times tools were replaced, and the number of times parameters were corrected. Based on these path discrepancies, the deviation between the current recommended path and the user's actual needs is determined, and this discrepancy information is used as data for subsequent optimization of the skill matching model and path scoring strategy.

[0106] The system statistically analyzes path execution characteristics based on problem type, intent, business object, or similar task clusters to obtain path statistical indicators, including: whether the path was successfully executed, whether the user confirmed it, the completion status of the task objective, execution anomalies, and path correction status. Based on the path effectiveness evaluation results, the system updates skill nodes, skill relationship edges, and path scoring parameters, giving higher priority to historically validated skill combinations and lowering the selection priority of skill combinations that fail to execute or are frequently modified.

[0107] After the path execution is complete, the system records: the failed tool, failure type, failure parameters, failure context, original path, alternative tool, alternative steps, repair path, repair success rate, repair cost, and repair time. Then, the system updates the model matching the problem skill based on newly generated problem nodes, user-confirmed paths, execution results, and failure repair results. The model uses user-confirmed skills and paths as positive samples, and skills that will be deleted or replaced by the user, skills that violate hard constraints, semantically similar skills but with incorrect business objects, skills that fail to execute in the same context, skills that cannot produce the target output, and paths explicitly rejected by the user as difficult negative samples. During the model update process: Overall benefits of system-defined path execution: , Where: Success indicates whether the path was executed successfully; UserConfirm indicates whether the user confirmed the path; GoalCompletion indicates the degree of completion of the task objective; Exception indicates the exception and its severity; Correction indicates the intensity of user correction; Cost indicates the execution cost; Time indicates the execution time; RepairCount indicates the number of repairs during the path execution process; μ1, μ2, μ3, μ4, μ5, μ6, μ7, and μ8 are the weight coefficients corresponding to each parameter.

[0108] The system adjusts the following parameters based on the average reward of paths in different task clusters: node recall weight, problem skill matching model, path rating weight, hard constraint rules, repairable constraint cost, path length penalty, failure risk weight, and Temperature parameter.

[0109] The system update strategy is as follows: When a user explicitly corrects the path, a serious tool failure occurs, or a local path repair is successfully completed, update the local nodes, edges, and failure markers.

[0110] The system aggregates multiple user issues, execution paths, and failed repair samples according to a preset cycle, and then updates the vector model, matching model, and scoring parameters in a unified manner.

[0111] An update is triggered when the cumulative number of path executions reaches a threshold, the path success rate decreases, the user correction rate increases, the tool anomaly rate exceeds a threshold, the skill file changes, the tool version changes, or the recovery path is successfully verified multiple times. The system prioritizes updating local policies within specific problem clusters, intent clusters, or business object clusters, and then promotes the validated policies to higher levels for updates.

[0112] In summary, in step 5, when the user confirms the recommended path or the path is executed successfully, the system creates the user's question as a question node and adds it to the skill association graph. Based on the differences between the user-confirmed path and the recommended path, as well as path execution statistics, the system updates the question skill matching model and path strategy parameters. Path execution statistics include path success rate, user confirmation rate, user correction rate, and path anomaly rate. Path strategy parameters include node recall weight, question skill matching model, path scoring weight, hard constraint rules, repairable constraint cost, path length penalty, failure risk weight, and the Temperature parameter.

[0113] The aforementioned path strategy parameters refer to a series of adjustable configuration items involved in the system's adaptive adjustment based on path execution rewards, specifically including but not limited to: Node recall weight, which is the scoring weight of each type of node in the recall phase; Problem-skill matching model, which is the matching algorithm and model parameters between problems and skills; Path score weights are the weight coefficients of each item in the basic utility and path penalty. Hard constraint rules are the hard conditions and thresholds for determining the feasibility of a path. Repairable constraint cost, i.e., the repair cost coefficient corresponding to various repairable constraints; Path length penalty, which refers to the relevant parameters in the path length penalty formula; Failure risk weight, which is the coefficient used in the calculation of failure penalty; Temperature parameter, which is the temperature control parameter T when selecting a path.

[0114] In other words, path strategy parameters are the collective term for all scoring weights, constraint thresholds, penalty coefficients, and selection strategy variables that the system relies on when evaluating, selecting, and optimizing skill execution paths. The system dynamically adjusts these parameters by continuously calculating the average returns of paths across different task clusters to improve the accuracy and success rate of subsequent path recommendations.

[0115] Step 5 above involves backfilling verified user issues and execution experience into the graph, and dynamically adjusting various strategy parameters based on user feedback differences and execution statistics. This enables the system to continuously learn and optimize itself, thereby continuously improving the accuracy of subsequent path recommendations and the success rate of execution as the number of uses increases. This achieves intelligent and precise processing across the entire chain from user issues to skill invocation.

[0116] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the present invention in any way, and all technical solutions obtained by equivalent substitution or equivalent transformation fall within the protection scope of the present invention.

Claims

1. A skill invocation method based on a skill graph, characterized in that, Includes the following steps: Construct a skill association graph, defining skill nodes, skill step nodes, tool nodes, tool input parameter nodes, tool output parameter nodes, intent nodes, problem nodes, and the edge relationships between each node; Receive and analyze user questions, and jointly recall candidate nodes based on multiple node types; Based on the candidate nodes, a heuristic search is performed on the hierarchical graph to generate candidate skill paths, and the candidate skill paths are constrained, verified, and scored to determine the target skill path. The target skill path is then selected to execute the skill call. When the target skill path fails to execute, a local path replanning is performed; The skill association graph is dynamically updated based on the execution results.

2. The skill invocation method based on skill graphs according to claim 1, characterized in that, The construction of the skill association graph includes: extracting skill information, skill step information, tool information and parameter information from skill files and importing them into corresponding nodes; establishing edge relationships between skill nodes and intent nodes, skill nodes and tool nodes, and tool nodes and parameter nodes; and vectorizing the node descriptions.

3. The skill invocation method based on skill graphs according to claim 2, characterized in that, The construction of the skill association graph also includes: performing intent recognition and classification on skill descriptions, and merging new intents with existing intents or creating new intent nodes based on intent merging scores; The intent merging score is calculated based on vector similarity, text similarity, keyword similarity, and rule consistency.

4. The skill invocation method based on skill graphs according to claim 1, characterized in that, The process of receiving and parsing user questions includes: Analyze user questions to obtain a question summary, user intent, target action, business object, known parameters, missing parameters, time and scope conditions, permissions and compliance conditions, expected output, and prohibited operations; Candidate nodes are jointly recalled from multiple node types based on recall scores; The recall score is calculated based on vector similarity, keyword matching degree, text semantic similarity, business object matching degree, task action matching degree, and business rule matching degree.

5. The skill invocation method based on skill graphs according to claim 4, characterized in that, When the combined matching score between the historical problem and the current problem reaches a high confidence threshold, the historical problem is associated with a verified skill path, the corresponding skills and tools are in a valid state, the parameters of the current problem are compatible with the historical problem, and there are no hard failure markers in the historical path, the corresponding historical skill path will be taken as the preferred candidate path.

6. The skill invocation method based on skill graphs according to claim 1, characterized in that, The heuristic search based on hierarchical graphs includes: The skill association graph is divided into the problem and intent layer, the skill ability layer, the execution step layer, and the tool and parameter layer in sequence. After determining the candidate skill combination based on the candidate node in the question and intent layer and the skill capability layer, the execution step layer and the tool and parameter layer are then entered to search for the skill steps, tools and parameters of the candidate skill combination in order to generate the candidate skill path.

7. The skill invocation method based on skill graphs according to claim 6, characterized in that, Based on the constraints, multiple skill nodes in the skill capability layer are aggregated into skill clusters to form a skill cluster layer; First, the corresponding skill cluster is searched in the skill cluster layer, and then the multiple skill nodes contained in the selected skill cluster are retrieved and sorted.

8. The skill invocation method based on skill graphs according to claim 6, characterized in that, The recalled candidate nodes constitute a recall anchor point set; The necessary set of completion nodes is determined based on the recall anchor point set, and then the actual set of nodes allowed to be searched is determined. While satisfying the requirements of skill integrity, dependency, and parameter passing, access to the set of necessary completion nodes outside the set of recall anchors is permitted.

9. The skill invocation method based on skill graphs according to claim 6, characterized in that, When there is a stable and verified intermediate sub-path between two non-adjacent level nodes, the system establishes a cross-level jump edge and performs a jump edge search.

10. The skill invocation method based on skill graphs according to claim 6, characterized in that, The heuristic search employs a search function: f(n) = g(n) + h(n), Where g(n) represents the cumulative cost from the initial state to the current state, and h(n) represents the expected remaining cost from the current state to the completion of the task objective. The h(n) is calculated based on the similarity between the current node and the user's task objective, the number of necessary constraints that have not yet been met, the expected failure risk, the number of uncovered task objectives, and the expected remaining search depth.

11. The skill invocation method based on skill graphs according to claim 1, characterized in that, The constraint verification of the candidate skill paths includes: The feasibility of candidate paths is determined based on a set of hard constraints, which include: skill prerequisites, step order, required tool parameters, calling permissions, input and output types of tools before and after, completeness of necessary steps, nodes that are prohibited from being called, and current business status. For any candidate skill path that violates at least one non-relaxable hard constraint, delete it or initiate a path repair process; The repair cost is calculated for the candidate skill path for path repair based on the set of repairable constraints, which includes: missing parameters, precondition constraints, and format constraints.

12. The skill invocation method based on skill graphs according to claim 11, characterized in that, The candidate skill paths are scored and calculated based on basic utility and path penalty. The basic utilities include: recall node coverage, intent matching, skill matching, step completeness, tool and parameter fit, constraint satisfaction, historical reliability, and output coverage. The path penalties include: repair cost, path length penalty, irrelevance penalty, execution risk, expected cost, expected delay, and failure penalty.

13. The skill invocation method based on skill graphs according to claim 1, characterized in that, Determining the target skill path includes: Within the allowed score difference range, select the path with the higher score to construct a high-quality candidate path set; The Temperature parameter is dynamically determined based on task risk, candidate path differences, and path historical stability. The selection strategy is determined based on the Temperature parameter: when Temperature is 0, the path with the highest score is selected; when Temperature is greater than 0, the target path is selected from the set of high-quality candidate paths according to the probability distribution.

14. The skill invocation method based on skill graphs according to claim 1, characterized in that, The local path replanning includes: marking failed paths as failures, generating a hard prohibition set and a soft penalty set, and calculating failure penalties for nodes and edges in the soft penalty set; The failed paths are split up, retaining the prefixes of successfully executed paths, and parameter repair or equivalent tool replacement is performed according to the failure type; Candidate repair paths are generated based on the objective function of the repair path. The objective function is calculated based on the task objective coverage, output reusability, reliability, constraint satisfaction, failure risk, structural differences between the repair path and the original remaining path, cost of adding calls, and cost of rolling back executed steps.

15. The skill invocation method based on skill graphs according to claim 14, characterized in that, The equivalent tool replacement is based on the alternative tool adaptation score, which is calculated based on capability similarity, input parameter compatibility, output result compatibility, historical reliability of the alternative tool, permission adaptation, current context failure risk, and tool switching cost.

16. The skill invocation method based on skill graphs according to claim 1, characterized in that, The dynamic update includes: when the user confirms the recommended path or the path is executed successfully, the user's question is created as a question node and stored in the skill association graph; based on the difference between the user's confirmed path and the recommended path and the path execution statistics, the question skill matching model and path strategy parameters are updated.

Images