Interactive processing methods, vehicles, and computer-readable storage media

By searching and testing the agent's skill library, target skills without logical conflicts were selected, which solved the problem of poor agent interaction processing and improved response speed and stability.

CN123115489APending Publication Date: 2026-11-17安徽开阳科技有限公司 +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611298678.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-25
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In related technologies, the scheduling of intelligent agent skills generally adopts retrieval and post-processing logic, which leads to redundant and conflicting skills occupying core computing power, resulting in resource waste and poor interaction processing effect of the target client.

Method used

The skill library is searched based on the interactive request text to filter candidate skills. Logic conflict tests are conducted to eliminate conflicting skills. The interactive request text is then processed using an execution model to generate the processing results.

Benefits of technology

It enables the rapid identification of potentially available skills related to user intent, reduces computing costs, improves the response speed and execution stability of intelligent agent interaction, and prevents invalid skills from entering the high-computing-power inference stage.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115489A_ABST
    Figure CN123115489A_ABST
Patent Text Reader

Abstract

This application provides an interaction processing method, a vehicle, and a computer-readable storage medium. The method includes: responding to receiving an interaction request text from a target client; retrieving multiple preset skills stored in a skill base based on the interaction request text to obtain multiple candidate skills; performing conflict testing on the multiple candidate skills based on the functional implementation logic of the multiple candidate skills to obtain conflict test results; filtering the multiple candidate skills based on the conflict test results to obtain a target skill; sending the target skill and the interaction request text to an execution model, and using the execution model to process the interaction request text according to the target skill to obtain a processing result for the interaction request text. This application solves the technical problem of poor interaction processing effect of the target client in related technologies.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to an interactive processing method, a vehicle, and a computer-readable storage medium. Background Technology

[0002] In the field of artificial intelligence, intelligent agents need to maintain dozens to hundreds of fine-grained discrete skills, covering data querying, execution, and tool invocation, to adapt to diverse vehicle control and interaction needs. However, the skill scheduling of intelligent agents in related technologies generally adopts retrieval and recall logic and post-processing logic, resulting in a large number of redundant and conflicting skills occupying core computing power and causing resource waste. Therefore, the interaction processing effect of the target client in related technologies is poor.

[0003] There is currently no good solution to the above problems. Summary of the Invention

[0004] This application provides an interactive processing method, a vehicle, and a computer-readable storage medium to at least solve the technical problem of poor interactive processing performance of the target client in related technologies.

[0005] According to one aspect of the embodiments of this application, an interaction processing method is provided, comprising: responding to receiving an interaction request text from a target client; retrieving multiple preset skills stored in a skill library based on the interaction request text to obtain multiple candidate skills; performing conflict testing on the multiple candidate skills based on the functional implementation logic of the multiple candidate skills to obtain conflict test results, wherein the conflict test results are used to characterize whether there is a conflict in the functional implementation logic of the multiple candidate skills; filtering the multiple candidate skills based on the conflict test results to obtain a target skill; sending the target skill and the interaction request text to an execution model, and using the execution model to process the interaction request text according to the target skill to obtain a processing result of the interaction request text.

[0006] Furthermore, based on the interaction request text, multiple preset skills stored in the skill library are retrieved to obtain multiple candidate skills, including: transforming the interaction request text based on a preset vector dimension to obtain a target query vector, wherein the preset vector dimension is used to represent the vector dimension of the function vectors corresponding to multiple preset skills; and retrieving multiple preset skills based on the interaction request text and the target query vector to obtain multiple candidate skills.

[0007] Furthermore, multiple preset skills are retrieved based on the interaction request text and the target query vector to obtain multiple candidate skills, including: matching the interaction request text, the target query vector, and multiple preset skills to obtain text matching results and vector matching results of multiple preset skills; and filtering multiple preset skills based on the text matching results and vector matching results of multiple preset skills to obtain multiple candidate skills.

[0008] Furthermore, the interaction request text, target query vector, and multiple preset skills are matched to obtain text matching results and vector matching results for multiple preset skills. This includes: performing text matching on the interaction request text and the text description information of multiple preset skills to obtain text matching results for multiple preset skills; and performing vector matching on the target query vector and the function vectors of multiple preset skills to obtain vector matching results for multiple preset skills.

[0009] Furthermore, based on the text matching results and vector matching results of multiple preset skills, the multiple preset skills are filtered to obtain multiple candidate skills, including: determining the skill evaluation scores of multiple preset skills based on the text matching results and vector matching results of multiple preset skills; ranking the multiple candidate skills based on the skill evaluation scores of multiple preset skills to obtain ranking results; and filtering the multiple preset skills based on the ranking results to obtain multiple candidate skills.

[0010] Furthermore, based on the text matching results and vector matching results of multiple preset skills, the skill evaluation scores of multiple preset skills are determined, including: determining the text matching scores of multiple preset skills based on the text matching results of multiple preset skills; determining the vector matching scores of multiple preset skills based on the vector matching results of multiple preset skills; and weighting and summing the text matching scores and vector matching scores based on a preset weighting ratio to obtain the skill evaluation scores of multiple preset skills.

[0011] Furthermore, multiple preset skills are filtered based on the ranking results to obtain multiple candidate skills, including: filtering multiple preset skills based on the ranking results to obtain multiple initial skills; performing similarity matching between any two initial skills to obtain similarity matching results, wherein the similarity matching results are used to indicate whether the similarity between any two initial skills is greater than a preset threshold; and removing duplicates from the multiple initial skills based on the similarity matching results to obtain multiple candidate skills.

[0012] Furthermore, multiple preset skills are filtered based on the ranking results to obtain multiple initial skills, including: determining the preset skills that are ranked before a preset position in the ranking results as multiple initial skills; or, extracting a preset number of preset skills based on the ranking results as multiple initial skills.

[0013] Furthermore, based on the similarity matching results, multiple initial skills are deduplicated to obtain multiple candidate skills, including: dividing the multiple initial skills based on the similarity matching results to obtain at least one first initial skill and multiple second initial skills, wherein the multiple second initial skills are associated initial skills with a similarity greater than a preset threshold, and at least one first initial skill is other skills among the multiple initial skills besides the multiple second initial skills; deduplicating the multiple second initial skills to obtain the deduplicated target skill; merging the deduplicated target skill and at least one first initial skill to obtain multiple candidate skills.

[0014] Furthermore, the multiple second initial skills are deduplicated to obtain the deduplicated target skills, including: dividing the multiple second initial skills into multiple skill combinations; for any one of the multiple skill combinations, deduplicating the initial skills in that skill combination based on the skill evaluation score of the initial skills in that skill combination to obtain the deduplicated skills corresponding to that skill combination; and determining the deduplicated target skills based on the deduplicated skills corresponding to the multiple skill combinations.

[0015] Furthermore, conflict testing is performed on multiple candidate skills based on their functional implementation logic to obtain conflict test results. This includes: comparing the compatibility of multiple candidate skills based on their functional implementation logic to obtain comparison results, where the comparison results indicate whether any of the following exists: mutually exclusive operation types, data read / write conflicts, or permission overriding. If the comparison results indicate that multiple candidate skills exhibit any of these three issues, the conflict test results determine that the functional implementation logic of the multiple candidate skills is conflicting. If the comparison results indicate that multiple candidate skills do not exhibit any of these three issues, the conflict test results determine that the functional implementation logic of the multiple candidate skills is not conflicting.

[0016] Furthermore, based on the functional implementation logic of multiple candidate skills, a compatibility comparison is performed on multiple candidate skills to obtain comparison results, including: parsing the functional implementation logic of multiple candidate skills to obtain the operation type, data access mode, and skill execution permission of multiple candidate skills; performing structured processing on the operation type, data access mode, and skill execution permission to obtain logical constraints; and performing a compatibility comparison on multiple candidate skills based on the logical constraints to obtain comparison results.

[0017] Furthermore, based on the conflict test results, multiple candidate skills are screened to obtain the target skill, including: in response to the conflict test results indicating that the functional implementation logic of multiple candidate skills does not conflict, multiple candidate skills are determined as the target skill; in response to the conflict test results indicating that the functional implementation logic of multiple candidate skills does conflict, any two candidate skills among the multiple candidate skills whose functional implementation logic conflicts are determined, and multiple candidate skills are screened based on the conflict type of any two candidate skills to obtain the target skill.

[0018] Furthermore, the conflict types include execution permission conflicts, read / write operation conflicts, and operation logic conflicts. Multiple candidate skills are filtered based on the conflict types of any two candidate skills to obtain the target skill. This includes: in response to an execution permission conflict, deleting the first candidate skill from any two candidate skills based on their permission levels to obtain the target skill, wherein the permission level of the first candidate skill is lower than that of the other candidate skill; and in response to read / write operation conflicts and operation logic conflicts, deleting the second candidate skill from any two candidate skills based on their call frequency to obtain the target skill, wherein the call frequency of the second candidate skill is lower than that of the other candidate skill.

[0019] According to another aspect of the embodiments of this application, an interactive processing apparatus is also provided, comprising: a retrieval module, configured to, in response to receiving an interactive request text from a target client, retrieve multiple preset skills stored in a skill library based on the interactive request text to obtain multiple candidate skills; a testing module, configured to perform conflict testing on the multiple candidate skills based on the functional implementation logic of the multiple candidate skills to obtain conflict test results, wherein the conflict test results are used to characterize whether there is a conflict in the functional implementation logic of the multiple candidate skills; a filtering module, configured to filter the multiple candidate skills based on the conflict test results to obtain a target skill; and a processing module, configured to send the target skill and the interactive request text to an execution model, and use the execution model to process the interactive request text according to the target skill to obtain a processing result of the interactive request text.

[0020] Furthermore, the retrieval module includes: a conversion unit, used to convert the interactive request text based on a preset vector dimension to obtain a target query vector, wherein the preset vector dimension is used to represent the vector dimension of the function vectors corresponding to multiple preset skills; and a retrieval unit, used to retrieve multiple preset skills based on the interactive request text and the target query vector to obtain multiple candidate skills.

[0021] Furthermore, the retrieval unit includes: a matching subunit, used to match the interactive request text, the target query vector, and multiple preset skills to obtain text matching results and vector matching results of multiple preset skills; and a filtering subunit, used to filter multiple preset skills based on the text matching results and vector matching results of multiple preset skills to obtain multiple candidate skills.

[0022] Furthermore, the matching subunit includes: a first matching subunit, used to perform text matching on the interactive request text and the text description information of multiple preset skills to obtain the text matching results of multiple preset skills; and a second matching subunit, used to perform vector matching on the target query vector and the function vectors of multiple preset skills to obtain the vector matching results of multiple preset skills.

[0023] Further, the filtering subunit includes: a first determining subunit, which determines the skill evaluation scores of multiple preset skills based on the text matching results and vector matching results of multiple preset skills; a sorting subunit, which sorts multiple candidate skills based on the skill evaluation scores of multiple preset skills to obtain a sorting result; and a second determining subunit, which filters multiple preset skills based on the sorting result to obtain multiple candidate skills.

[0024] Further, the first determining subunit includes: a first score determining subunit, used to determine the text matching score of multiple preset skills based on the text matching results of multiple preset skills; a second score determining subunit, used to determine the vector matching score of multiple preset skills based on the vector matching results of multiple preset skills; and a weighted summation subunit, used to perform weighted summation of the text matching score and the vector matching score based on a preset weighting ratio to obtain the skill evaluation score of multiple preset skills.

[0025] Further, the second determining subunit includes: a third determining subunit, used to filter multiple preset skills based on the ranking result to obtain multiple initial skills; a matching subunit, used to perform similarity matching between any two initial skills among the multiple initial skills to obtain a similarity matching result, wherein the similarity matching result is used to indicate whether the similarity between any two initial skills is greater than a preset threshold; and a deduplication subunit, used to deduplicate multiple initial skills based on the similarity matching result to obtain multiple candidate skills.

[0026] According to another aspect of the embodiments of this application, a vehicle is also provided, including: a memory storing an executable program; and a processor for running the program, wherein the program executes the methods in various embodiments of this application when it runs.

[0027] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored executable program, wherein, when the executable program is running, it controls the device where the computer-readable storage medium is located to perform the methods of various embodiments of this application.

[0028] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program that, when executed by a processor, implements the methods of various embodiments of this application.

[0029] According to another aspect of the embodiments of this application, a computer program product is also provided, including a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods in various embodiments of this application.

[0030] According to another aspect of the embodiments of this application, a computer program is also provided, which, when executed by a processor, implements the methods of the various embodiments of this application.

[0031] In this embodiment, in response to receiving an interaction request text from the target client, multiple preset skills stored in the skill library are retrieved based on the interaction request text to obtain multiple candidate skills. Conflict testing is then performed on the multiple candidate skills based on their functional implementation logic to obtain conflict test results, whereby the conflict test results characterize whether there are conflicts in the functional implementation logic of the multiple candidate skills. Based on the conflict test results, the multiple candidate skills are filtered to obtain the target skill. The target skill and the interaction request text are sent to the execution model, which processes the interaction request text according to the target skill to obtain the processing result of the interaction request text. This embodiment filters multiple candidate skills from the skill library, enabling rapid identification of potentially available skills related to user intent, providing an initial candidate set for subsequent processing. Subsequently, conflict testing is performed on the candidate set to identify potential conflicts such as operational mutual exclusion, data competition, or permission overriding between skills before sending them to the execution model. Then, the multiple candidate skills are filtered based on the conflict detection results, reducing the number of skills entering the subsequent inference stage. Finally, based on the selected target skills, the execution model processes the interaction request text to obtain the processing result. This eliminates the need for the execution model to handle complex conflict arbitration and redundant reasoning, thereby shortening the prompt assembly and reasoning time, improving the response speed and execution stability of the agent interaction, and reducing the overall computing cost. This embodiment of the application achieves the goal of obtaining a set of target skills without logical conflicts and with high adaptability, thus avoiding invalid skills from entering the high-computing-power reasoning stage, and solving the technical problem of poor interaction processing effect of the target client in related technologies. Attached Figure Description

[0032] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0033] Figure 1 This is a flowchart of an interactive processing method according to an embodiment of this application;

[0034] Figure 2 This is a schematic diagram of an optional interactive processing method according to an embodiment of this application;

[0035] Figure 3 This is a schematic diagram of an optional interactive processing method according to an embodiment of this application;

[0036] Figure 4 This is a schematic diagram of an optional interactive processing method according to an embodiment of this application;

[0037] Figure 5 This is a schematic diagram of an optional interactive processing method according to an embodiment of this application;

[0038] Figure 6 This is a schematic diagram of an interactive processing device according to an embodiment of this application. Detailed Implementation

[0039] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0040] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0041] According to an embodiment of this application, an embodiment of an interactive processing method is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0042] This embodiment provides an interactive processing method. Figure 1 This is a flowchart of an interactive processing method according to an embodiment of this application, such as... Figure 1 As shown, the process includes the following steps:

[0043] Step S102: In response to receiving the interaction request text from the target client, the system retrieves multiple preset skills stored in the skill library based on the interaction request text to obtain multiple candidate skills.

[0044] The target client mentioned above refers to the user terminal, mobile device, or application that initiates the interaction request. The target client has network communication capabilities, enabling it to receive natural language commands or descriptive text from the user and transmit the interaction request text to the server cluster where the background execution model resides via the communication network. The target client includes, but is not limited to, hardware devices such as smartphones, tablets, in-vehicle terminals, or personal computers. It is responsible for collecting the user's intent information and displaying the final processing result, serving as the starting point and data source of the entire intelligent interaction process.

[0045] The aforementioned interaction request text refers to natural language instructions, questions, or descriptive text input by the user through the target client. The interaction request text contains the user's specific task intent, such as querying data, performing an operation, or obtaining information. Interaction request text typically exists in string form and is preprocessed for matching with preset skills in the skill library. The content of interaction request text is semantically diverse, containing keywords, contextual information, or ambiguous instructions. The interaction processing system needs to perform word segmentation, noise reduction, and vectorization processing on the interaction request text to transform it into a machine-understandable format. This enables rapid location of relevant functional units within a massive skill set, ensuring the accuracy and relevance of the retrieval and improving the user experience.

[0046] The aforementioned skill library refers to a database, vector database, or distributed storage system that stores multiple preset skills. The skill library is responsible for maintaining the metadata, functional descriptions, vector representations, and logical constraints of available skills. It supports high-concurrency read and write operations, ensuring a rapid response to user requests. The internal structure of the skill library typically includes text and vector indexes to support mixed retrieval strategies. The skill library also possesses dynamic update capabilities, allowing the addition, modification, or deletion of preset skills to adapt to changes in logic and iterative upgrades of skills, ensuring the timeliness and completeness of the skill library content and providing rich resource support for the intelligent agent.

[0047] The aforementioned preset skills refer to predefined executable units in the skill library that possess specific functional implementation logic and parameters. Each preset skill corresponds to a specific function, such as data querying, file processing, or system control. Preset skills include attribute information such as identifiers, names, descriptive text, function vectors, and permission levels. Preset skills undergo standardization before being added to the library to ensure clear functional descriptions and accurate function vector representations. Preset skills are the fundamental components for intelligent agents to perform tasks. Through retrieval and filtering mechanisms, skill combinations that best meet user needs can be selected from numerous preset skills to achieve automated task processing and improve the level of intelligence.

[0048] The aforementioned candidate skills refer to a set of multiple pre-defined skills retrieved from the skill base based on the interactive request text and related to the request. The candidate skill set typically contains dozens or more pre-defined skills that score highly in text matching or vector similarity. Generating candidate skills is a preliminary screening process aimed at narrowing down the scope for subsequent accurate matching and conflict detection. Candidate skills are semantically related to the interactive request text but have not yet undergone logical conflict testing or priority ranking. The candidate skill set provides initial processing objects for subsequent skill evaluation, deduplication, and conflict resolution, serving as a crucial intermediate link connecting the retrieval results and the final execution decision, thus improving system resource utilization.

[0049] In one optional embodiment, the interaction processing system receives the interaction request text sent by the target client. First, it cleans and normalizes the interaction request text, removing irrelevant characters and stop words, and extracting core semantic features. Then, the interaction processing system inputs the normalized interaction request text into a retrieval engine for similarity matching. The retrieval engine traverses multiple preset skills stored in the skill base, calculating the relevance score between each preset skill and the interaction request text. Based on a preset recall quantity, the interaction processing system selects multiple preset skills with higher scores as candidate skills and returns a set of candidate skills and their corresponding matching scores for subsequent steps, ensuring retrieval efficiency. The aforementioned retrieval engine is a technical component used to quickly locate and recall relevant information from a large-scale dataset. In this embodiment, it refers to an intelligent agent skill retrieval module that supports hybrid retrieval, capable of parallelly executing text matching based on the Best Matching 25 (BM25) algorithm and semantic similarity matching based on vector space. By comprehensively calculating the relevance score, it efficiently recalls a set of candidate skills from the skill base that closely match the user's request.

[0050] In another optional embodiment, after receiving the interaction request text from the target client, the interaction processing system first uses an embedding model to convert the interaction request text into a high-dimensional vector representation. Simultaneously, the interaction processing system performs text matching between the interaction request text and the text description information of preset skills in the skill library, obtaining a text matching result. Next, the interaction processing system performs vector matching between the vector representation of the interaction request text and the function vectors of preset skills in the skill library, obtaining a vector matching result. Based on the text matching result and the vector matching result, a weighted fusion strategy is used to calculate the comprehensive score of each preset skill. Finally, the interaction processing system ranks the preset skills according to the comprehensive score and selects the top-ranked preset skills as a candidate skill set. The aforementioned embedding model is a machine learning model that converts natural language text into a numerical representation in a high-dimensional continuous vector space. The embedding model can capture semantic information, making semantically similar texts closer in the vector space. In this embodiment, the embedding model is used to convert user interaction requests into query vectors so that similarity calculations can be performed with the function vectors of skills, achieving semantic-based intelligent retrieval.

[0051] This application embodiment uses a hybrid retrieval strategy to quickly recall relevant candidate skills from a massive number of skills, narrowing the scope of subsequent processing, reducing computational overhead, and improving retrieval efficiency and the response speed of the interactive processing system.

[0052] Step S104: Based on the functional implementation logic of multiple candidate skills, perform conflict testing on the multiple candidate skills to obtain conflict test results. The conflict test results are used to characterize whether there are conflicts in the functional implementation logic of the multiple candidate skills.

[0053] The aforementioned functional implementation logic refers to the internal logical structure describing the specific execution steps, operation types, data access methods, and permission requirements of a preset skill. This logic defines how a skill interacts with external systems, including which data to read, which results to write, and what permissions are required. Functional implementation logic is typically stored in structured data or code form, containing key constraints such as operation types, data access modes, and execution permissions. By parsing the functional implementation logic, the interaction processing system can extract key characteristics of skill execution, used to determine whether there are issues such as mutually exclusive operations, data conflicts, or permission overlaps between different skills. This allows for an accurate assessment of the feasibility of skill combinations and prevents illegal or conflicting operations from being executed.

[0054] The conflict test results mentioned above refer to status indicators or data results that characterize whether there are mutual exclusions, conflicts, or compatibility issues among multiple candidate skills in terms of functional implementation logic. The conflict test results are generated by the system after compatibility comparison analysis and are used to indicate whether the current skill set meets the conditions for simultaneous execution. Conflict test results can be Boolean values ​​or structured data, explicitly indicating whether there are situations such as mutually exclusive operation types, data read / write conflicts, or permission overriding. Conflict test results serve as a decision-making basis, guiding the next step of the skill selection process. If conflicts exist, the interactive processing system will eliminate low-priority or high-risk skills according to the conflict type and arbitration rules, ensuring that the final output target skill combination is safe and logically consistent.

[0055] In one optional embodiment, the interaction processing system receives multiple candidate skills. First, it parses the functional implementation logic of each candidate skill, extracting the operation type, data access mode, and skill execution permissions. The interaction processing system transforms these elements into standardized logical constraints and establishes a set of constraints. Subsequently, the interaction processing system compares the logical constraints between the multiple candidate skills pairwise to check for mutually exclusive operation types, data read / write conflicts, or permission overriding. If any of these conflicts are found, the interaction processing system determines that a conflict exists. If no conflict is found, the conflict test result is determined to be non-conflicting, and this result is output to the subsequent filtering module.

[0056] In another optional embodiment, after receiving multiple candidate skills, the interaction processing system constructs a compatibility comparison matrix to record the logical relationships between each pair of candidate skills. The interaction processing system traverses the compatibility comparison matrix, performing in-depth analysis on each candidate skill based on the operation type, data access mode, and skill execution permissions in the functional implementation logic. The interaction processing system determines whether each pair of skills constitutes a mutually exclusive relationship based on a preset conflict rule base. Subsequently, it summarizes the results of all adversarial tests. If any skill pair is found to be in conflict, the conflict test result is marked as conflicting, and the specific conflict type is recorded. If no skill pair is in conflict, the conflict test result is marked as non-conflicting, ensuring the overall compatibility of the skill set for subsequent steps. The aforementioned compatibility comparison matrix represents a structured data table for systematically recording and analyzing the pairwise logical relationships between multiple candidate skills. By traversing the compatibility comparison matrix, the interaction processing system can efficiently identify operational mutual exclusion, data competition, or permission conflicts between skills, ensuring that the selected skill set does not interfere with each other in functional execution and is logically compatible. The aforementioned conflict rule base refers to a static configuration set of pre-defined skill mutual exclusion conditions in the storage logic, covering deterministic rules such as mutual exclusion of functions, mutual exclusion of read and write operations, and mutual exclusion of permissions. The interactive processing system uses this conflict rule base to quickly compare candidate skills, determine the conflict type, and perform arbitration, avoiding reliance on large model reasoning and improving the accuracy and efficiency of conflict resolution.

[0057] This application embodiment identifies and intercepts mutually exclusive or conflicting skill combinations in advance through conflict testing based on functional implementation logic, preventing invalid skills from entering the execution stage, reducing computing power waste, and improving the security and accuracy of intelligent agent decision-making.

[0058] Step S106: Based on the conflict test results, multiple candidate skills are screened to obtain the target skill.

[0059] The aforementioned target skills refer to one or more candidate skills ultimately determined for handling interaction requests after conflict testing and screening. Target skills are the set of superior skills retained after eliminating conflicting or low-priority skills from multiple candidate skills. Target skills must be highly relevant to the user interaction request, have internal logical compatibility, and not have issues such as mutually exclusive operations, data conflicts, or permission violations. Target skills will serve as input to the execution model, guiding it to generate specific processing results.

[0060] In one optional embodiment, the interaction processing system first determines whether the conflict test results indicate that there is no conflict among the multiple candidate skills. If the conflict test results indicate no conflict, the interaction processing system will directly identify the multiple candidate skills as a whole as multiple target skills without additional elimination operations. If the conflict test results indicate a conflict, the interaction processing system will identify the specific skill pairs with conflict and filter them according to preset conflict arbitration rules. Based on the conflict type, the priority indicators of related skills are compared, and conflicting skills with lower priority are eliminated, while higher priority skills are retained. Finally, a set of target skills without conflict is generated. The aforementioned conflict types can refer to execution permission conflicts, read / write operation conflicts, or operation logic conflicts, etc. The aforementioned priority indicators can refer to permission levels, call frequency, etc.

[0061] In another optional embodiment, the interaction processing system divides multiple candidate skills into conflict-free subsets and conflicting subsets. For the conflict-free subsets, the interaction processing system includes them in the target skill candidate pool. For the conflicting subsets, the system executes a fine-grained filtering strategy, extracting conflict type information for each group of conflicting skills. If the conflict type is an execution permission conflict, the system compares the permission levels of related skills and deletes conflicting skills with lower permission levels. If the conflict type is a read / write operation conflict or an operation logic conflict, the system compares the historical call frequency of related skills and deletes conflicting skills with lower call frequencies. The interaction processing system merges the remaining filtered skills to form the final target skill set. The aforementioned fine-grained filtering strategy refers to a differentiated adjudication mechanism for skill subsets with logical conflicts. Depending on the specific type of conflict, the fine-grained filtering strategy employs either priority comparison based on permission levels or statistical comparison based on historical call frequencies to eliminate conflicting skills with low priority or low frequency of use, thereby eliminating logical contradictions while retaining high-value skills and generating a conflict-free target skill set.

[0062] This application embodiment uses a screening mechanism based on conflict test results to eliminate skills with logical conflicts, reduce ineffective computing power consumption, improve the accuracy and security of agent decision-making, and ensure stable system operation.

[0063] Step S108: Send the target skill and interaction request text to the execution model. The execution model processes the interaction request text according to the target skill to obtain the processing result of the interaction request text.

[0064] The aforementioned execution model refers to a large language model, intelligent agent, or specialized task execution engine used to semantically understand and process interactive requests based on the target skill. The execution model possesses powerful natural language processing capabilities, enabling it to parse the deep semantics of the interactive request text. It can combine the functional implementation logic of the target skill to generate response content that meets user expectations. Execution models typically contain complex neural network structures capable of contextual reasoning and logical judgment. Upon receiving the target skill and the interactive request text, the execution model merges the two, constructing specific prompts or instructions to drive the model to generate the final processing result, realizing the transformation from user intent to concrete action.

[0065] The processing result of the aforementioned interaction request text refers to the final output content generated by the execution model after processing the interaction request text according to the target skill. The processing result includes text responses, structured data, executed commands, or system feedback information, directly responding to the specific needs raised by the user in the interaction request text. The processing result undergoes comprehensive reasoning by the execution model, combining the functional logic of the target skill and the semantic information of the interaction request text to ensure the accuracy and relevance of the output content. The processing result will be returned to the target client for the user to view or further operate, completing the closed loop of the entire intelligent interaction process. The quality of the processing result is the ultimate manifestation of the user's perceived system value and determines the quality of the user experience.

[0066] In one optional embodiment, the interaction processing system first converts the functional implementation logic of the target skill into structured instructions, which serve as system prompts for the execution model. Subsequently, the interaction processing system concatenates the interaction request text with the structured instructions to construct a complete input context. Upon receiving the input context, the execution model utilizes its pre-trained language understanding capabilities to parse the key intentions and constraints within the interaction request text. Based on the functional implementation logic of the target skill, the execution model invokes the corresponding internal parameters or external interfaces to simulate the skill execution process. The execution model generates a logically consistent processing result, which includes a direct response to the user's request and operational feedback. Finally, the processing result is output to the system interface. The aforementioned pre-trained language understanding capabilities refer to the deep semantic parsing and logical reasoning capabilities developed by the large language model after training on massive amounts of corpus data.

[0067] In another optional embodiment, the interactive processing system employs a chained reasoning strategy, decomposing the target skill into multiple sub-steps and sequentially inputting the logical constraints of each sub-step into the execution model. The execution model processes the interactive request text step-by-step, verifying the compatibility of the interactive request text with the target skill in each sub-step. The execution model dynamically adjusts the reasoning path to ensure that the final processing result meets the execution requirements of all target skills. When generating the processing result, the execution model introduces a self-verification mechanism to check for logical contradictions or operational conflicts. If no logical contradictions or operational conflicts exist, the execution model encapsulates the final processing result into a standard format and returns it to the caller. If logical contradictions or operational conflicts exist, the execution model will re-perform the reasoning or return an error message to ensure the reliability and consistency of the processing result and improve the robustness of the system.

[0068] This application embodiment accurately processes interactive request text through an execution model, combines the logical constraints of the target skill, generates high-quality processing results, and improves response accuracy and user experience.

[0069] In this embodiment, in response to receiving an interaction request text from the target client, multiple preset skills stored in the skill library are retrieved based on the interaction request text to obtain multiple candidate skills. Conflict testing is then performed on the multiple candidate skills based on their functional implementation logic to obtain conflict test results, whereby the conflict test results characterize whether there are conflicts in the functional implementation logic of the multiple candidate skills. Based on the conflict test results, the multiple candidate skills are filtered to obtain the target skill. The target skill and the interaction request text are sent to the execution model, which processes the interaction request text according to the target skill to obtain the processing result of the interaction request text. This embodiment filters multiple candidate skills from the skill library, enabling rapid identification of potentially available skills related to user intent, providing an initial candidate set for subsequent processing. Subsequently, conflict testing is performed on the candidate set to identify potential conflicts such as operational mutual exclusion, data competition, or permission overriding between skills before sending them to the execution model. Then, the multiple candidate skills are filtered based on the conflict detection results, reducing the number of skills entering the subsequent inference stage. Finally, based on the selected target skills, the execution model processes the interaction request text to obtain the processing result. This eliminates the need for the execution model to handle complex conflict arbitration and redundant reasoning, thereby shortening the prompt assembly and reasoning time, improving the response speed and execution stability of the agent interaction, and reducing the overall computing cost. This embodiment of the application achieves the goal of obtaining a set of target skills without logical conflicts and with high adaptability, thus avoiding invalid skills from entering the high-computing-power reasoning stage, and solving the technical problem of poor interaction processing effect of the target client in related technologies.

[0070] Optionally, multiple preset skills stored in the skill library are retrieved based on the interaction request text to obtain multiple candidate skills, including: transforming the interaction request text based on a preset vector dimension to obtain a target query vector, wherein the preset vector dimension is used to represent the vector dimension of the function vectors corresponding to multiple preset skills; and retrieving multiple preset skills based on the interaction request text and the target query vector to obtain multiple candidate skills.

[0071] The aforementioned preset vector dimension refers to the vector dimension used to represent the function vectors corresponding to multiple preset skills. The preset vector dimension is a fixed value that determines the size and complexity of the vector space. When building the skill library, the interaction processing system generates function vectors for each preset skill. The length of these function vectors must be consistent with the preset vector dimension to ensure the consistency of the vector space. The preset vector dimension is a key attribute of the embedded model output, affecting the accuracy and efficiency of vector retrieval. A higher preset vector dimension can capture richer semantic information but increases computational overhead. A lower preset vector dimension provides coarser semantic information but has lower computational overhead. During the design phase, the interaction processing system needs to select an appropriate preset vector dimension based on requirements and hardware resources to ensure the efficiency and accuracy of subsequent retrieval and similarity calculation, providing a foundation for hybrid retrieval strategies.

[0072] The aforementioned target query vector refers to the vector representation of the interaction request text, transformed through an embedding model, with the same dimension as the preset skill function vector. The target query vector is the mathematical expression of the interaction request text in semantic space, used for similarity matching with preset skills in the skill base. The generation of the target query vector relies on the embedding model, which can transform natural language into high-dimensional numerical vectors while preserving the core semantic information of the text. The dimension of the target query vector is consistent with the dimension of the preset vector, ensuring that distance calculations can be performed between them in the same vector space.

[0073] The aforementioned multiple candidate skills refer to a set of pre-defined skills retrieved from the skill base based on the interaction request text and related to the request. These multiple candidate skills are preliminary results generated under a hybrid retrieval strategy, including pre-defined skills that score highly in text matching and vector similarity. The number of multiple candidate skills is typically determined by the parameters of the retrieval algorithm, aiming to cover relevant skills and avoid missed recalls. These multiple candidate skills have not yet undergone logical conflict testing or priority ranking, serving as input data for subsequent processing. The generation process of multiple candidate skills combines semantic and keyword information to ensure the accuracy of recall.

[0074] The aforementioned function vectors refer to fixed-dimensional numerical arrays generated by transforming the functional descriptions of skills through a pre-trained semantic embedding model. Function vectors are a mathematical expression of the semantic information of skills, preserving the meaning characteristics of skills at the functional level, enabling computers to measure the semantic distance between skills through numerical calculations. In agent skill management, function vectors are stored in the index as vector retrieval fields, supporting efficient semantic similarity matching using the K-Nearest Neighbors (KNN) algorithm. By comparing the cosine similarity between function vectors, skills with overlapping functional semantics can be identified, thereby achieving semantic deduplication, ensuring the independence and diversity of the candidate skill set, and reducing the computational consumption of subsequent processing.

[0075] In one optional embodiment, after receiving the interaction request text, the interaction processing system first encodes the text using a pre-trained embedding model. The embedding model maps the text to a high-dimensional vector space, generating a target query vector. The dimension of the target query vector matches the dimension of the function vectors of preset skills in the skill base, satisfying the requirement of the preset vector dimension. Subsequently, the interaction processing system activates a hybrid retrieval engine. On one hand, it uses an inverted index to perform keyword matching on the interaction request text; on the other hand, it uses an approximate nearest neighbor search algorithm to find the preset skill closest to the target query vector in the vector index. The interaction processing system merges the text matching results and the vector matching results, filtering out multiple preset skills with higher scores to form multiple candidate skill sets for subsequent processing. The aforementioned inverted index refers to an unstructured text indexing technology that establishes a reverse mapping relationship between keywords and document identifiers. For example, the inverted index uses the BM25 algorithm to perform keyword matching on the interaction request text to locate skills containing specific semantic entities.

[0076] This application embodiment achieves accurate matching of natural language and skill semantics through a unified transformation of preset vector dimensions, improves the semantic understanding capability of retrieval, ensures high relevance and comprehensiveness of candidate skill recall, and lays a solid foundation for subsequent conflict resolution and accurate execution.

[0077] Optionally, multiple preset skills are retrieved based on the interaction request text and the target query vector to obtain multiple candidate skills, including: matching the interaction request text, the target query vector, and multiple preset skills to obtain text matching results and vector matching results of multiple preset skills; and filtering the multiple preset skills based on the text matching results and vector matching results of multiple preset skills to obtain multiple candidate skills.

[0078] The aforementioned text matching results refer to the relevance scores or ranking lists obtained after keyword or full-text searches based on the interactive request text and the textual descriptions of multiple preset skills. Text matching results reflect the degree of similarity between the interactive request text and the preset skills in a literal sense. Text matching results can capture accurate information such as entity names, specific terms, or numerical ranges in the interactive request text, compensating for the shortcomings of vector retrieval in detailed matching. The combined effect of text matching and vector matching results ensures that the search results possess both semantic relevance and literal accuracy, serving as a crucial basis for selecting multiple candidate skills and improving the overall precision of the search.

[0079] The vector matching results mentioned above refer to the relevance scores or ranking lists obtained after calculating the similarity between the target query vector and the functional vectors of multiple preset skills. Vector matching results reflect the semantic proximity between the interactive request text and the preset skills, typically calculated using metrics such as cosine similarity. Vector matching results can capture the deep semantic connections between the interactive request text and the preset skills; even if they do not share words literally, their relevance can be determined through vector distance. As another output of the hybrid retrieval strategy, vector matching results compensate for the limitations of text matching in semantic understanding. When fused with text matching results, they can more comprehensively evaluate the matching degree between preset skills and user intent, improving retrieval recall.

[0080] In one optional embodiment, the interaction processing system first performs text matching and vector matching in parallel. In text matching, the system utilizes inverted indexing technology to segment and extract keywords from the interaction request text, and retrieves preset skills matching the keywords from the skill base, generating multiple text matching results for preset skills. In vector matching, the system uses an approximate nearest neighbor search algorithm to calculate the cosine similarity between the target query vector and the function vector of each preset skill in the skill base, generating multiple vector matching results for preset skills. Subsequently, the system employs a reciprocal ranking fusion algorithm to deduplicate and merge the text matching and vector matching results, ranking the preset skills based on the fused comprehensive score, and selecting multiple preset skills with higher scores as candidate skills to ensure the diversity and accuracy of the search results. The aforementioned reciprocal ranking fusion algorithm is an ensemble algorithm for merging multiple search lists. The reciprocal ranking fusion algorithm evaluates the comprehensive relevance of each document by calculating the sum of the reciprocals of its position in the list. In the embodiments of this application, the inverse ranking fusion algorithm deduplicates and merges the text retrieval results and the vector semantic retrieval results, giving higher weight to high-frequency and high-ranking skills, thereby generating a high-quality hybrid retrieval candidate set.

[0081] This application embodiment achieves both literal accuracy and semantic relevance by fusing and filtering text matching results and vector matching results, thereby improving the accuracy and comprehensiveness of retrieval and recall, ensuring the high quality of multiple candidate skills, and providing a foundation for subsequent conflict resolution and precise execution.

[0082] Optionally, the interaction request text, the target query vector, and multiple preset skills are matched to obtain text matching results and vector matching results of multiple preset skills, including: performing text matching on the interaction request text and the text description information of multiple preset skills to obtain text matching results of multiple preset skills; and performing vector matching on the target query vector and the function vectors of multiple preset skills to obtain vector matching results of multiple preset skills.

[0083] The aforementioned text description information refers to natural language text data associated with multiple preset skills, used to describe the function, purpose, and operation method of the skills. The text description information covers key information such as the core capabilities, applicable scenarios, and input / output parameters of the skills, aiming to allow the system to understand the specific role of the skills through natural language. During the retrieval process, the interactive processing system extracts keywords and semantic features from the text description information and compares them with the interactive request text to assess the literal relevance of the two. The accuracy and completeness of the text description information determine the reliability of the text matching results, providing accurate literal matching basis for hybrid retrieval.

[0084] In one optional embodiment, the interaction processing system first preprocesses the interaction request text, including word segmentation and stop word removal, and extracts key entities and feature words. Then, using inverted indexing technology, the system searches the skill base for preset skills that match these key feature words, calculates the text matching score for each preset skill, and generates multiple text matching results for preset skills. Simultaneously, the system uses an approximate nearest neighbor search algorithm to find the function vectors of preset skills that are close to the target query vector in the vector index, calculates the cosine similarity, and generates multiple vector matching results for preset skills. The system stores both matching results separately and labels them with corresponding preset skill identifiers to prepare for subsequent fusion and filtering, ensuring the efficiency and parallelism of the retrieval process.

[0085] This application embodiment balances literal accuracy and semantic relevance by independently calculating and fusing text matching results and vector matching results, thereby improving the accuracy and comprehensiveness of retrieval and ensuring the high quality of multiple candidate skills.

[0086] Optionally, based on the text matching results and vector matching results of multiple preset skills, multiple preset skills are filtered to obtain multiple candidate skills, including: determining the skill evaluation scores of multiple preset skills based on the text matching results and vector matching results of multiple preset skills; ranking the multiple candidate skills based on the skill evaluation scores of multiple preset skills to obtain ranking results; and filtering the multiple preset skills based on the ranking results to obtain multiple candidate skills.

[0087] The skill assessment score mentioned above refers to a numerical score calculated by combining text matching and vector matching results to measure the relevance of a preset skill to the interactive request text. The skill assessment score is typically obtained by weighted fusion of text matching and vector matching scores, reflecting the comprehensive matching degree of the preset skill in both literal and semantic dimensions. A higher skill assessment score indicates a stronger relevance between the preset skill and the user's intent. The skill assessment score quantifies the value of each preset skill, helping the system quickly identify high-quality candidates from a vast pool of skills, ensuring that the final selected set of candidate skills is highly relevant and practical.

[0088] The ranking result mentioned above refers to an ordered list generated by arranging multiple preset skills according to their skill evaluation scores from highest to lowest or lowest to highest. The ranking result shows the relative strength and weakness of each preset skill in the candidate pool, allowing the system to prioritize high-scoring skills. The ranking result typically includes the identifier of the preset skill, its corresponding ranking position, and its evaluation score. Through the ranking result, it is possible to clearly determine which skills are at the top and which are at the bottom, thus providing a clear reference standard for subsequent truncation or threshold screening. The ranking result is a preliminary step in the screening process; by ordering the skills, it simplifies the screening logic, improves processing efficiency, and ensures that highly relevant skills are prioritized.

[0089] In one optional embodiment, the interactive processing system first receives text matching results and vector matching results, and extracts the text matching score and vector matching score corresponding to each preset skill. Then, according to preset fusion weights, the text matching score and vector matching score are weighted and summed to calculate the skill evaluation score for each preset skill. Next, the interactive processing system normalizes the skill evaluation scores of all preset skills to eliminate the influence of different units of measurement. Finally, based on the normalized skill evaluation scores, the interactive processing system sorts the multiple preset skills from highest to lowest score, generating a ranking result. Based on the ranking result, the interactive processing system selects the top-ranked preset skills as multiple candidate skills.

[0090] This application embodiment achieves quantitative fusion and orderly arrangement of multi-path retrieval results through the determination and sorting of skill assessment scores, thereby improving the accuracy and efficiency of candidate skill selection, ensuring that subsequent processing is based on a high-quality set of relevant skills, and improving the overall performance of agent retrieval and recall.

[0091] Optionally, based on the text matching results and vector matching results of multiple preset skills, the skill evaluation scores of multiple preset skills are determined, including: determining the text matching scores of multiple preset skills based on the text matching results of multiple preset skills; determining the vector matching scores of multiple preset skills based on the vector matching results of multiple preset skills; and weighting and summing the text matching scores and vector matching scores based on a preset weighting ratio to obtain the skill evaluation scores of multiple preset skills.

[0092] The text matching score mentioned above refers to a numerical score obtained after retrieving text fields such as query statements and skill names. The text matching score reflects the degree of literal overlap between the query statement and the skill text description, as well as the strength of keyword matching; a higher score indicates stronger text relevance. The text matching score is used to capture accurate hard-match information such as entity names, specific terms, or numerical ranges. It compensates for the shortcomings of vector retrieval in literal matching, ensuring that even accurately expressed but semantically scattered skills can be accurately recalled, thus guaranteeing the precision of search results.

[0093] The aforementioned vector matching score refers to a numerical score obtained by calculating the similarity between the query vector and the skill function vector. The vector matching score is typically measured using cosine similarity, reflecting the semantic proximity between the query intent and the core function of the skill; a higher score indicates a stronger semantic relevance. The vector matching score is used to capture the deep semantic association between user queries and skill functions, identifying synonyms or skills with similar functions but different descriptions. The vector matching score provides meaning-based retrieval capabilities, ensuring that even if the query term and skill name are inconsistent, highly relevant skills can be retrieved, improving the recall rate and semantic understanding depth of search results.

[0094] In one optional embodiment, the interactive processing system first receives text matching results and vector matching results, and extracts the text matching score and vector matching score corresponding to each preset skill. Then, the interactive processing system calculates the skill evaluation score for each preset skill by weighted summation according to preset fusion weights. The interactive processing system normalizes the skill evaluation scores of the preset skills to eliminate the influence of different units of measurement. Next, the interactive processing system sorts the multiple preset skills from highest to lowest score based on the normalized skill evaluation scores, generating a ranking result. Finally, based on the ranking result, the interactive processing system selects the top-ranked preset skills as multiple candidate skills, ensuring that the candidate skill set includes highly relevant skill units, achieving accurate screening based on scores.

[0095] This application embodiment achieves quantitative fusion and orderly arrangement of multi-path retrieval results by determining and sorting skill assessment scores, thereby improving the accuracy and efficiency of candidate skill selection, ensuring that subsequent processing is based on a high-quality set of relevant skills, and improving the overall performance of agent retrieval and recall.

[0096] Optionally, multiple preset skills are filtered based on the ranking results to obtain multiple candidate skills, including: filtering multiple preset skills based on the ranking results to obtain multiple initial skills; performing similarity matching between any two initial skills to obtain a similarity matching result, wherein the similarity matching result is used to indicate whether the similarity between any two initial skills is greater than a preset threshold; and removing duplicates from the multiple initial skills based on the similarity matching result to obtain multiple candidate skills.

[0097] The aforementioned initial skills refer to a set of pre-selected skills that rank highly, initially selected from multiple pre-defined skills based on the ranking results. These initial skills are the direct processing objects in the semantic deduplication stage, including skill units that scored high in the comprehensive evaluation but have not yet undergone semantic overlap checks. The generation of these initial skills relies on truncation or threshold filtering of the ranking results, aiming to narrow the scope of subsequent calculations and improve deduplication efficiency. While these initial skills differ literally, they exhibit a high degree of functional semantic overlap. As input to the deduplication algorithm, these initial skills need further redundancy identification through similarity matching to ensure the semantic uniqueness and diversity of the final output candidate skill set, providing concise and high-quality input data for subsequent conflict resolution.

[0098] The similarity matching result mentioned above refers to a quantitative indicator representing the degree of semantic overlap between any two initial skills. The similarity matching result is typically obtained by calculating the cosine similarity or other distance metrics between the corresponding function vectors of two initial skills, used to measure the closeness of their functional descriptions. If the similarity matching result is greater than a preset threshold, the two initial skills are determined to have semantic overlap or functional redundancy. If it is less than or equal to the preset threshold, they are determined to be semantically independent. The similarity matching result serves as the basis for deduplication logic, determining which skills will be considered duplicates and eliminated, and which skills will be retained, ensuring that the final candidate skill set does not contain redundant skills with similar functions, thus improving the purity of the skill library.

[0099] The aforementioned similarity matching result refers to a quantitative indicator or judgment state used to characterize the degree of semantic overlap between two initial skills by calculating the cosine similarity of their functional vectors. The similarity matching result is typically represented by a value between 0 and 1; the closer the value is to 1, the higher the semantic overlap in the functional descriptions of the two skills, and the greater the probability of functional redundancy. In the deduplication process, the interactive processing system compares the similarity matching result with a preset semantic overlap threshold. If the similarity matching result exceeds the threshold, the two initial skills are determined to be semantically redundant skills, requiring further deduplication screening. If the threshold is not exceeded, they are determined to be semantically independent. The similarity matching result serves as the basis for performing the lightweight vector similarity semantic deduplication step, directly determining which skills will be eliminated to ensure the purity of the candidate skill pool and prevent functionally similar skills from entering subsequent high-computational stages.

[0100] The aforementioned preset threshold refers to the numerical similarity limit used to determine whether any two initial skills constitute semantic duplication. The preset threshold is a pre-defined scalar value, representing the upper limit of the interactive processing system's tolerance for skill semantic similarity. The size of the preset threshold affects the degree of deduplication; a higher preset threshold allows more similar skills to coexist, while a lower preset threshold filters out skills with lower similarity. Setting a reasonable preset threshold can balance the skill coverage and deduplication effect, and is a key parameter for improving the quality of candidate skills.

[0101] In one optional embodiment, the interaction processing system first receives multiple initial skills and their corresponding function vectors. The system constructs a double-loop structure, iterating through each pair of initial skills. For each pair of initial skills, the cosine similarity between the function vectors is calculated to obtain a similarity matching result. The system compares this similarity matching result with a preset threshold. If the similarity matching result is greater than the preset threshold, the skill evaluation scores corresponding to the two initial skills are compared, and the initial skill with the lower score is eliminated, while the initial skill with the higher score is retained. The system repeats the comparison and elimination process until the similarity matching result between the initial skill pairs is no greater than the preset threshold. Finally, the remaining set of initial skills constitutes multiple candidate skills.

[0102] This application embodiment uses a deduplication mechanism based on similarity matching results to accurately eliminate redundant skills with semantic overlap, reducing the computational power consumption of subsequent processing, improving the purity and diversity of the candidate skill set, and ensuring the efficiency and stability of the agent's execution process.

[0103] Optionally, multiple preset skills are filtered based on the ranking results to obtain multiple initial skills, including: determining the preset skills that are ranked before a preset position in the ranking results as multiple initial skills; or, extracting a preset number of preset skills based on the ranking results as multiple initial skills.

[0104] The aforementioned preset ranking refers to the ranking threshold value used to define the initial skill selection range in the sorting results. The preset ranking is an integer index representing a specific position in the skill assessment score ranking list. Preset skills ranking higher than the preset ranking indicate that they performed better in the overall score and are located at the top of the list. The process of determining the preset ranking is usually based on the system's expected control over the size of the candidate skills, aiming to ensure that skills with higher rankings are selected for subsequent stages. The setting of the preset ranking determines the upper limit of the initial skill set size; the higher the position, the higher the relevance of the initial skills. By setting the preset ranking, the system can dynamically adapt to different numbers of search results with a fixed ranking threshold, ensuring that the skill set input to the deduplication module always remains in a high-quality range, providing a stable data foundation for subsequent processing.

[0105] The aforementioned preset quantity refers to a fixed numerical indicator used to limit the size of the initial skill set. The preset quantity is a positive integer representing the maximum number of skill items allowed to enter the deduplication process. Extracting a preset number of skills based on the ranking results indicates that regardless of how many skills are retrieved, the interactive processing system retains the top-rated skills. The setting of the preset quantity comprehensively considers computational resources, response latency, and the time complexity of the deduplication algorithm. By fixing the preset quantity, the interactive processing system can ensure the predictability and stability of the processing flow, avoiding performance fluctuations caused by variations in the number of retrieval results. The preset quantity affects the computational overhead of subsequent deduplication and conflict resolution, ensuring that the system maintains an efficient operating rhythm even in scenarios with massive skill sets.

[0106] In one optional embodiment, the interactive processing system first receives the ranking results after skill assessment scores, which include preset skills participating in the scoring and their ranking indices. Then, the interactive processing system determines a specific preset rank index value based on preset configuration parameters. The interactive processing system iterates through the ranking results, identifying preset skills whose rank index value is less than or equal to the preset rank index value. The interactive processing system collects the identified preset skills to form a new set. The interactive processing system performs an integrity check on this set to ensure no omissions or duplications. Finally, this set consisting of the top-ranked preset skills is defined as multiple initial skills.

[0107] This application embodiment extracts high-quality initial skills from the sorting results through a screening strategy based on a preset rank or preset number, controls the scale of subsequent processing, improves deduplication efficiency, and ensures high relevance and processing stability of the candidate skill set.

[0108] Optionally, multiple initial skills are deduplicated based on similarity matching results to obtain multiple candidate skills, including: dividing multiple initial skills based on similarity matching results to obtain at least one first initial skill and multiple second initial skills, wherein the multiple second initial skills are associated initial skills with a similarity greater than a preset threshold, and at least one first initial skill is other skills among the multiple initial skills besides the multiple second initial skills; deduplicating the multiple second initial skills to obtain deduplicated target skills; and merging the deduplicated target skills and at least one first initial skill to obtain multiple candidate skills.

[0109] The aforementioned first initial skill refers to the set of skills other than the multiple second initial skills among the multiple initial skills. The first initial skill is those skill units that, in similarity matching, do not exhibit a similarity greater than a preset threshold with other initial skills. At least one first initial skill is determined to be semantically independent, without functional overlap or redundancy, and therefore can be retained without complex deduplication. The first initial skill ensures that unique and conflict-free high-scoring skills are retained. The first initial skill is merged with the deduplicated target skills to form the final multiple candidate skills, ensuring the comprehensiveness and purity of the candidate skill set, reducing unnecessary computational overhead, and improving processing efficiency.

[0110] The aforementioned multiple second initial skills refer to the set of associated initial skills among multiple initial skills whose similarity to other initial skills exceeds a preset threshold. These second initial skills constitute a subset of semantically overlapping skills, indicating functional redundancy or high similarity among them. Second initial skills are marked as requiring further processing due to their similarity exceeding the preset threshold; typically, a deduplication algorithm is used to select superior representatives. Identifying multiple second initial skills is a crucial step in the deduplication process, aiming to locate all potential conflict points. Second initial skills cannot be directly used as final candidate skills because they lack uniqueness; further cleaning and optimization are necessary to ensure that the final output skill set is semantically independent and non-repetitive, thereby guaranteeing the efficiency of the agent's execution.

[0111] The deduplicated target skills mentioned above refer to the superior set of skills obtained after deduplicating multiple second initial skills. The deduplication process typically involves comparing the similarity matching results and skill evaluation scores among multiple second initial skills, retaining skills with higher scores and similarities below a preset threshold with other retained skills. The deduplicated target skills represent the best representatives in each semantic overlap cluster, ensuring that highly relevant functions are not lost while eliminating redundancy. The target skills are the result of a selection process, avoiding duplicate execution or logical conflicts. The deduplicated target skills share the same semantic independence characteristics as at least one first initial skill, forming part of the conflict-free candidate skill pool and determining the quality and accuracy of the agent's subsequent task execution.

[0112] In one optional embodiment, the interaction processing system first traverses each pair of initial skill combinations from multiple initial skills, calculating the similarity matching result. Based on a comparison of the similarity matching result with a preset threshold, the interaction processing system identifies initial skill pairs with similarity scores greater than the preset threshold. The interaction processing system labels the initial skills involving these similar pairs as second initial skills, which constitute a semantically overlapping network. The interaction processing system retains the remaining initial skills not labeled as second initial skills as at least one first initial skill. Subsequently, the interaction processing system performs cluster analysis on the multiple second initial skills. Within each cluster, the interaction processing system extracts the initial skills with higher skill evaluation scores as deduplicated target skills. Finally, the interaction processing system combines the deduplicated target skills with at least one first initial skill to obtain multiple candidate skills, ensuring that the final set is free of redundancy and contains independent high-scoring skills.

[0113] This application's embodiments accurately identify and eliminate semantically overlapping skills by segmenting and deduplicating based on similarity matching results, while retaining independent high-scoring skills. This improves the purity and diversity of the candidate skill set, avoids execution conflicts, and ensures the efficiency and accuracy of the agent's decision-making.

[0114] Optionally, deduplication is performed on multiple second initial skills to obtain deduplicated target skills, including: dividing the multiple second initial skills into multiple skill combinations; for any one of the multiple skill combinations, deduplicating the initial skills in that skill combination based on the skill evaluation scores of the initial skills in that skill combination to obtain the deduplicated skills corresponding to that skill combination; and determining the deduplicated target skills based on the deduplicated skills corresponding to the multiple skill combinations.

[0115] The aforementioned multiple skill combinations refer to several non-overlapping subsets formed by dividing multiple second initial skills based on semantic relevance. Each skill combination contains a group of initial skill units that are highly similar in function and semantics or have overlapping relationships. This division is typically based on a connected graph constructed from similarity matching results or a clustering algorithm, ensuring that skills within the same combination are related to each other, while skills in different combinations are semantically far apart. Dividing multiple second initial skills into multiple skill combinations aims to decompose the complex overall deduplication problem into multiple independent local deduplication tasks, thereby reducing computational complexity. This grouping strategy allows the system to process each combination in parallel or sequentially, ensuring that redundant skills within each semantic cluster are processed independently, improving the efficiency and logical clarity of large-scale skill deduplication.

[0116] The deduplicated skills corresponding to any of the above skill combinations refer to the subset of superior initial skills selected for a specific skill combination by comparing the skill evaluation scores of the initial skills within that combination. The deduplicated skills typically include one or more initial skills with high scores within the combination, depending on the degree of overlap within the combination and the preset deduplication strategy. If the skills within the combination highly overlap, only those with higher scores are retained. If there is partial overlap, multiple non-overlapping high-scoring skills are retained. This process eliminates semantic redundancy within the combination, ensuring that the output skills are unique and representative within the local semantic cluster.

[0117] In one optional embodiment, the interaction processing system first traverses multiple second initial skills and uses a clustering algorithm to divide them into multiple skill combinations based on similarity matching results. The skill similarity within each combination is higher than a preset threshold, while the similarity between combinations is lower. For any given skill combination, the interaction processing system extracts the skill evaluation scores of the initial skills within the combination. The system sorts the scores in descending order and selects the initial skills with higher scores as the deduplicated skills for the combination. If the higher-scoring skill has a high degree of overlap with other skills, it is directly retained. If there is a skill with the second-highest score and its overlap is lower than the threshold, it is also retained. The interaction processing system repeats the above process to generate corresponding deduplicated skills for each skill combination. Finally, the system merges the deduplicated skills generated for each combination to form the deduplicated target skill, ensuring global redundancy.

[0118] This application's embodiments use a grouping deduplication strategy to transform global redundancy elimination into local better selection, reducing computational complexity and ensuring that highly relevant skills are retained to the maximum extent while eliminating semantic overlap, thereby improving the quality and processing efficiency of the candidate skill set.

[0119] Optionally, conflict testing is performed on multiple candidate skills based on their functional implementation logic to obtain conflict test results. This includes: comparing the compatibility of multiple candidate skills based on their functional implementation logic to obtain comparison results, wherein the comparison results indicate whether any one of the following exists: mutually exclusive operation types, data read / write conflicts, and permission overriding. In response to the comparison results indicating that multiple candidate skills have any one of these issues, the conflict test results determine that the functional implementation logic of the multiple candidate skills is conflicting. In response to the comparison results indicating that multiple candidate skills do not have any one of these issues, the conflict test results determine that the functional implementation logic of the multiple candidate skills is not conflicting.

[0120] The aforementioned functional implementation logic refers to the technical specifications that define the specific execution behavior, operation steps, and internal processing rules of each candidate skill. The functional implementation logic describes the process from receiving input to generating output, including data flow, control conditions, and execution actions. It serves as the basis for evaluating the compatibility between skills and determines how a skill interacts with the system environment and other skills during runtime. By analyzing the functional implementation logic of multiple candidate skills, a deeper understanding of the specific mechanism of each skill can be achieved.

[0121] The aforementioned compatibility comparison refers to the process of comparing the functional implementation logic of multiple candidate skills pairwise or as a whole to assess whether they can work collaboratively in a coexisting environment. Compatibility comparison aims to identify resource contention, state conflicts, or operational contradictions that exist between different skills during execution. This process typically involves analyzing skill input / output interfaces, resource access permissions, and operational timing. Through compatibility comparison, the interactive processing system can quantitatively or qualitatively determine the executability of skill combinations. If the comparison results show that skills can collaborate seamlessly, they are considered compatible. If potential conflicts are found, they are considered incompatible. Compatibility comparison determines the selection of subsequent conflict resolution strategies, ensuring the logical self-consistency of the skill combination.

[0122] The comparison results mentioned above refer to the judgment data regarding whether there are specific types of conflicts among multiple candidate skills, obtained through compatibility comparison. The comparison result is a logical state or list that clearly identifies specific issues such as mutually exclusive operation types, data read / write conflicts, or permission overriding. If the comparison result indicates a conflict, the specific type of conflict and the skills involved are listed. If the comparison result indicates no conflict, the skills are confirmed to be functionally compatible. The comparison result serves as an intermediate decision-making basis, directly triggering the subsequent conflict test result judgment logic. The comparison result provides accurate conflict information, enabling the interaction processing system to handle conflicts specifically, rather than blindly rejecting or accepting everything, thus improving the efficiency of conflict handling and providing reliable data support for ultimately determining the effectiveness of the skill combination.

[0123] The aforementioned operation type mutual exclusion refers to a conflicting state where multiple candidate skills cannot be executed simultaneously or consecutively due to their opposite operation types. Operation type mutual exclusion typically occurs when two skills attempt to perform opposite operations on the same resource, such as one skill performing a write operation while another performs a delete operation, or one skill starting a transaction while another commits it. This mutual exclusion can lead to data inconsistency or system state errors. Identifying operation type mutual exclusion helps the system provide early warnings during the skill scheduling phase, preventing the execution of illegal operation sequences. By detecting operation type mutual exclusion, interactive processing systems can improve the skill execution order or eliminate conflicting skills, ensuring operational consistency and serving as a mechanism for maintaining system data integrity and logical correctness.

[0124] The aforementioned data read / write conflict refers to the risk of data inconsistency or loss caused by improper read / write order or permissions when multiple candidate skills access shared data resources. Data read / write conflicts are common in concurrent scenarios, such as when one skill reads data while another skill modifies that data, without a locking mechanism or version control. Such conflicts can lead to reading stale data, overwriting uncommitted changes, or data loss. Detecting data read / write conflicts requires analyzing the data dependencies and access patterns of skills. By identifying data read / write conflicts, the system can introduce concurrency control strategies or adjust execution logic to ensure the security and consistency of data access. This is crucial for ensuring data accuracy when agents collaborate with multiple skills, preventing logical errors or performance degradation caused by concurrent access, and improving the robustness and reliability of the system.

[0125] The aforementioned permission overlap refers to security policy conflicts or unauthorized access risks arising from permission requirements or changes during the execution of multiple candidate skills. Permission overlap manifests as a high-privilege skill overriding the settings of a low-privilege skill, or a low-privilege skill attempting to perform operations requiring high privileges. This conflict leads to security vulnerabilities or permission confusion, preventing the system from correctly implementing access control. Detecting permission overlap requires analyzing the permission declarations and actual execution context of each skill. By identifying permission overlap, interactive processing systems can ensure consistent permission management and prevent unauthorized access or operations.

[0126] In one optional embodiment, the interaction processing system first parses the functional implementation logic of multiple candidate skills, extracting the operation type, data access scope, and permission requirements of each skill. The interaction processing system constructs a conflict detection matrix, pairing multiple candidate skills together and comparing their compatibility one by one. For each pair of skills, the interaction processing system checks for mutually exclusive operation types. Next, the interaction processing system checks for data read / write conflicts, analyzing the time series of data access and locking mechanisms to determine if there are concurrent access risks. Finally, the interaction processing system checks for permission overriding, comparing the permission levels and declarations of the two skills to confirm whether there are permission downgrades or unauthorized accesses. Finally, the above check results are summarized to generate a comparison result. If the comparison result marks any conflict, the conflict test result is determined to be conflict-free. If the comparison result marks no conflict, the conflict test result is determined to be conflict-free.

[0127] This application's embodiments accurately identify potential risks in the functional implementation logic through compatibility comparison and multiple conflict detection, ensuring that conflict-free skill combinations enter the execution phase, thereby improving the stability and security of intelligent agent task execution and avoiding system anomalies or data errors caused by logical conflicts.

[0128] Optionally, a compatibility comparison is performed on multiple candidate skills based on their functional implementation logic to obtain comparison results. This includes: parsing the functional implementation logic of multiple candidate skills to obtain their operation types, data access modes, and skill execution permissions; structuring the operation types, data access modes, and skill execution permissions to obtain logical constraints; and performing a compatibility comparison on multiple candidate skills based on the logical constraints to obtain comparison results.

[0129] The aforementioned operation type refers to the technical attribute identifier describing the specific action category performed by a candidate skill. The operation type clarifies whether the skill involves basic operations such as reading, writing, modifying, deleting, or creating data or resources. Operation types determine whether there are mutually exclusive factors between skills, such as simultaneous write and delete operations leading to inconsistent data states. By extracting the operation types of multiple candidate skills, a semantic mapping of skill actions can be established, providing a foundational dimension for subsequent conflict detection. The structured representation of operation types enables the system to automatically compare the action compatibility between different skills, identify potential logical conflicts, ensure the safety and consistency of actions during the execution of skill combinations, and avoid the generation of illegal or contradictory operation sequences.

[0130] The aforementioned data access patterns refer to a set of behavioral characteristics describing how candidate skills interact with the data storage system. Data access patterns encompass specific behaviors such as data reading, data writing, data locking, data sharing, and data isolation. Data access patterns reflect the skill's consumption of data resources and its concurrency characteristics. By analyzing data access patterns, a data dependency graph can be constructed, and potential data read / write conflicts can be predicted. This pattern recognition mechanism ensures that the agent can correctly manage data resources when multiple skills collaborate, guaranteeing data consistency and integrity, and preventing logical errors or performance bottlenecks caused by improper concurrent access.

[0131] The aforementioned skill execution permissions refer to the system resource permissions that candidate skills can access, the scope of operations they can perform, and the required security level. Skill execution permissions clearly define the role and operational boundaries of a skill within the system. Permission information is the basis for determining whether there are risks of permission overlap or unauthorized access between skills. By parsing skill execution permissions, the system can assess whether permission conflicts or security policy violations will occur when multiple skills run simultaneously. For example, a low-privilege skill attempting to override the security settings of a high-privilege skill can lead to system vulnerabilities. Structured processing of skill execution permissions ensures that each skill operates within its authorized scope, maintains the system's security isolation mechanism, and prevents security incidents or policy failures caused by permission confusion.

[0132] The aforementioned logical constraints refer to a set of rules used to determine skill compatibility, formed by structuring operation types, data access patterns, and skill execution permissions. These logical constraints include elements such as resource identifiers, operation restrictions, dependencies, and permission requirements. These logical constraints constitute the benchmark for compatibility comparison; the interactive processing system uses these constraints to evaluate whether multiple candidate skills will conflict in a coexisting environment. Logical constraints transform unstructured functional descriptions into machine-understandable logical rules. By comparing based on logical constraints, the interactive processing system can accurately identify issues such as operation mutual exclusion, data conflicts, and permission overriding, ensuring the accuracy and efficiency of conflict detection and providing a reliable decision-making basis for subsequent skill selection.

[0133] In one optional embodiment, the interaction processing system first receives functional implementation logic descriptions of multiple candidate skills and uses natural language processing technology to extract the operation type, data access mode, and skill execution permissions for each skill. The interaction processing system maps this extracted information to a predefined ontology model to construct standardized skill feature vectors. Subsequently, the interaction processing system structurally encapsulates the operation type, data access mode, and skill execution permissions to form logical constraints. The interaction processing system traverses multiple candidate skills, extracts the logical constraints corresponding to each skill, and constructs a constraint comparison matrix. Based on the constraints in the matrix, it checks whether there are resource overlaps and incompatible operations, conflicting access modes, or overlapping permissions between any two skills. Finally, the interaction processing system summarizes the tags generated during the checking process, generates comparison results, and clearly identifies whether any type of conflict exists.

[0134] This application's embodiments realize logic through parsing and transform it into structured logical constraints, thereby automating compatibility comparison, accurately identifying potential conflicts at the operation, data, and permission levels, ensuring that logically compatible skill combinations enter the execution phase, and improving the stability and security of intelligent agent task execution.

[0135] Optionally, multiple candidate skills are screened based on the conflict test results to obtain the target skill, including: in response to the conflict test results indicating that the functional implementation logic of multiple candidate skills does not conflict, multiple candidate skills are determined as the target skill; in response to the conflict test results indicating that the functional implementation logic of multiple candidate skills does conflict, any two candidate skills among the multiple candidate skills whose functional implementation logic conflicts, and multiple candidate skills are screened based on the conflict type of any two candidate skills to obtain the target skill.

[0136] In one optional embodiment, the interaction processing system first receives the conflict test results. If the conflict test results indicate that the functional implementation logic of multiple candidate skills does not conflict, the system directly determines that the set of multiple candidate skills meets the execution conditions and identifies the entire set of candidate skills as the target skill, without requiring additional filtering. If the conflict test results indicate that a conflict exists, the interaction processing system locates any two candidate skills based on the conflict detection results and reads their conflict type. The interaction processing system arbitrates the conflict between any two candidate skills based on preset resolution rules. The interaction processing system removes conflicting skills determined to be of low priority or high risk from the multiple candidate skills, retaining another compatible skill. The interaction processing system repeats the above process until the conflict is resolved, and the final set of remaining conflict-free skills is identified as the target skill.

[0137] This application embodiment uses a dynamic filtering mechanism based on conflict test results to intelligently handle both conflict-free and conflict-containing scenarios, ensuring that the final output set of target skills is functionally compatible, effectively avoiding execution conflicts, and improving the stability and success rate of intelligent agent task execution.

[0138] Optionally, the conflict types include execution permission conflicts, read / write operation conflicts, and operation logic conflicts. Multiple candidate skills are filtered based on the conflict types of any two candidate skills to obtain the target skill. This includes: in response to an execution permission conflict, deleting the first candidate skill from any two candidate skills based on their permission levels to obtain the target skill, wherein the permission level of the first candidate skill is lower than that of the other candidate skill; and in response to read / write operation conflicts and operation logic conflicts, deleting the second candidate skill from any two candidate skills based on their call frequency to obtain the target skill, wherein the call frequency of the second candidate skill is lower than that of the other candidate skill.

[0139] The aforementioned execution permission conflict refers to the security policy conflict or unauthorized access risk caused by permission requirements or changes during the execution of multiple candidate skills. Execution permission conflicts manifest as high-privilege skills overriding low-privilege skills, or low-privilege skills attempting to perform operations requiring high privileges. Execution permission conflicts can lead to security vulnerabilities or permission confusion, preventing the system from correctly implementing access control. Identifying execution permission conflicts helps the system provide early warnings during the skill scheduling phase, preventing the execution of illegal operation sequences. By detecting execution permission conflicts, the system can improve the skill execution order or eliminate conflicting skills, ensuring operational consistency. This is a crucial mechanism for maintaining system data integrity and logical correctness, preventing interruptions or data corruption due to operational conflicts.

[0140] The aforementioned permission levels refer to the quantitative identifiers of the system resources a candidate skill can access, the scope of operations it can perform, and the required security level. Permission levels clearly define the skill's role and operational boundaries, such as read-only, read-write, or administrator privileges. By resolving permission levels, the interactive processing system can assess whether permission conflicts or security policy violations will occur when multiple skills run simultaneously. For example, a low-privilege skill attempting to override the security settings of a high-privilege skill can lead to system vulnerabilities. The structured representation of permission levels ensures that each skill operates within its authorized scope, maintains the system's security isolation mechanism, and prevents security incidents or policy failures caused by permission confusion.

[0141] The aforementioned first-choice skill refers to a candidate skill deemed to have a lower permission level and should be deleted in a permission conflict scenario. The first-choice skill conflicts with another candidate skill in terms of execution permissions, and its permission level is lower than the other candidate skill. Based on the permission level, the interactive processing system determines that the first-choice skill is at a disadvantage in terms of security policy; retaining the first-choice skill would lead to security risks or confusion in permission management. Deleting the first-choice skill ensures that each skill in the final target skill set has permissions matching its responsibilities, preventing low-permission skills from interfering with the execution of high-permission tasks, thereby improving system security and stability.

[0142] The aforementioned read-write conflict refers to the risk of data inconsistency or loss caused by improper read-write order or permissions when multiple candidate skills access shared data resources. Read-write conflicts are common in concurrent scenarios, such as when one skill reads data while another skill modifies that data, without a locking mechanism or version control. Such conflicts can lead to reading stale data, overwriting uncommitted changes, or data loss. Detecting read-write conflicts requires analyzing the data dependencies and access patterns of skills. By identifying read-write conflicts, interactive processing systems can introduce concurrency control strategies or adjust execution logic to ensure the security and consistency of data access.

[0143] The aforementioned operational logic conflict refers to a logical contradiction arising during the execution of multiple candidate skills due to inconsistencies in rules, state machine transitions, or causal dependencies. Operational logic conflicts manifest as one skill expecting the system to be in one state, while another skill forces the system to transition to another state, causing subsequent operations to fail or produce incorrect results. Operational logic conflicts do not involve resource contention or permission issues, but rather the coherence of logic. Identifying operational logic conflicts requires analyzing the skill's operational sequence and context. By detecting operational logic conflicts, the skill execution order can be improved or skills that do not conform to the current logic can be eliminated, ensuring the correctness of the task flow and preventing interruptions or a decline in user experience due to logical breaks.

[0144] The aforementioned call frequency refers to a statistical indicator reflecting the frequency with which candidate skills are called by users or the system during historical operations. A higher call frequency generally indicates that the skill is more frequently used, more reliable, or more in line with user habits. Call frequency is used to evaluate skill value and popularity. Interaction processing systems tend to retain skills with high call frequencies because they have been proven effective and popular with users in past practices. By filtering based on call frequency, interaction processing systems can prioritize retaining high-value skills and eliminate low-frequency or unpopular conflicting skills, thereby improving the overall utility of skill combinations and enhancing the practicality of agent responses and user satisfaction.

[0145] The aforementioned second candidate skill refers to a candidate skill that is determined to have a low call frequency and should be deleted in read / write operation conflicts or operational logic conflicts. A second candidate skill conflicts with another candidate skill, and its call frequency is lower than the other candidate skill. Based on the call frequency, the interaction processing system determines that the second candidate skill is at a disadvantage in terms of value and retains it. A second candidate skill may lead to resource waste or reduce overall execution efficiency. Deleting a second candidate skill is to eliminate logical or data-level conflicts, ensuring that the skills in the final target skill set have high usability and reliability, thereby improving the success rate of agent task execution and user experience.

[0146] In one optional embodiment, the interaction processing system first determines that the conflict type is an execution permission conflict. The system extracts the permission levels of any two candidate skills and compares their levels. Then, the system identifies the skill with the lower permission level as the first candidate skill. The system removes the first candidate skill from the set of multiple candidate skills, retaining the other candidate skill with the higher permission level. Next, the system checks whether there are any other permission conflicts after the removal; if not, the remaining set of conflict-free skills is identified as the target skill.

[0147] This application embodiment distinguishes conflict types and adopts differentiated filtering strategies. For permission conflicts, permission level is prioritized, and for read / write and logic conflicts, call frequency is prioritized. Low-value or high-risk skills are removed, thereby improving the security, reliability, and utility of the target skill set and ensuring the efficient execution of intelligent agent tasks.

[0148] Conventional deduplication methods in related technologies target literal information such as skill identifiers (IDs) and names to remove duplicate entries. However, they fail to identify skills with overlapping semantics or mutually exclusive execution, easily leading to multi-skill conflicts and repeated execution, thus reducing task efficiency and accuracy. Furthermore, conflict handling in these technologies is often placed during the skill execution phase or accurate matching stage, requiring a large number of candidate skills to be fed into a large model for prompt assembly and semantic reasoning. Invalid skills consume core computing power and GPU memory resources, resulting in wasted computing power and significantly increasing interaction response latency, impacting user experience.

[0149] The existing technologies rely on large models for semantic deduplication and conflict determination, which suffer from problems such as model illusion, unstable inference results, and untraceable black-box decision-making. These technologies exhibit poor controllability and stability, and high maintenance costs. Furthermore, the deduplication and conflict determination logic is disconnected from the front-end retrieval process, failing to fully reuse retrieval data and incurring additional redundant computational overhead, making them unsuitable for the high-efficiency operation requirements of scenarios with massive skill volumes.

[0150] Currently, there is no lightweight pre-processing solution in related technologies that embeds skill conflict resolution and semantic deduplication logic into the hybrid retrieval of Elasticsearch (ES) before accurate matching. Furthermore, there is no purely lightweight conflict arbitration and deduplication solution for retrieving candidate sets without relying on large models, making it difficult to solve problems such as wasted computing power, slow response, and misjudgment of conflicts in scenarios with massive skill sets for intelligent agents.

[0151] The relevant technologies cannot meet the requirements for efficient and stable operation in scenarios with a large number of skills. Furthermore, deduplication and conflict verification are placed after the large model is accurately matched. All candidate skills recalled by ES need to be sent into the large model to complete the whole set of inference. A large number of redundant conflicting skills are directly eliminated after participating in high-computation inference, resulting in ineffective computing power investment and increasing the overall response latency.

[0152] In addition, the reliance on large models to complete semantic deduplication and conflict determination in related technologies has problems such as model illusion, random fluctuations in inference results, and untraceable black boxes of decision-making. Furthermore, the deployment and operation costs of large models are high, and the Prompt needs to be improved or the model needs to be fine-tuned in sync after the skill library is updated, resulting in high maintenance costs.

[0153] The hybrid ES retrieval process in related technologies is independent of the subsequent deduplication and conflict handling process. It does not reuse core data such as retrieval scores and requires additional recalculation. Furthermore, it does not perform lightweight pre-processing on the retrieval candidate set, making it impossible to intercept invalid skills in advance. The overall process is redundant and difficult to adapt to dynamic skill expansion and high-concurrency scenarios.

[0154] To address the shortcomings of related technologies in maintaining massive skill sets for intelligent agents, such as high computational cost in post-processing of skill conflicts, poor coarse-grained deduplication, insufficient stability of deduplication involving large models, and high response latency, this application provides a hybrid retrieval-based method for pre-emptive resolution of skill conflicts and deduplication filtering. By embedding conflict resolution and semantic deduplication between retrieval and matching, and employing lightweight computational logic without large models, this method achieves early interception of invalid skills and precise removal of redundant conflicting skills, reducing subsequent computational cost from the source and improving the interactive response speed and operational stability of the intelligent agent.

[0155] This application discloses a method for pre-emptive resolution and deduplication filtering of intelligent agent skill conflicts based on hybrid retrieval. The method includes: standardized entry of intelligent agent skills into the database and construction of hybrid indexes in Elasticsearch (ES), preprocessing of user requests and recall in hybrid ES retrieval, pre-weighting of candidate skills and low-score screening, lightweight vector similarity semantic deduplication, conflict arbitration resolution driven by preset rules, and distribution and closed-loop execution of a streamlined skill pool.

[0156] In the stage of standardizing and storing agent skills and building a hybrid Elasticsearch (ES) index, the massive number of skills maintained by the agent are standardized and sorted out, extracting attribute information such as identifier, skill name, core function description, execution permission, priority, and mutually exclusive skills for each skill. An ES hybrid index is built based on these skill attributes, with fields such as skill name and execution permission set as text retrieval fields. The core function description field is converted into a fixed-dimensional static function vector using a pre-trained semantic embedding model and stored in the ES index as a vector retrieval field. Simultaneously, priority and mutually exclusive skill pairs are stored in the index, achieving integrated storage of skill text information, function vectors, and identifiers. Furthermore, the completed ES index supports incremental skill updates; adding, modifying, or deleting a single skill does not require rebuilding the index, ensuring index real-time performance and retrieval efficiency.

[0157] In the user request preprocessing and ES hybrid retrieval phase, after receiving the user's interaction request text, the agent cleans and normalizes the text, removing redundant symbols and stop words, and extracting the core query statement. The normalized query statement is then synchronously converted into a target query vector of the same dimension as the skill function vector. The standard query statement and the query vector are input into the ES engine together, triggering parallel hybrid retrieval. One path performs text retrieval based on BM25 to obtain text matching scores. The other path performs vector semantic retrieval based on the KNN algorithm to obtain vector similarity scores. Finally, a predetermined number of highly relevant candidate skills are retrieved, and the text matching score, vector similarity score, priority, and mutual exclusion for each candidate skill are obtained.

[0158] In the candidate skill pre-weighting and low-score screening stage, the text matching score and vector matching score returned by the Elasticsearch hybrid retrieval are directly reused and weighted and summed according to a preset fixed ratio to obtain the basic comprehensive score of each candidate skill, without the need for additional model inference or vector calculation. At the same time, a preset low-score threshold is set to eliminate marginal invalid candidate skills with a basic comprehensive score below the preset threshold, narrowing the scope of subsequent processing, further reducing the amount of computation, and retaining high-scoring candidate skills to form a candidate set to be processed.

[0159] In the lightweight vector similarity semantic deduplication stage, for the candidate set to be processed, the static function vectors of each skill pre-stored in the ES index are extracted, and a cosine similarity algorithm is used for pairwise comparison. No large model is involved throughout the process; only pure numerical calculations are performed. A preset semantic overlap threshold is set, and skills with a cosine similarity exceeding the threshold are judged as semantically redundant and functionally overlapping skills. For overlapping skills, skills with higher basic comprehensive scores are retained, while redundant skills with lower scores are directly removed, completing semantic deduplication and preventing functionally similar skills from repeatedly entering subsequent stages.

[0160] In the conflict arbitration and resolution phase driven by pre-defined rules, a logically pre-defined skill mutual exclusion rule base is used. The rules cover three categories of deterministic rules: mutual exclusion of operations with the same function, mutual exclusion of read and write operations, and mutual exclusion of execution permissions. Mutual exclusion relationships are stored in a static configuration table. The deduplicated candidate skill set is scanned, and mutual exclusion pairs are identified by combining them with mutual exclusion identifiers stored in the Elasticsearch index. The conflict arbitration mechanism is triggered, retaining higher-priority skills and removing other mutually exclusive skills according to a predetermined order of priority, execution security, and historical call frequency, thus completing the pre-conflict resolution. Finally, a streamlined candidate skill pool without duplicates or conflicts is generated.

[0161] In the streamlined skill pool distribution and closed-loop execution phase, the streamlined candidate skill pool, after pre-processing deduplication and conflict resolution, is distributed to the agent matching and execution module. Only a small number of skills within the streamlined pool undergo subsequent prompt assembly and large-scale model matching. Compared to related technologies, this significantly reduces the number of skills entering the high-computation-power stage, completely avoiding the computational waste caused by ineffective skills and improving overall response speed. Simultaneously, the deduplication and conflict resolution results are recorded to improve semantic overlap thresholds and mutual exclusion rules, ensuring long-term operational effectiveness.

[0162] This application's embodiments prioritize conflict resolution and deduplication, intercepting invalid and redundant skills in advance, preventing a large number of candidate skills from entering the large model matching stage, significantly reducing invalid computing power consumption, and only occupying the computing power of the basic processor (Central Processing Unit, CPU), without additional video memory overhead.

[0163] The embodiments of this application have strong operational stability, do not rely on large models, adopt pure numerical calculation and deterministic hard rules, have no model illusion and no inference fluctuations, and the results are traceable and controllable.

[0164] The embodiments of this application improve response speed, reduce computation time to milliseconds, significantly reduce the number of candidate skills, shorten subsequent prompt assembly and model inference time, and greatly reduce overall interaction response latency.

[0165] The embodiments of this application provide accurate deduplication, taking into account both literal repetition and semantic overlap, while also covering mutual exclusion and conflict arbitration. Compared with coarse-grained deduplication, the accuracy is improved, effectively avoiding skill conflicts and duplicate execution.

[0166] The embodiments of this application have low maintenance costs, requiring only the maintenance of threshold parameters and mutual exclusion rule tables. Skill updates do not require fine-tuning of the model or rebuilding of the index, adapting to the needs of dynamic expansion of massive skills.

[0167] In the agent skill standardization and Elasticsearch (ES) hybrid index construction phase, hundreds of general skills were standardized in advance, and six attributes were extracted for each skill: unique ID, skill name, core function description, priority, and mutually exclusive skills. Fields were then defined for each attribute: the skill name was set as the ES text retrieval field, the core function description was converted into a static function vector as the vector retrieval field, and priority and mutually exclusive skills were set as additional attribute fields. Subsequently, an ES hybrid index library was built, and the incremental index synchronization function was enabled. Modification, addition, or obsolescence of individual skill information only requires updating the corresponding index entry synchronously, without requiring a full index rebuild. In this embodiment, semantically overlapping skills and mutually exclusive skill pairs are clearly defined.

[0168] During the user request reception and preprocessing phase, the agent receives interaction requests initiated by employees. For example, if a user inputs "Help me check the company's order data for this month," the system first preprocesses the request text, removing modal particles and redundant spaces, and standardizes it into the standard query statement "Query the company's order data for this month." Then, it uses a semantic vector model to convert this statement into a query vector, maintaining consistency with the vector dimension of the skill library, thus completing the data preparation for the query end.

[0169] In the Elasticsearch (ES) hybrid retrieval and candidate skill recall phase, standard query statements and query vectors are simultaneously input into the ES cluster, triggering parallel hybrid retrieval. One path performs accurate matching of text fields based on algorithms, while the other path calculates vector semantic similarity using the KNN approximate nearest neighbor algorithm. Both retrieval paths are executed synchronously, with the time consumption controlled within milliseconds. Finally, highly relevant candidate skills are recalled, and the text matching score, vector similarity score, priority, and mutual exclusion of each candidate skill are returned, completing the initial coarse screening.

[0170] In the candidate skill pre-weighting and low-score elimination stage, the two scores returned by ES are directly reused, and the text score and vector score are weighted and summed according to a fixed ratio to calculate the basic comprehensive score of each candidate skill. In addition, a low-score elimination threshold is set to eliminate a number of marginal irrelevant skills with scores below the threshold, and high-scoring skills are retained to form a candidate set to be processed, which further narrows the scope of subsequent processing and reduces the amount of computation.

[0171] In the lightweight vector similarity semantic deduplication stage, for the remaining skills to be processed, the pre-stored static function vectors in the ES index are extracted, and pairwise comparisons are performed using the cosine similarity algorithm. A preset semantic overlap threshold is set, and several groups of semantically overlapping skills are identified through calculation. Overlapping skills are those with similar functions but different expressions. Following the principle of retaining high-scoring skills, low-scoring redundant skills are eliminated, and candidate skills without semantic overlap are retained. The entire process involves only pure numerical calculations, without the participation of a large model.

[0172] In the pre-defined rule conflict arbitration and simplification pool generation stage, based on the pre-defined mutual exclusion rule base, the remaining skills are scanned, and several pairs of mutually exclusive skills are identified. According to the arbitration order of priority, execution security, and historical call frequency, low-priority mutually exclusive skills are eliminated. Finally, several simplified candidate skill pools with no duplication, no conflicts, and high adaptability are generated, completing the pre-filtering process.

[0173] During the execution and performance comparison phase, the streamlined skills are distributed to the agent matching module. Only a small number of skills need to be assembled into prompts and matched against a large model. Ultimately, a small number of superior skills are selected to execute the task. Compared to related technical solutions where all skills are sent to the high-computational-power inference stage, the embodiment of this application significantly reduces invalid computational power consumption, and there are no skill conflicts or duplicate executions, resulting in improved operational stability and user experience.

[0174] Figure 2 This is a schematic diagram of an optional interactive processing method according to an embodiment of this application, such as... Figure 2 As shown, the process includes the following steps: skill entry into the database, user request processing, elastic search engine retrieval, pre-weighted filtering, deduplication, rule conflict arbitration, and deployment of the execution model.

[0175] First, skill data is standardized and stored in a database, and a hybrid index is built to form the foundation for retrieval. Next, user request processing cleanses and vectorizes the query statements to extract user intent. An elastic search engine then retrieves a set of highly relevant candidate skills in parallel to initially identify potential matches. Then, pre-weighted filtering eliminates low-scoring, marginally invalid skills to significantly narrow down the scope of subsequent processing. Simultaneously, deduplication is performed using pure numerical calculations to remove semantically overlapping redundant skills, improving the purity of the candidate set. Rule conflict arbitration eliminates operational conflicts based on pre-defined mutual exclusion rules to ensure execution security. Finally, the execution model submits the streamlined, conflict-free skill pool to the main model, reducing computational consumption and improving overall response speed.

[0176] Figure 3 This is a schematic diagram of an optional interactive processing method according to an embodiment of this application, such as... Figure 3 As shown, the process includes the following steps: generating static vectors of skills, calculating cosine similarity, determining if the similarity exceeds a preset threshold, removing skills, or classifying them as mutually exclusive skills.

[0177] First, static vector generation of skills constructs functional semantic representations of skills to support subsequent comparisons. Then, cosine similarity calculation quantifies the semantic closeness between skills to provide numerical evidence. Next, a threshold value greater than the cosine similarity is used to identify skill pairs with highly overlapping functions, thus locking in redundant targets. If the cosine similarity is less than or equal to the threshold, the skill is removed. If the cosine similarity is greater than the threshold, the skills are determined to be mutually exclusive.

[0178] Figure 4 This is a schematic diagram of an optional interactive processing method according to an embodiment of this application, such as... Figure 4 As shown, the method includes the following steps:

[0179] Step S402: In response to receiving the interaction request text from the target client, the interaction request text is transformed based on a preset vector dimension to obtain the target query vector.

[0180] Among them, the preset vector dimension is used to represent the vector dimension of the function vectors corresponding to multiple preset skills.

[0181] Step S404: Based on the interactive request text and the target query vector, multiple preset skills are retrieved to obtain multiple candidate skills.

[0182] Step S406: Based on the functional implementation logic of multiple candidate skills, perform conflict testing on multiple candidate skills to obtain conflict test results.

[0183] Among them, the conflict test results are used to characterize whether there is a conflict in the functional implementation logic of multiple candidate skills.

[0184] Step S408: Based on the conflict test results, multiple candidate skills are screened to obtain the target skill.

[0185] Step S410: Send the target skill and interaction request text to the execution model. The execution model processes the interaction request text according to the target skill to obtain the processing result of the interaction request text.

[0186] Figure 5 This is a schematic diagram of an optional interactive processing method according to an embodiment of this application, such as... Figure 5 As shown, the method includes the following steps:

[0187] Step S502: In response to receiving the interaction request text from the target client, the interaction request text is transformed based on a preset vector dimension to obtain the target query vector.

[0188] Step S504: Perform text matching on the interactive request text and the text description information of multiple preset skills to obtain the text matching results of multiple preset skills; perform vector matching on the target query vector and the function vectors of multiple preset skills to obtain the vector matching results of multiple preset skills.

[0189] Step S506: Based on the text matching results and vector matching results of multiple preset skills, the multiple preset skills are filtered to obtain multiple candidate skills.

[0190] Step S508: Based on the functional implementation logic of multiple candidate skills, perform conflict testing on multiple candidate skills to obtain conflict test results.

[0191] Step S510: Based on the conflict test results, multiple candidate skills are screened to obtain the target skill.

[0192] Step S512: Send the target skill and interaction request text to the execution model. The execution model processes the interaction request text according to the target skill to obtain the processing result of the interaction request text.

[0193] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0194] According to an embodiment of this application, an embodiment of an interactive processing apparatus is provided. Figure 6 This is a schematic diagram of an interactive processing device according to an embodiment of this application, such as... Figure 6 As shown, it should be noted that this device can be used to execute the above-described interactive processing method. The device includes: a retrieval module 602, a testing module 604, a filtering module 606, and a processing module 608.

[0195] The retrieval module 602 is used to respond to the received interaction request text from the target client, and retrieve multiple preset skills stored in the skill library based on the interaction request text to obtain multiple candidate skills; the testing module 604 is used to perform conflict testing on the multiple candidate skills based on the functional implementation logic of the multiple candidate skills to obtain conflict test results, wherein the conflict test results are used to characterize whether there is a conflict in the functional implementation logic of the multiple candidate skills; the filtering module 606 is used to filter the multiple candidate skills based on the conflict test results to obtain the target skill; the processing module 608 is used to send the target skill and the interaction request text to the execution model, and use the execution model to process the interaction request text according to the target skill to obtain the processing result of the interaction request text.

[0196] The retrieval module includes a conversion unit and a retrieval unit. The retrieval unit includes a matching subunit and a filtering subunit. The matching subunit includes a first matching subunit and a second matching subunit.

[0197] The filtering subunit includes a first determining subunit, a sorting subunit, and a second determining subunit. The first determining subunit includes a first score determining subunit, a second score determining subunit, and a weighted summation subunit. The second determining subunit includes a third determining subunit, a matching subunit, and a deduplication subunit.

[0198] The conversion unit is used to convert the interactive request text based on the preset vector dimension to obtain the target query vector, wherein the preset vector dimension is used to represent the vector dimension of the function vectors corresponding to multiple preset skills; the retrieval unit is used to retrieve multiple preset skills based on the interactive request text and the target query vector to obtain multiple candidate skills.

[0199] The matching subunit is used to match the interactive request text, the target query vector, and multiple preset skills to obtain the text matching results and vector matching results of multiple preset skills. The filtering subunit is used to filter multiple preset skills based on the text matching results and vector matching results of multiple preset skills to obtain multiple candidate skills.

[0200] The first matching subunit is used to perform text matching between the interactive request text and the text description information of multiple preset skills to obtain the text matching results of multiple preset skills; the second matching subunit is used to perform vector matching between the target query vector and the function vectors of multiple preset skills to obtain the vector matching results of multiple preset skills.

[0201] The first determining subunit is used to determine the skill evaluation scores of multiple preset skills based on the text matching results and vector matching results of multiple preset skills; the ranking subunit is used to rank multiple candidate skills based on the skill evaluation scores of multiple preset skills to obtain the ranking results; and the second determining subunit is used to filter multiple preset skills based on the ranking results to obtain multiple candidate skills.

[0202] The first score determination subunit is used to determine the text matching score of multiple preset skills based on the text matching results of multiple preset skills; the second score determination subunit is used to determine the vector matching score of multiple preset skills based on the vector matching results of multiple preset skills; and the weighted summation subunit is used to perform weighted summation of the text matching score and the vector matching score based on a preset weighting ratio to obtain the skill evaluation score of multiple preset skills.

[0203] The third determining subunit is used to filter multiple preset skills based on the sorting results to obtain multiple initial skills; the matching subunit is used to perform similarity matching between any two initial skills among the multiple initial skills to obtain similarity matching results, wherein the similarity matching results are used to indicate whether the similarity between any two initial skills is greater than a preset threshold; the deduplication subunit is used to deduplicate multiple initial skills based on the similarity matching results to obtain multiple candidate skills.

[0204] The third determining subunit is also used to determine that the preset skills that are ranked before the preset position in the sorting result are multiple initial skills; or, to extract a preset number of preset skills as multiple initial skills based on the sorting result.

[0205] The deduplication subunit is further used to divide multiple initial skills based on similarity matching results, obtaining at least one first initial skill and multiple second initial skills. The multiple second initial skills are associated initial skills with a similarity greater than a preset threshold, and the at least one first initial skill is other skills among the multiple initial skills besides the multiple second initial skills. The multiple second initial skills are deduplicated to obtain the deduplicated target skill. The deduplicated target skill and at least one first initial skill are merged to obtain multiple candidate skills.

[0206] The deduplication subunit is also used to divide multiple second initial skills into multiple skill combinations; for any skill combination in the multiple skill combinations, the initial skills in the initial skills of ...

[0207] The testing module is also used to perform compatibility comparisons on multiple candidate skills based on their functional implementation logic, and obtain comparison results. The comparison results indicate whether any of the following exist: mutually exclusive operation types, data read / write conflicts, and permission overriding. If the comparison result indicates that multiple candidate skills have any of these issues, the conflict test result determines that the functional implementation logic of the multiple candidate skills is conflicting. If the comparison result indicates that multiple candidate skills do not have any of these issues, the conflict test result determines that the functional implementation logic of the multiple candidate skills is not conflicting.

[0208] The testing module is also used to parse the functional implementation logic of multiple candidate skills to obtain the operation type, data access mode, and skill execution permission of multiple candidate skills; to perform structured processing on the operation type, data access mode, and skill execution permission to obtain logical constraints; and to perform compatibility comparison on multiple candidate skills based on the logical constraints to obtain comparison results.

[0209] The filtering module is further used to determine multiple candidate skills as target skills in response to the conflict test result indicating that there is no conflict in the functional implementation logic of multiple candidate skills; and to determine any two candidate skills among the multiple candidate skills whose functional implementation logic is conflicting in response to the conflict test result indicating that there is conflict in the functional implementation logic of multiple candidate skills, and to filter the multiple candidate skills based on the conflict type of any two candidate skills to obtain the target skill.

[0210] The filtering module is further configured to, in response to a conflict type of execution permission conflict, delete the first candidate skill from any two candidate skills based on the permission levels of any two candidate skills to obtain the target skill, wherein the permission level of the first candidate skill is lower than that of the other candidate skill; and in response to a conflict type of read / write operation conflict and operation logic conflict, delete the second candidate skill from any two candidate skills based on the call frequency of any two candidate skills to obtain the target skill, wherein the call frequency of the second candidate skill is lower than that of the other candidate skill.

[0211] Embodiments of this application also provide a vehicle, including: a memory storing an executable program; and a processor for running the program, wherein the program executes the methods described in various embodiments of this application when it runs.

[0212] Embodiments of this application also provide a computer-readable storage medium including a stored executable program, wherein, when the executable program is running, it controls the device where the computer-readable storage medium is located to perform the methods of various embodiments of this application.

[0213] Embodiments of this application also provide a computer program product, including a computer program that, when executed by a processor, implements the methods of various embodiments of this application.

[0214] Embodiments of this application also provide a computer program product, including a non-volatile computer-readable storage medium for storing a computer program that, when executed by a processor, implements the methods in various embodiments of this application.

[0215] Embodiments of this application also provide a computer program that, when executed by a processor, implements the methods described in the various embodiments of this application.

[0216] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0217] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0218] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0219] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0220] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0221] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. An interactive processing method, characterized in that, include: In response to receiving an interaction request text from the target client, the system retrieves multiple preset skills stored in the skill library based on the interaction request text to obtain multiple candidate skills; Based on the functional implementation logic of the multiple candidate skills, conflict tests are performed on the multiple candidate skills to obtain conflict test results, wherein the conflict test results are used to characterize whether there is a conflict in the functional implementation logic of the multiple candidate skills; Based on the conflict test results, the multiple candidate skills are screened to obtain the target skill; The target skill and the interaction request text are sent to the execution model, which then processes the interaction request text according to the target skill to obtain the processing result of the interaction request text.

2. The method according to claim 1, characterized in that, Based on the interactive request text, multiple preset skills stored in the skill library are retrieved to obtain multiple candidate skills, including: The interactive request text is transformed based on a preset vector dimension to obtain a target query vector, wherein the preset vector dimension is used to represent the vector dimension of the function vectors corresponding to the multiple preset skills; Based on the interaction request text and the target query vector, the multiple preset skills are retrieved to obtain the multiple candidate skills.

3. The method according to claim 2, characterized in that, Based on the interaction request text and the target query vector, the plurality of preset skills are retrieved to obtain the plurality of candidate skills, including: The interaction request text, the target query vector, and the multiple preset skills are matched to obtain the text matching results and the vector matching results of the multiple preset skills. Based on the text matching results and vector matching results of the multiple preset skills, the multiple preset skills are filtered to obtain the multiple candidate skills.

4. The method according to claim 3, characterized in that, The interaction request text, the target query vector, and the multiple preset skills are matched to obtain text matching results and vector matching results for the multiple preset skills, including: The interactive request text and the text description information of the multiple preset skills are matched to obtain the text matching results of the multiple preset skills; Vector matching is performed on the target query vector and the function vectors of the multiple preset skills to obtain the vector matching results of the multiple preset skills.

5. The method according to claim 3, characterized in that, Based on the text matching results and vector matching results of the multiple preset skills, the multiple preset skills are filtered to obtain the multiple candidate skills, including: Based on the text matching results and vector matching results of the multiple preset skills, the skill evaluation scores of the multiple preset skills are determined; Based on the skill evaluation scores of the multiple preset skills, the multiple candidate skills are ranked to obtain the ranking result; Based on the ranking results, the multiple preset skills are filtered to obtain the multiple candidate skills.

6. The method according to claim 5, characterized in that, Based on the text matching results and vector matching results of the multiple preset skills, the skill evaluation scores of the multiple preset skills are determined, including: Based on the text matching results of the multiple preset skills, determine the text matching score of the multiple preset skills; Based on the vector matching results of the multiple preset skills, determine the vector matching score of the multiple preset skills; The text matching score and the vector matching score are weighted and summed based on a preset weighting ratio to obtain the skill evaluation score of the multiple preset skills.

7. The method according to claim 5, characterized in that, Based on the ranking results, the multiple preset skills are filtered to obtain the multiple candidate skills, including: Based on the sorting results, the multiple preset skills are filtered to obtain multiple initial skills; A similarity matching is performed between any two initial skills from the plurality of initial skills to obtain a similarity matching result, wherein the similarity matching result is used to indicate whether the similarity between any two initial skills is greater than a preset threshold; Based on the similarity matching results, the multiple initial skills are deduplicated to obtain the multiple candidate skills.

8. The method according to claim 7, characterized in that, Based on the ranking results, the multiple preset skills are filtered to obtain multiple initial skills, including: The preset skills that are ranked before a preset position in the sorting results are identified as the plurality of initial skills; Alternatively, a preset number of preset skills may be extracted based on the sorting results as the plurality of initial skills.

9. The method according to claim 7, characterized in that, Based on the similarity matching results, the multiple initial skills are deduplicated to obtain the multiple candidate skills, including: Based on the similarity matching results, the multiple initial skills are divided to obtain at least one first initial skill and multiple second initial skills. The multiple second initial skills are associated initial skills with a similarity greater than the preset threshold, and the at least one first initial skill is other skills among the multiple initial skills besides the multiple second initial skills. The multiple second initial skills are deduplicated to obtain the deduplicated target skills; The deduplicated target skill and the at least one first initial skill are merged to obtain the plurality of candidate skills.

10. The method according to claim 9, characterized in that, The multiple second initial skills are deduplicated to obtain the deduplicated target skills, including: The multiple second initial skills are divided into multiple skill combinations; For any one of the multiple skill combinations, the initial skills in the initial skill combination are deduplicated based on the skill evaluation score of the initial skills in the initial skill combination to obtain the deduplicated skills corresponding to the initial skill combination. Based on the deduplicated skills corresponding to the multiple skill combinations, the deduplicated target skill is determined.

11. The method according to any one of claims 1 to 10, characterized in that, Based on the functional implementation logic of the multiple candidate skills, conflict testing is performed on the multiple candidate skills to obtain conflict test results, including: Based on the functional implementation logic of the multiple candidate skills, a compatibility comparison is performed on the multiple candidate skills to obtain a comparison result. The comparison result is used to indicate whether any one of the following exists: mutual exclusion of operation types, data read / write conflict, and permission overriding. In response to the comparison result indicating that any one of the following exists among the multiple candidate skills: mutual exclusion of operation types, conflict of data read / write, and overriding of permissions, the conflict test result is determined to indicate that the functional implementation logic of the multiple candidate skills is conflicted. In response to the comparison result indicating that none of the following exist among the multiple candidate skills: mutual exclusion of operation types, conflict of data read / write, and overriding of permissions, the conflict test result is determined to indicate that the functional implementation logic of the multiple candidate skills does not conflict.

12. The method according to claim 11, characterized in that, Based on the functional implementation logic of the multiple candidate skills, a compatibility comparison is performed on the multiple candidate skills to obtain the comparison results, including: The functional implementation logic of the multiple candidate skills is analyzed to obtain the operation type, data access mode, and skill execution permission of the multiple candidate skills. The operation type, the data access mode, and the skill execution permission are structured to obtain logical constraints. The compatibility comparison of the multiple candidate skills is performed based on the logical constraints to obtain the comparison results.

13. The method according to any one of claims 1 to 10, characterized in that, Based on the conflict test results, the multiple candidate skills are screened to obtain target skills, including: In response to the conflict test result indicating that there is no conflict in the functional implementation logic of the plurality of candidate skills, the plurality of candidate skills are determined as the target skill; In response to the conflict test result indicating that the functional implementation logic of the plurality of candidate skills is conflicted, any two candidate skills among the plurality of candidate skills whose functional implementation logic is conflicted are identified, and the plurality of candidate skills are filtered based on the conflict type of the any two candidate skills to obtain the target skill.

14. The method according to claim 13, characterized in that, The conflict types include execution permission conflicts, read / write operation conflicts, and operation logic conflicts. Based on the conflict types of any two candidate skills, the multiple candidate skills are filtered to obtain the target skill, including: In response to the conflict type being the execution permission conflict, the first candidate skill among the two candidate skills is deleted based on the permission levels of the two candidate skills to obtain the target skill, wherein the permission level of the first candidate skill is lower than that of the other candidate skill among the two candidate skills. In response to the conflict type being the read-write operation conflict and the operation logic conflict, the second candidate skill among the two candidate skills is deleted based on the calling frequency of the two candidate skills to obtain the target skill, wherein the calling frequency of the second candidate skill is less than that of the other candidate skill among the two candidate skills.

15. An interactive processing device, characterized in that, include: The retrieval module is used to respond to the received interactive request text from the target client, and retrieve multiple preset skills stored in the skill library based on the interactive request text to obtain multiple candidate skills; The testing module is used to perform conflict testing on the multiple candidate skills based on the functional implementation logic of the multiple candidate skills, and obtain conflict test results, wherein the conflict test results are used to characterize whether there is a conflict in the functional implementation logic of the multiple candidate skills; A filtering module is used to filter the multiple candidate skills based on the conflict test results to obtain the target skill; The processing module is used to send the target skill and the interaction request text to the execution model, and use the execution model to process the interaction request text according to the target skill to obtain the processing result of the interaction request text.

16. The apparatus according to claim 15, characterized in that, The search module includes: A conversion unit is used to convert the interactive request text based on a preset vector dimension to obtain a target query vector, wherein the preset vector dimension is used to represent the vector dimension of the function vectors corresponding to the plurality of preset skills; The retrieval unit is used to retrieve the plurality of preset skills based on the interaction request text and the target query vector to obtain the plurality of candidate skills.

17. The apparatus according to claim 16, characterized in that, The retrieval unit includes: The matching subunit is used to match the interactive request text, the target query vector, and the multiple preset skills to obtain the text matching results and the vector matching results of the multiple preset skills. The filtering subunit is used to filter the multiple preset skills based on the text matching results and the vector matching results of the multiple preset skills to obtain the multiple candidate skills.

18. The apparatus according to claim 17, characterized in that, Matching subunits include: The first matching subunit is used to perform text matching on the interactive request text and the text description information of the multiple preset skills to obtain the text matching results of the multiple preset skills; The second matching subunit is used to perform vector matching on the target query vector and the function vectors of the multiple preset skills to obtain the vector matching result of the multiple preset skills.

19. The apparatus according to claim 17, characterized in that, The filtering sub-unit includes: The first determining subunit determines the skill evaluation score of the multiple preset skills based on the text matching results and the vector matching results of the multiple preset skills. The sorting subunit is used to sort the multiple candidate skills based on the skill evaluation scores of the multiple preset skills, and obtain the sorting result; The second determining subunit is used to filter the plurality of preset skills based on the sorting results to obtain the plurality of candidate skills.

20. The apparatus according to claim 19, characterized in that, The first defined subunit includes: The first score determination subunit is used to determine the text matching score of the multiple preset skills based on the text matching results of the multiple preset skills; The second score determination subunit is used to determine the vector matching score of the multiple preset skills based on the vector matching results of the multiple preset skills; The weighted summation subunit is used to perform a weighted summation of the text matching score and the vector matching score based on a preset weighting ratio to obtain the skill evaluation scores of the multiple preset skills.

21. The apparatus according to claim 19, characterized in that, The second defined subunit includes: The third determining subunit is used to filter the multiple preset skills based on the sorting results to obtain multiple initial skills; A matching subunit is used to perform similarity matching between any two initial skills among the plurality of initial skills to obtain a similarity matching result, wherein the similarity matching result is used to indicate whether the similarity between any two initial skills is greater than a preset threshold; The deduplication subunit is used to deduplicatize the multiple initial skills based on the similarity matching results to obtain the multiple candidate skills.

22. A vehicle, characterized in that, include: Memory, which stores executable programs; A processor for running the program, wherein the program, when running, performs the method according to any one of claims 1 to 14.

23. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored executable program, wherein, when the executable program is executed, it controls the device on which the storage medium is located to perform the method according to any one of claims 1 to 14.