Information processing method and information processing apparatus

Abstract

Embodiments of the present specification provide an information processing method and an information processing device, wherein the information processing method comprises: obtaining a to-be-processed intention and at least one reference skill corresponding to the to-be-processed intention, and determining an initial execution link corresponding to the to-be-processed intention according to the at least one reference skill; generating a reference execution result corresponding to the to-be-processed intention according to the initial execution link, and recording link attribute information corresponding to the initial execution link; in a case where the reference execution result does not satisfy a preset verification condition, determining a target failure type according to the link attribute information; adjusting the initial execution link according to the target failure type, generating a target execution link, and generating a target execution result corresponding to the to-be-processed intention according to the target execution link. According to the present specification, the accuracy of skill maintenance is improved.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of artificial intelligence technology, and in particular to an information processing method and an information processing device. Background Technology

[0002] With the continuous development of artificial intelligence technology, people are demanding higher and higher accuracy from the intelligent processing unit's processing links.

[0003] In multi-skill intelligent processing unit maintenance scenarios, it is usually necessary to manually identify and optimize the problems of each skill. However, this approach is time-consuming, costly, and lacks traceability. Furthermore, when product API calls change, skills need to be continuously synchronized and adapted; otherwise, routing, retrieval, and responses will become invalid.

[0004] Therefore, how to efficiently and accurately verify skills has become an urgent problem to be solved. Summary of the Invention

[0005] In view of the above, embodiments of this specification provide an information processing method. One or more embodiments of this specification also relate to an information processing apparatus, a computing device, a computer-readable storage medium, and a computer program product, to address the technical deficiencies existing in the prior art.

[0006] According to a first aspect of the embodiments of this specification, an information processing method is provided, comprising: Obtain the intent to be processed and at least one reference skill corresponding to the intent to be processed, and determine the initial execution link corresponding to the intent to be processed based on at least one reference skill; Generate a reference execution result corresponding to the intent to be processed based on the initial execution link, and record the link attribute information corresponding to the initial execution link; If the reference execution result does not meet the preset verification conditions, the target failure type is determined based on the link attribute information; Based on the target failure type, the initial execution chain is adjusted to generate the target execution chain, and the target execution result corresponding to the intent to be processed is generated based on the target execution chain.

[0007] According to a second aspect of the embodiments of this specification, an information processing apparatus is provided, comprising: The acquisition unit is configured to acquire the intent to be processed and at least one reference skill corresponding to the intent to be processed, and determine the initial execution link corresponding to the intent to be processed based on the at least one reference skill. The recording unit is configured to generate a reference execution result corresponding to the intent to be processed based on the initial execution link, and to record the link attribute information corresponding to the initial execution link; The determining unit is configured to determine the target failure type based on the link attribute information when the reference execution result does not meet the preset verification conditions. The generation unit is configured to adjust the initial execution link according to the target failure type, generate the target execution link, and generate the target execution result corresponding to the intent to be processed based on the target execution link.

[0008] According to a third aspect of the embodiments of this specification, an information processing method is provided, applied to a cloud-side device, comprising: The receiving end-side device sends a pending request, parses the pending request, obtains the pending intent and at least one reference skill corresponding to the pending intent, and determines the initial execution link corresponding to the pending intent based on at least one reference skill. Generate a reference execution result corresponding to the intent to be processed based on the initial execution link, and record the link attribute information corresponding to the initial execution link; If the reference execution result does not meet the preset verification conditions, the target failure type is determined based on the link attribute information; Based on the target failure type, the initial execution chain is adjusted to generate the target execution chain, and the target execution result corresponding to the intent to be processed is generated based on the target execution chain; The target execution result is sent to the end device.

[0009] According to a fourth aspect of the embodiments of this specification, a computing device is provided, comprising: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions. When the computer programs / instructions are executed by the processor, they implement the steps of the above-described information processing method.

[0010] According to a fifth aspect of the embodiments of this specification, an electronic device is provided, comprising: A memory and a processor, the memory and the processor being connected via a bus; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the above-described information processing method.

[0011] According to a sixth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores a computer program / instructions that, when executed by a processor, implement the steps of the information processing method described above.

[0012] According to a seventh aspect of the embodiments of this specification, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the steps of the information processing method described above.

[0013] According to the information processing method provided in this manual, by acquiring the intent to be processed and its corresponding reference skills, and determining the initial execution link accordingly, the appropriate skills can be automatically matched and execution paths can be constructed based on the intent to be processed, thus avoiding the inefficiency and high cost of manually configuring execution links. By generating reference execution results based on the initial execution link and recording link attribute information, complete recording and traceability of the execution process are achieved, providing sufficient data support for subsequent failure analysis. When the reference execution results do not meet the preset verification conditions, the target failure type is determined through link attribute information, enabling precise location and typological attribution of the failure cause, allowing different types of failures to be handled separately, rather than rewriting the failure execution link indiscriminately. By adjusting the initial execution link according to the target failure type and generating the target execution link, targeted link adjustment based on failure type is achieved, avoiding overfitting to samples that trigger failures, and reducing the risk of damaging historical correct capabilities and interfering with adjacent skills. Therefore, according to the information processing method provided in this manual, while taking into account the repair effect, historical stability, and cross-skill interference control, the long-term stable evolution of the skill library is supported, thereby achieving efficient and accurate skill verification. Attached Figure Description

[0014] Figure 1 This specification shows an architecture diagram of an information processing system provided in one embodiment; Figure 2 A flowchart of an information processing method according to an embodiment of this specification is shown; Figure 3 A flowchart illustrating the processing procedure of an information processing method applied to the proposal control plane according to an embodiment of this specification is shown. Figure 4 A flowchart illustrating the processing procedure of an information processing method applied to an access control plane according to an embodiment of this specification is shown. Figure 5 This specification shows a schematic diagram of the structure of an information processing apparatus according to one embodiment. Figure 6A structural block diagram of a computing device according to one embodiment of this specification is shown; Figure 7 A structural block diagram of an electronic device provided according to one embodiment of this specification is shown. Detailed Implementation

[0015] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0016] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0017] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0018] Furthermore, 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 one or more embodiments of this specification are obtained through open-source datasets or public datasets that comply with their license agreements, or are obtained with full authorization from the relevant parties. Moreover, 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.

[0019] In one or more embodiments of this specification, a large model refers to a deep learning model with a large number of model parameters, typically containing hundreds of millions, tens of billions, hundreds of billions, trillions, or even tens of trillions of model parameters. A large model can also be called a foundation model. It is pre-trained using large-scale unlabeled corpora to produce a pre-trained model with hundreds of millions of parameters. Such models can adapt to a wide range of downstream tasks and have good generalization ability. Examples include Large Language Models (LLMs) and multi-modal pre-training models.

[0020] In practical applications, large models only require a small number of samples to fine-tune the pre-trained model before they can be applied to different tasks. Large models can be widely used in fields such as Natural Language Processing (NLP) and Computer Vision. Specifically, they can be applied to computer vision tasks such as Visual Question Answering (VQA), Image Captioning (IC), and Image Generation, as well as natural language processing tasks such as text-based sentiment classification, text summarization, and machine translation. The main application scenarios of large models include digital assistants, intelligent robots, search, online education, office software, e-commerce, and intelligent design.

[0021] The information processing methods provided in this manual are applicable to scenarios involving skill maintenance, such as scenarios where intelligent processing units invoke multiple skills to generate question-and-answer results. Examples include intelligent customer service scenarios and knowledge-based question-and-answer scenarios.

[0022] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0023] An intelligent processing unit, also known as an agent, is a computer program that invokes large language models, retrieval tools, task tools, and runtime control modules to complete query understanding, skill routing, evidence retrieval, and answer generation. An intelligent agent can automatically select appropriate skills and orchestrate execution flows based on user-input query requests, thereby completing complex task processing.

[0024] A skill refers to an external capability unit that can be selected and reused by an agent. It may include route descriptions, execution methods, retrieval strategies, response constraints, scripts, reference materials, and testing experience. Skills can be stored in Lightweight Data Interchange Format (JavaScript Object Notation, JSON), Data Serialization Format (YAML Ain't MarkupLanguage, YAML), database records, or other machine-readable formats. Multiple skills, their versions, route information, constraints, source data, and test records constitute a skill library, forming a shared skill set. In this specification, the skill library can be understood as an intent skill library, which can be divided into different versions according to its evolution. For ease of explanation, this specification uses the current intent skill library to explain the skills constituted in one version.

[0025] A routing description is a machine-readable description used to characterize the applicable intent, product, task, or query boundary of a skill, and to enable a router to select the skill. A routing description may contain intent identifiers, keywords, semantic vectors, and other information, used to match the most relevant skill to the query request within the skill base.

[0026] An execution chain refers to the sequence of skill invocations and execution paths that an agent undertakes when processing a specific intent. An execution chain can include invocations of one or more skills, as well as intermediate results such as routing scores, retrieval results, and generated outputs for each skill. The quality of the execution chain directly affects the accuracy and completeness of the final answer.

[0027] With the development of computer and internet technologies, intelligent agent technology has been widely applied. Intelligent agents typically complete tasks such as query understanding, skill routing, evidence retrieval, and answer generation by invoking large language models, retrieval tools, task tools, and runtime control modules. During execution, the intelligent agent adapts to different specialized tasks through multiple reusable skill units. Each skill unit can include routing descriptions, execution methods, retrieval strategies, and answer constraints.

[0028] In multi-skill intelligent agent systems, multiple skills share routing descriptions and execution environments. As product documentation, interface versions, and user expressions change, skill units need continuous synchronization and adaptation; otherwise, issues such as incorrect routing, outdated evidence, and missing answers may occur. Currently, failure cases are typically identified manually and the corresponding skills are fixed. This maintenance method is time-consuming, costly, and lacks traceability. Furthermore, when product API calls change, skills need continuous synchronization and adaptation; otherwise, routing, retrieval, and responses may all become invalid.

[0029] However, the aforementioned manual repair method carries several risks: firstly, the repair may overfit to samples that failed to trigger the error, failing to generalize to similar unseen problems; secondly, modifications may disrupt historical correctness, causing previously working skills to regress; and thirdly, modifications may alter the routing boundaries of adjacent skills, triggering cross-skill interference. Performing full-database regression testing after each modification would incur high retrieval, model invocation, and time costs. Therefore, an effective solution is needed to maintain the accuracy and effectiveness of various skills within a multi-skill agent during interaction.

[0030] To address the aforementioned technical problems, this specification provides an information processing method. This specification also relates to an information processing apparatus, a computing device, a computer-readable storage medium, and a computer program product, which will be described in detail in the following embodiments.

[0031] See Figure 1 , Figure 1 This specification illustrates an architecture diagram of an information processing system according to one embodiment of the present specification. The information processing system may include a client 100 and a server 200. Client 100 is used to send pending intents to server 200; Server 200 is configured to acquire a pending intent and at least one reference skill corresponding to the pending intent, and determine an initial execution link corresponding to the pending intent based on the at least one reference skill; generate a reference execution result corresponding to the pending intent based on the initial execution link, and record the link attribute information corresponding to the initial execution link; if the reference execution result does not meet preset verification conditions, determine a target failure type based on the link attribute information; adjust the initial execution link based on the target failure type, generate a target execution link, and generate a target execution result corresponding to the pending intent based on the target execution link; and send the target execution result to client 100.

[0032] Client 100 is also used to receive the target execution result sent by server 200.

[0033] The information processing system may include multiple clients 100 and a server 200. Clients 100 can be referred to as end-side devices, and the server 200 can be referred to as cloud-side devices. Multiple clients 100 can establish communication connections through the server 200. In the information processing scenario, the server 200 is used to provide information processing services between the multiple clients 100. Each client 100 can act as a sender or receiver, communicating through the server 200.

[0034] Users can interact with server 200 through client 100 to receive data sent by other clients 100, or send data to other clients 100, etc. In information processing scenarios, users can publish data streams to server 200 through client 100, server 200 can generate target execution results based on the data stream, and push the target execution results to other clients that have established communication.

[0035] In this system, client 100 and server 200 establish a connection via a network. The network provides the medium for communication between client 100 and server 200. The network can include various connection types, such as wired or wireless communication links or fiber optic cables. Data transmitted by client 100 may need to undergo encoding, transcoding, compression, or other processing before being published to server 200.

[0036] Client 100 can be a browser, application (APP), or web application such as H5 (HyperText Markup Language 5) application, or a lightweight application (also known as a mini-program), or cloud application, etc. Client 100 can be developed based on the software development kit (SDK) of the corresponding service provided by server 200, such as based on the Real-Time Communication (RTC) SDK. Client 100 can be deployed in a computing device and depends on the device to run or on certain APPs on the device. The computing device may have a display screen and support information browsing, such as a personal mobile terminal such as a mobile phone, tablet computer, or personal computer. Various other types of applications can also be configured in the computing device, such as human-computer interaction applications, model training applications, text processing applications, web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0037] Server 200 may include servers providing various services, such as servers providing communication services to multiple clients, servers supporting backend training of models used on clients, and servers processing data sent by clients. It should be noted that server 200 can be implemented as a distributed server cluster composed of multiple servers, or as a single server. The server can also be a server in a distributed system, or a server integrated with blockchain. The server can also be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms, or an intelligent cloud computing server or intelligent cloud host with artificial intelligence technology.

[0038] It is worth noting that the information processing methods provided in the embodiments of this specification are generally executed by the server. However, in other embodiments of this specification, the client may also have similar functions to the server, thereby executing the information processing methods provided in the embodiments of this specification. In other embodiments, the information processing methods provided in the embodiments of this specification may also be executed jointly by the client and the server.

[0039] See Figure 2 , Figure 2 A flowchart of an information processing method according to an embodiment of this specification is shown, which specifically includes the following steps.

[0040] Step 202: Obtain the intent to be processed and at least one reference skill corresponding to the intent to be processed, and determine the initial execution link corresponding to the intent to be processed based on at least one reference skill.

[0041] The intent to be processed can be understood as the intention or purpose expressed in the user query that the intelligent processing unit needs to respond to during the current processing. This intent can be obtained through semantic analysis of the user's query request or selected from an intent skill library.

[0042] A reference skill can be understood as a skill that needs to be invoked to execute the intent to be processed, determined based on the intent to be processed. For example, at least one reference skill corresponding to the intent to be processed can be determined based on a pre-determined correspondence between intents and intent skills, or at least one reference skill matching the semantics of the intent to be processed can be determined based on the semantics of the intent to be processed. At least one initial execution chain can be constructed based on the reference skills.

[0043] The initial execution path can be understood as the initial execution path constructed for the intent to be processed based on reference skills. The initial execution path includes information such as the selected reference skills and their invocation order, the routing scores corresponding to each reference skill, and the retrieval strategy configuration. The initial execution path reflects the system's initial judgment on how to process the intent in the current state, and its quality depends on the accuracy of the reference skill selection and the current version state of the intent skill library.

[0044] In one specific embodiment provided in this specification, a processing intent is obtained, and at least one reference skill corresponding to the processing intent is determined based on the processing intent. On this basis, an initial execution chain is generated according to the processing intent and each reference skill.

[0045] In one specific implementation, the intent semantics corresponding to the intent to be processed are determined, and at least one reference skill matching the intent semantics is identified among multiple skills. In this case, the execution order corresponding to each reference skill is determined based on the degree of matching between the intent semantics and each reference skill, thereby obtaining the initial execution chain.

[0046] In one scenario, the initial execution chain can be determined by executing each reference skill sequentially according to its corresponding execution order, or it can be determined by selecting a subset of reference skills according to their corresponding execution order to facilitate parallel and sequential execution.

[0047] For ease of understanding, this specification explains the method of determining the intent to be processed and the method of determining at least one reference skill corresponding to the intent to be processed in the following manner.

[0048] In one specific embodiment provided in this specification, obtaining the intent to be processed and at least one reference skill corresponding to the intent to be processed includes S2022-S2024: S2022. Obtain the current intent skill library, determine the intent to be processed in the current intent skill library, and obtain at least one reference skill corresponding to the intent to be processed, wherein the current intent skill library is constructed according to the correspondence between intents and intent skills.

[0049] The current intent skill library can be understood as a skill index structure built based on the correspondence between intents and intent skills. The current intent skill library also includes current version information. It's important to note that the current version information of the intent skill library can be updated based on the reference execution results of subsequent initial execution chains. Furthermore, intents and intent skills in the current intent skill library can be invoked independently. Moreover, this specification introduces a version management mechanism into the intent skill library, assigning a version identifier to each intent and intent skill to ensure that the correct correspondence between intents and intent skills can be traced after the intent skill library is updated.

[0050] Optionally, the current intent skill library can be constructed using hash tables, vector indexes, or simply graphs.

[0051] In one specific embodiment provided in this specification, a pre-constructed current intent skill library based on the correspondence between intents and intent skills is obtained. From the pre-constructed current intent skill library, a corresponding intent is randomly selected or selected according to preset selection conditions as the intent to be processed. The intent skill corresponding to the intent to be processed in the current intent skill library is used as a reference skill.

[0052] Furthermore, in scenarios where intent acquisition is performed using the current intent skill library, the quality of the intent skill library's construction directly affects the accuracy of intent matching and skill invocation.

[0053] In one specific embodiment provided in this specification, the current intent skill library is constructed in the following manner: Obtain multiple historical query requests and at least one intent skill corresponding to each historical query request; Based on the request semantics of each historical query request, generate at least one intent cluster; Based on at least one intent skill corresponding to each historical query request in each intent cluster, construct the correspondence between intents and intent skills; Generate the current intention skill library based on the correspondence between intentions and intention skills.

[0054] Historical query requests can be understood as user query data accumulated during historical operation. These historical query requests can originate from users' actual usage records, test cases, or open-source datasets.

[0055] The request skill can be understood as the skill that is actually invoked when processing historical query requests.

[0056] Request semantics can be understood as the semantic representation obtained after semantic analysis of historical query requests, which can be obtained through vector embedding models, keyword extraction, or semantic parsing.

[0057] Intent clustering can be understood as grouping historical query requests with similar semantics into the same category to form an intent set.

[0058] In one specific embodiment provided in this specification, multiple historical query requests are obtained, and at least one intent skill invoked during the execution of each historical query request is acquired. In this case, clustering is performed based on the request semantics of each historical query request to obtain at least one intent cluster. Each intent cluster includes at least one historical query request. The intent skill corresponding to a historical query request belonging to the same intent cluster is determined as the intent skill of that intent cluster. It should be noted that, since there may be cases where different historical query requests correspond to the same intent skill within the same intent cluster, this specification can deduplicate multiple intent skills within the same intent cluster before determining at least one intent skill corresponding to that intent cluster, to avoid duplicate construction of at least two identical intent skills within the same intent cluster. Based on the deduplicated intent skills and intent clusters, a correspondence between intents and intent skills is constructed, thereby generating the current intent skill library.

[0059] Optionally, this specification may also introduce a quality filtering mechanism during the clustering process, such as retaining only intent clusters with a query count exceeding a preset threshold and an average execution quality that meets the standard, to ensure the generated intent skill base has high reliability. The intent skill base can also be updated periodically or triggered automatically, for example, automatically triggering re-clustering and mapping updates when the accumulated new query data exceeds a certain level.

[0060] In one example, multiple historical query requests are received and clustered to obtain at least one intent cluster. Based on the intent cluster, a target execution path is generated according to the tools, operation methods, and corresponding target results invoked during the execution of each historical query request within that intent cluster. This target execution path can be understood as a standardized execution process for that intent cluster within the intelligent processing unit. Throughout this process, the target results all meet preset verification conditions. Furthermore, the relationships between the intent (i.e., source information), query requests, and routing descriptions corresponding to the intent skills are saved. It should be noted that the source information corresponding to each intent skill is not limited to unsupervised clustering; it can also be obtained through manual classification, task directories, rules, or supervised models. This further determines the correspondence between each intent and intent skill, and assigns a version identifier and content summary to the current intent skill library.

[0061] According to a specific implementation method provided in this specification, by performing semantic analysis and clustering on historical query requests, the natural distribution patterns of user intents can be automatically discovered, and queries with similar semantics can be grouped into the same intent cluster. Then, based on the request skills actually invoked by historical queries in each intent cluster, a mapping relationship between intents and skills can be automatically constructed, avoiding subjective biases and maintenance costs caused by manual configuration. This historical data-driven construction method enables the intent skill library to reflect the actual distribution of user needs and skill invocation patterns.

[0062] In one specific embodiment provided in this specification, the intent to be processed can also be determined based on the user's query request.

[0063] S2024. In response to detecting a target query request, determine the intent to be processed based on the target query request, and determine at least one reference skill corresponding to the intent to be processed.

[0064] The target query request can be understood as the real-time query input of the user monitored by the system, which can be a natural language query, a structured command, or a multimodal input.

[0065] In one specific implementation provided in this specification, upon detecting a user's real-time query input target query request, a pending intent is determined based on the user's target query request. This pending intent may be the same as or different from intents in the current intent skill library. Based on this, at least one corresponding reference skill is determined according to the intent semantics corresponding to the pending intent.

[0066] In one specific embodiment provided in this specification, determining at least one reference skill corresponding to the intent to be processed includes: Determine the intent semantics corresponding to the intent to be processed, and obtain the skill semantics corresponding to each intent skill in the current intent skill library; Based on the intent semantics and the skill semantics corresponding to each intent skill, at least one initial reference skill corresponding to the intent to be processed is determined, and the initial routing score corresponding to each initial reference skill is determined. The initial reference skill corresponding to the initial route score that meets the preset route threshold is determined as the reference skill corresponding to the intent to be processed.

[0067] Intent semantics can be understood as the semantic representation obtained after semantic encoding of the intent to be processed, which can be represented in the form of vector embedding, semantic tags or structured semantic framework.

[0068] Skill semantics can be understood as the semantic representation of each intention skill in the intention skill library, which is usually generated by combining information such as the skill's routing description, applicable scenarios, and functional description.

[0069] The initial route score can be understood as a quantified value of the degree of matching between intent semantics and skill semantics, which can be calculated using cosine similarity, Euclidean distance or other distance metrics.

[0070] The preset routing threshold can be understood as a threshold value used to filter the initial routing scores, which can be adjusted according to the task scenario and accuracy requirements.

[0071] In one specific implementation provided in this specification, by calculating the degree of matching between intent semantics and skill semantics and combining it with routing thresholds for filtering, skills related to the current intent can be accurately selected from the skill library, while skills with low relevance are filtered out, thus avoiding interference from irrelevant skills on the execution chain and improving the accuracy and efficiency of reference skill selection.

[0072] In one specific implementation, when the intelligent processing unit determines the intent to be processed, it matches the intent semantics corresponding to the intent to be processed with the skill semantics corresponding to each intent skill in the current intent skill library to obtain at least one initial reference skill. At this time, the routing score corresponding to each initial reference skill is determined, and the initial reference skill whose initial routing score meets a preset routing threshold is determined as the reference skill corresponding to the intent to be processed.

[0073] Optionally, a multi-dimensional scoring mechanism can be introduced on top of semantic matching, such as comprehensively considering factors like the skill's historical execution success rate and response time, to improve the comprehensiveness of the routing score. Dynamic routing thresholds can also be used, automatically adjusting the preset routing threshold size based on the current skill library size and load.

[0074] Furthermore, the two intent acquisition methods can be combined in this specification. For example, candidate intents can be obtained from the current intent skill library first, and then the real-time semantics of the target query request can be combined for secondary verification and filtering to improve the accuracy of intent recognition.

[0075] According to a specific implementation method provided in this specification, by providing two complementary intent acquisition paths, it is possible to both utilize the built intent skill library for systematic intent matching and directly respond to real-time query requests for dynamic processing, thereby adapting to the intent acquisition needs under different task scenarios and improving the flexibility and coverage of intent acquisition to be processed.

[0076] Step 204: Generate a reference execution result corresponding to the intent to be processed based on the initial execution link, and record the link attribute information corresponding to the initial execution link.

[0077] The reference execution result can be understood as the output obtained after the initial execution chain is actually executed. The reference execution result reflects the accuracy of each reference skill call on the initial execution chain, thereby determining whether the preset verification conditions are met. In this specification, the reference execution result may include response text, retrieval evidence, tool call results, etc. By comparing the reference execution result with the preset verification conditions, it can be determined whether the current execution chain can produce output that meets the quality requirements.

[0078] The preset verification conditions can be understood as pre-defined conditions used to determine whether the reference execution result is qualified. These can include correctness, completeness, relevance, evidence-based citation correctness, and the initial link quality indicators reaching the corresponding quality indicator thresholds.

[0079] Link attribute information can be understood as various attribute data recorded at different points in time during the initial execution of the link. It's important to clarify that link attribute information provides the data foundation for subsequent failure type determination; that is, it provides the data basis for failure attribution, enabling the system to analyze the reasons for execution failure from multiple dimensions.

[0080] In one specific embodiment provided in this specification, the link attribute information includes at least one of the following: The current version information corresponding to the current intent skill library, wherein the current intent skill library is constructed based on the correspondence between intents and intent skills; The intent identifier corresponding to the intent to be processed; Each reference skill and its corresponding routing score; The execution attribute information corresponding to the reference execution result; Initial link quality metrics.

[0081] The current version information can be understood as the version identifier and content summary information corresponding to the current intent skill library. The version identifier can be a version number, version hash, generation time, parent version identifier, etc.; the content summary can be a hash value or summary of content such as route description, constraints, retrieval configuration, execution script, etc.

[0082] It's important to clarify that the current version information is used to ensure that subsequent replay outputs can be associated with the accurate version, supporting version comparison, atomic commits, rollbacks, and auditing. When attributing failures, it can determine whether the failure is related to a specific version; when isolating replays, it can clearly distinguish between the current version and candidate versions.

[0083] An intent identifier can be understood as identification information used to uniquely identify the intent to be processed, such as intent ID, semantic family identifier, query identifier, etc. It can be used to distinguish different intents and supports clustering, deduplication, time isolation, and replay partitioning.

[0084] It should be clarified that, based on the intent identifier, the semantic family to which the target failure type belongs can be determined in the scenario.

[0085] The routing score can be understood as a quantitative score of the degree of matching between the intent to be processed and the reference skill, usually calculated based on the similarity between the semantics of the intent and the semantics of the skill. The higher the routing score, the better the reference skill matches the intent to be processed.

[0086] It's important to clarify that routing scores are used to determine whether reference skills meet preset routing thresholds, thereby deciding which skills enter the candidate list and which skill is ultimately selected for execution. In failure attribution, routing scores can be used to determine if a routing failure has occurred, such as a high-scoring skill not being selected or a low-scoring skill being mistakenly selected.

[0087] Execution attribute information can be understood as the execution process attribute information corresponding to the reference execution result, that is, the process data generated when the reference execution result is generated. It can include: retrieval request; returned evidence; reference links and evidence time; number of search calls; input and output tokens; latency; error status; selected skills, candidate skill list, etc.

[0088] It is important to clarify that execution attribute information is used for failure attribution and admission determination. For example: retrieval requests and returned evidence can identify retrieval failures; citation links and evidence timestamps can identify citation errors or outdated evidence; search call counts, tokens, and latency can be used for cost hard constraints; and error states can identify execution anomalies.

[0089] Initial link quality metrics can be understood as indicators obtained after evaluating the quality of the reference execution results, used to determine whether the reference execution results meet preset verification conditions. These may include: correctness; completeness; relevance; correctness of evidence-based citations; and presentation quality.

[0090] It is important to clarify that the initial link quality metrics are the key basis for determining the failure type. If the initial link quality metrics do not meet the quality metric thresholds in the preset verification conditions, the failure attribution and link adjustment process will commence.

[0091] In one specific embodiment provided in this specification, upon obtaining link attribute information, the skill performance of each skill in the initial execution link is determined based on the link attribute information.

[0092] In one specific embodiment provided in this specification, the method further includes: Determine whether the reference execution result meets the preset verification conditions; If the reference execution result meets the preset verification conditions, the intent to be processed and each reference skill are recorded. The current intent skill library is updated based on the intent to be processed and each reference skill, wherein the current intent skill library is constructed based on the correspondence between intents and intent skills.

[0093] The preset verification conditions can be understood as pre-defined conditions used to determine whether the reference execution result is qualified. These conditions may include one or more quality indicators such as correctness, completeness, relevance, evidence-based assessment, citation correctness, and presentation quality reaching a corresponding threshold.

[0094] In one specific embodiment provided in this specification, the quality of the reference execution results generated based on the initial execution chain is evaluated to determine whether they meet preset verification conditions. If the reference execution results meet the preset verification conditions, the intent to be processed and each reference skill participating in this execution are saved. This provides a data foundation for subsequent updates to the skill library. The recorded content may include query identifiers, semantic family identifiers, routing scores, execution attributes, etc. Using the recorded successful cases, the current intent skill library is updated. Update methods may include: adding new source queries, strengthening the correspondence between intents and skills, updating skill routing descriptions, supplementing testing experience, and adjusting the initial test set, etc.

[0095] According to a specific implementation method provided in this specification, successful cases are incorporated into the current intent skill base update, enabling the current intent skill base to learn from both failures and successes simultaneously, forming a complete dual-timescale closed loop. By recording successful intent-skill correspondences, the source clustering and representative queries of the skill base can be supplemented, making the correspondence between intents and skills more accurate and improving the hit rate of subsequent route matching. Furthermore, successful cases can serve as positive samples for history partitions, used during subsequent isolation and replay to detect whether candidate changes cause historical correct cases to degrade, helping to protect the quality of existing data and avoiding the destruction of historical correct cases while fixing real-time issues.

[0096] Step 206: If the reference execution result does not meet the preset verification conditions, determine the target failure type based on the link attribute information.

[0097] Pre-defined verification can be understood as the criteria for judging the accuracy of each reference skill invoked in the initial execution chain. These pre-defined verification criteria can be evaluated using multiple dimensions of indicators, including the correctness, completeness, relevance, and sufficiency of evidence in the judgment of the reference execution results. Quality assessment can be achieved through manual review, rule-based judgment, or an independent evaluation model.

[0098] Target failure types can be understood as specific failure categories determined by analyzing link attribute information. Failure types can include routing failures, retrieval failures, generation failures, presentation failures, and noise failures. Routing failures refer to incorrect skill selection, such as skill leakage, misselection, or product confusion; retrieval failures refer to missing evidence, outdated links, or insufficient coverage; generation failures refer to overconfident conclusions or logical inconsistencies despite sufficient evidence; presentation failures refer to language mismatches, leakage of internal reasoning, or formatting errors; and noise failures refer to misattribution of query results due to prior experience or unreliable citation tags.

[0099] A target execution path can be understood as a new execution path obtained by adjusting the initial execution path after determining the target failure type. The specific adjustment method depends on the target failure type. For example, for routing failure, it may be necessary to change or adjust the reference skill; for retrieval failure, it may be necessary to modify the retrieval strategy or update the evidence source; for generation failure, it may be necessary to adjust the answer constraints. The generation of the target execution path may include determining the skills to be determined and generating the target execution path if each skill to be determined meets the skill launch conditions.

[0100] In one specific implementation provided in this specification, if the reference execution result does not meet the preset verification conditions, the target failure type is determined based on the link attribute information in order to perform failure attribution.

[0101] When the reference execution result does not meet the preset verification conditions, it is necessary to accurately identify the specific target failure type from the link attribute information so that targeted link adjustments can be made subsequently based on the failure type. To avoid the failure type determination method being too crude, such as uniformly classifying all failures as "execution exception", which would fail to provide precise guidance for link adjustments, resulting in a lack of targeted adjustment strategies, inability to effectively solve actual execution problems, or even the introduction of new problems.

[0102] In view of this, in a specific embodiment provided in this specification, determining the target failure type based on the link attribute information includes: Parse the link attribute information to determine the execution failure description information; Based on the execution failure description information, determine the target failure type corresponding to the execution failure description information from at least one failure type.

[0103] Among them, the execution failure description information can be understood as structured information obtained by parsing and extracting the link attribute information, which describes the specific phenomena and characteristics of execution failure.

[0104] The failure description information may include the specific stage where the failure occurred (such as the routing stage, retrieval stage, generation stage, or presentation stage), the specific manifestation of the failure (such as no skill match, zero search results, missing key fields or format errors in the generated content, etc.), the skill identifier and intent identifier associated with the failure, as well as the original execution log fragments and error codes corresponding to the failure.

[0105] Failure types can be understood as categories within a predefined set of failure categories, including at least routing failures, retrieval failures, generation failures, presentation failures, and noise failures. Each failure type corresponds to a set of predefined failure feature description templates, used for matching and comparing with execution failure description information. Specifically: Routing failures: omissions or misselections of specialized skills, or confusion regarding products or intentions; Search failures: missing evidence, outdated evidence, invalid links, insufficient source coverage; Errors include: insufficient evidence but incorrect conclusions, unsupported information combinations, or missing key points. Faults may occur due to language mismatch, internal reasoning leakage, redundancy, indirection, or incorrect formatting. Noise fault: Failed queries were incorrectly categorized into current skill experience, or the reference tags are unreliable.

[0106] In one specific embodiment provided in this specification, link attribute information is read, and routing information, retrieval information, evidence information, response information, quality indicators, cost information, etc., are analyzed to identify the specific manifestations and possible causes of the execution failure, forming execution failure description information. This execution failure description information is then matched with a preset or dynamically maintained set of failure types to determine the target failure type corresponding to the current failure.

[0107] Optionally, a confidence scoring mechanism can be introduced during the matching process between execution failure descriptions and failure types. A confidence score is calculated for each matched failure type. When the highest confidence score exceeds a preset confidence threshold, the corresponding failure type is identified as the target failure type. When all confidence scores are below the threshold, the failure is marked as an unknown type and a manual review process is triggered. A multi-label classification mechanism can also be introduced, allowing a single execution failure to belong to multiple failure types simultaneously, such as a composite failure scenario involving both routing and retrieval failures. An independent adjustment strategy can be assigned to each failure type. Furthermore, historical failure cases can be used as reference samples to improve the accuracy of failure type matching through case-based reasoning.

[0108] According to a specific implementation method provided in this specification, by first parsing the link attribute information to extract structured execution failure description information, and then matching the execution failure description information with predefined failure types, accurate identification of failure types can be achieved. This two-stage failure type determination method first transforms unstructured link attribute information into structured failure descriptions, and then performs type matching based on standardized failure descriptions, avoiding the uncertainty caused by directly judging types from raw attribute information. The structured processing of execution failure description information makes failure features clearer and more comparable, while the matching mechanism based on predefined failure types ensures the consistency and traceability of type judgment, providing an accurate failure attribution basis for subsequent generation of target constraints.

[0109] Step 208: Adjust the initial execution chain according to the target failure type, generate the target execution chain, and generate the target execution result corresponding to the intent to be processed according to the target execution chain.

[0110] When adjusting the initial execution chain based on the target failure type, a reasonable adjustment strategy is needed to avoid the risk of overfitting caused by adjusting based on a single failure.

[0111] In one specific embodiment provided in this specification, adjusting the initial execution chain according to the target failure type includes: Based on the target failure type, at least one historical failure type is obtained within a preset historical query time period, wherein the semantic similarity between each historical failure type and the target failure type satisfies a similarity threshold. If the target failure type and each historical failure type meet the preset adjustment conditions, the target constraint conditions are determined based on the target failure type. The initial execution chain is adjusted according to the target constraints.

[0112] The preset historical query time period can be understood as a time window used to find historical failure records. It can be set according to task requirements, such as the past week, the past month, or a custom time period.

[0113] Historical failure types can be understood as historical failure events recorded within a certain time period that are semantically similar to the target failure type.

[0114] Semantic similarity can be understood as being calculated through methods such as vector matching of failed descriptions, keyword overlap, or failure mechanism classification.

[0115] Preset adjustment conditions can be understood as the conditions that need to be met to trigger link adjustments, such as the cumulative number of similar failures exceeding a threshold, or failures covering multiple different queries.

[0116] Target constraints refer to the constraint rules generated based on failure analysis, used to guide link adjustments.

[0117] In one specific implementation provided in this specification, the target failure type is used as the retrieval basis. At least one historical failure type is obtained within a preset historical query time period. A semantic similarity threshold is used to filter historical failure types that are semantically similar to the current target failure type, thereby linking the current single failure with historical cumulative failure evidence and avoiding hasty adjustments based solely on a single anomaly. When the target failure type and all historical failure types jointly meet preset adjustment conditions, target constraints are determined. These preset adjustment conditions can be understood as cumulative evidence conditions, such as a certain number of similar historical failure types, originating from multiple semantic families, spanning multiple time points or multiple frozen checkpoints, or meeting attribution confidence requirements. Target constraints may include positive and / or negative constraints to guide subsequent answering strategies or prevent error-prone areas. The initial execution chain is adjusted based on the target constraints. For example, the target constraints are applied to relevant skills, routing descriptions, retrieval configurations, answer coverage dimensions, or prohibition rules in the initial execution chain, enabling the adjusted execution chain to perform targeted repairs for the target failure type and its similar historical failure types, rather than making unfounded global modifications.

[0118] Optionally, a failure severity assessment mechanism can be introduced. For failures with a wide impact or strong user feedback, the cumulative threshold for triggering adjustments can be lowered to accelerate problem fixing. Furthermore, target constraints can be linked to the positive and negative constraints of skills, allowing adjustment results to be directly reflected in the skill's constraint configuration.

[0119] According to a specific implementation provided in this specification, lifecycle operations are proposed using a dual time scale based on current failures and accumulated historical failure evidence. These operations include update, creation, splitting, merging, pruning, decommissioning, and delaying. Structural changes require accumulated evidence conditions to prevent widespread structural disturbances. By searching for historical failures similar to the target failure type within a historical time period and triggering link adjustments only when preset adjustment conditions are met, overreaction to occasional failures is avoided, while ensuring timely responses to cumulative and systemic problems. The generation of target constraints provides clear guiding principles for link adjustments, rather than simply rewriting the execution link, improving the accuracy and controllability of the adjustments.

[0120] In one specific implementation provided in this specification, when adjusting the initial execution link according to the target failure type, it is necessary to accurately locate the specific skill in the link that caused the failure and consider the interference effect that the adjustment of this skill may have on other related skills. If only the specific skill that caused the failure is repaired in isolation without considering the interdependence and routing boundary relationships between skills, it may lead to new conflicts between the repaired skill and adjacent skills, causing cross-skill interference or regression problems, thereby reducing the stability and reliability of the overall execution link.

[0121] In one specific embodiment provided in this specification, the method further includes: Based on the target failure type, identify the skills to be adjusted in the initial execution chain; Based on the skill to be adjusted, at least one candidate interference skill is determined in the skill interference graph, and the initial execution link is adjusted according to each candidate interference skill, wherein the skill interference graph is constructed based on the intention skill.

[0122] In this context, "skills to be adjusted" can be understood as skill units in the initial execution chain that are identified as causing execution failure or contributing the most to failure. Skills to be adjusted can be determined by analyzing the routing scores, execution logs, and output quality metrics of each skill in the chain attribute information. For example, skills with abnormally low routing scores, large deviations between output results and verification conditions, or error markers in the execution logs are identified as skills to be adjusted.

[0123] A skill interference graph can be understood as a graph structure built with intention skills as nodes and the relationships between skills as edges. It is used to represent the possible mutual influence relationships between various intention skills, including the relationships in dimensions such as overlapping routing boundaries, shared evidence sources, and semantic confusion.

[0124] Candidate interference skills can be understood as skills that have an associated edge with the skill to be adjusted in the skill interference graph and whose association strength meets the preset interference threshold. These skills may be affected by route drift, evidence conflict or semantic boundary changes after the skill to be adjusted is modified.

[0125] In one specific implementation provided in this specification, failure attribution analysis is first performed on each skill in the initial execution chain based on the target failure type to determine the skill with the highest failure contribution as the skill to be adjusted. Then, candidate interfering skills directly connected to the skill to be adjusted are searched in the skill interference graph, and the association weight between each candidate interfering skill and the skill to be adjusted is obtained. The set of skills requiring synchronous adjustment is determined based on the association weight ranking, and protective constraints for candidate interfering skills are incorporated when generating target constraints. For example, the scope of route description modifications for the skill to be adjusted is limited to prevent intrusion into the semantic boundaries of the candidate interfering skills. After adjustment, regression testing is performed on the candidate interfering skills to confirm that their historical correctness is unaffected.

[0126] Optionally, after identifying candidate interfering skills, they can be categorized based on the type and strength of their associations. For example, the candidate interfering skills with the highest association strength can be included in the synchronous adjustment range, regression testing can be performed on skills with medium association strength, and skills with low association strength can only be monitored and marked. A time decay mechanism can also be introduced into the skill interference graph, giving higher weights to associations of skills that have recently changed, while gradually decreasing the weights of older associations, thereby improving the timeliness and accuracy of interference analysis. Furthermore, the adjustment results of candidate interfering skills can be fed back into the skill interference graph to update the weights and types of associations.

[0127] According to a specific implementation method provided in this specification, by accurately locating the skill to be adjusted in the initial execution chain and searching for candidate interfering skills associated with it in the skill interference graph, the impact on associated skills can be considered simultaneously when adjusting the execution chain, thereby avoiding global problems caused by local repairs. During the adjustment process, not only can the skill to be adjusted be modified in a targeted manner, but the routing descriptions, constraints, or retrieval strategies of candidate interfering skills can also be preventively adjusted to ensure that the modified execution chain maintains overall consistency and coordination.

[0128] To facilitate understanding, this manual explains the construction method of the skill interference graph in the following manner.

[0129] In one specific embodiment provided in this specification, the skill interference graph is constructed in the following manner: In the current intent skill library, retrieve the intent skills corresponding to each intent, and determine the skill attribute information corresponding to each intent skill.

[0130] Skill attribute information can be understood as information describing the characteristics of intentional skills, used to determine the relationships between intentional skills. Skill attribute information may include: the intention corresponding to the skill; skill semantics, i.e., route description, skill description, etc.; historical execution links; bidirectional confusion statistics in historical routes; the frequency with which updating one skill led to the degradation of another skill in the past; source clustering, experience memory, etc.

[0131] In one specific embodiment provided in this specification, the intent skills corresponding to each intent are read from the current intent skill library, and the skill attribute information of each intent skill is extracted or calculated. The skill attribute information includes the intent corresponding to the skill, skill semantics, historical execution chain, bidirectional obfuscation statistics, historical regression effects, etc.

[0132] Based on skill attribute information, determine the relationships between various intentional skills; In this context, the correlation can be understood as the degree of association between two intentional skills in terms of routing confusion, semantic similarity, and historical regression effects. A stronger correlation indicates that the two skills are more likely to interfere with each other, and modifying one requires greater attention to the other. By leveraging correlations, hidden boundary confusions and cross-skill influences within the skill library can be made explicit.

[0133] In one specific implementation provided in this specification, when determining the relationship between various intentional skills based on skill attribute information, it is necessary to comprehensively evaluate the degree of mutual influence between skills from multiple dimensions.

[0134] In one specific embodiment provided in this specification, the association between various intentional skills is determined based on skill attribute information, including: Based on the historical execution path of each intent skill, determine the bidirectional obfuscation parameters and historical regression impact parameters between each intent skill; Based on the intent corresponding to each intent skill and the skill semantics corresponding to each intent skill, determine the similarity parameters between each intent skill; The relationships between intention skills are determined based on the bidirectional confusion parameter, historical regression influence parameter, and similarity parameter among the intention skills.

[0135] The bidirectional obfuscation parameter can be understood as a quantitative indicator of the frequency and extent to which two intent skills are incorrectly invoked by each other during historical execution. The bidirectional obfuscation parameter includes obfuscation measures in two directions: the probability that the first intent skill is incorrectly routed to the second intent skill, and the probability that the second skill is incorrectly routed to the first skill. Together, they constitute the bidirectional obfuscation parameter.

[0136] The historical regression impact parameter can be understood as a statistical indicator of the regression problems that occur in the historical execution of other related skills after an intentional skill is modified. It reflects the actual degree of impact of skill modification on the historical correctness of related skills.

[0137] The similarity parameter can be understood as the comprehensive similarity between two intention skills in the intention semantic space and the skill semantic space, which is obtained by weighted fusion of semantic similarity at the intention level and semantic similarity at the skill level.

[0138] The historical execution chain can be understood as the actual call records and execution data of each intent skill during its historical operation.

[0139] In one specific implementation provided in this specification, execution records are extracted from the historical execution logs of each intent skill, and the number and direction of bidirectional obfuscation between each pair of skills are statistically analyzed to calculate bidirectional obfuscation parameters. Simultaneously, the changes in the regression test pass rate of associated skills after each skill modification are analyzed to generate historical regression impact parameters. A semantic embedding model is also used to calculate the intent semantic vector and skill semantic vector corresponding to each intent skill, and similarity parameters are obtained through cosine similarity calculation. Finally, the bidirectional obfuscation parameters, historical regression impact parameters, and similarity parameters are weighted and fused according to a preset weight ratio to obtain a comprehensive correlation degree. When the comprehensive correlation degree exceeds a preset correlation threshold, a correlation relationship is determined between the two skills.

[0140] Optionally, different weight coefficients can be assigned to the bidirectional obfuscation parameter, historical regression influence parameter, and similarity parameter, respectively. The overall correlation degree can be calculated through weighted summation or nonlinear fusion, and the determination of whether to establish a correlation and the weight of the edge can be based on the overall correlation degree. The weight coefficients can be optimized according to the task scenario; for example, in scenarios with high routing stability requirements, the weight of the bidirectional obfuscation parameter can be increased. A time window mechanism can also be introduced, using recent historical execution link data to calculate the bidirectional obfuscation parameter and historical regression influence parameter, reflecting the latest correlation status between skills and avoiding interference from outdated data in the correlation determination.

[0141] According to a specific implementation provided in this specification, by extracting bidirectional obfuscation parameters and historical regression impact parameters from historical execution chains, the degree of interference between skills can be objectively assessed from the perspective of actual operational data, rather than relying solely on static attribute comparisons. The bidirectional obfuscation parameter reveals the ambiguity of routing boundaries between skills; a higher obfuscation frequency indicates that the routing boundaries of two skills are more likely to intrude into each other. The historical regression impact parameter reflects the historical chain reaction of skill modifications; a greater regression impact indicates a stronger dependency between two skills. The similarity parameter supplements the judgment of the intrinsic correlation between skills at the semantic level. Combining the bidirectional obfuscation parameter, historical regression impact parameter, and similarity parameter to determine correlations allows for a comprehensive assessment of the mutual influence between skills at both the behavioral and semantic levels, making the judgment of correlations more accurate and comprehensive.

[0142] The intention skills are defined as nodes, and the relationships between each intention skill are defined as edges between nodes, thus constructing a skill interference graph.

[0143] In one specific embodiment provided in this specification, each intentional skill is treated as a node in a graph, and the relationships between nodes are represented as edges, thereby constructing a skill interference graph. Edges may have weights, which represent the strength of the relationships.

[0144] According to a specific implementation provided in this specification, a comprehensive and accurate skill interference graph can be constructed by systematically extracting attribute information of each intention skill from the current intention skill library and quantitatively analyzing the correlation between skills based on this attribute information. This graph structure includes not only static correlations between skills (such as shared evidence sources, overlapping routing descriptions, etc.) but also dynamic correlations based on historical execution data (such as historical confusion frequency, regression impact, etc.). The graph structure, with intention skills as nodes and correlation relationships as edges, can intuitively present the influence network between skills, providing structured decision support for subsequent interference analysis and link adjustment, thereby avoiding the subjectivity and incompleteness of relying on human experience to judge skill correlations.

[0145] In one specific embodiment provided in this specification, generating the target execution chain includes: Based on the adjusted initial execution chain, at least one skill to be determined is identified; If each skill to be determined meets the skill launch conditions, a target execution link is generated based on each skill to be determined.

[0146] The skill deployment conditions can be understood as the conditions used to determine whether a skill can transition from a candidate state to a formally effective state. These conditions can include performance comparison results, historical regression rate, adjacency interference rate, cost difference, number of constraint conflicts, and coverage loss.

[0147] In one specific implementation provided in this specification, the adjusted initial execution chain is read, and the skills generated by operations such as addition, modification, splitting, merging, trimming, and retirement are identified. These skills that have not yet officially taken effect are designated as skills to be determined. There can be one or more skills to be determined. For each skill to be determined, it is determined whether it meets the skill launch conditions. If it does, these skills to be determined are included in the target execution chain; if there are skills to be determined that do not meet the skill launch conditions, they can be rejected, postponed, or rolled back, and not included in the target execution chain, or the current version can be retained.

[0148] In one specific embodiment provided in this specification, the skill activation conditions are determined in the following manner: Obtain the current skill performance information for each skill to be determined; Configure the isolation skills corresponding to each skill to be determined in the isolation environment, and determine the performance information of each isolation skill. Compare the isolated skill performance information with the current skill performance information to determine the performance comparison result; Based on the performance comparison results, the conditions for skill activation are determined.

[0149] An isolated environment can be understood as a runtime environment that is isolated from the production online environment, does not receive production online traffic, and does not overwrite the current intent skill library. It can include temporary directories, versioned databases, containers, branches, or immutable object storage, etc.

[0150] The performance comparison result can be understood as the result obtained by comparing the performance information of the isolation skill with the performance information of the current skill. The comparison may include: whether the isolation skill is better than the current skill, equal to the current skill, or worse than the current skill; whether new bad cases are introduced; and whether the historical regression rate, adjacency interference rate, cost difference, etc. exceed the threshold.

[0151] In one specific implementation provided in this specification, for each skill to be determined, its current skill performance information in the current intent skill library is obtained. This performance information can come from historical execution records, replay results, or online monitoring data. The necessary state of the current intent skill library is copied in an isolated environment, and each skill to be determined is configured as an isolated skill. Candidate incremental or structural operations are applied to obtain an isolated intent skill library. The isolated intent skill library is executed in the isolated environment to obtain the performance information corresponding to each isolated skill. It should be noted that in the isolated environment, for each isolated skill in the isolated intent skill library, the execution conditions corresponding to each isolated skill should be the same as or similar to the conditions for obtaining the current skill performance information, such as the same router, generation model, retrieval backend, reordering threshold, search budget, and evaluator version. The isolated skill performance information is compared item by item with the current skill performance information to determine the performance comparison result. The comparison content may include target partition quality improvement, historical regression rate, adjacency interference rate, cost difference, token difference, latency difference, constraint conflict number, coverage loss, etc. Based on the performance comparison result, it is determined whether the skill to be determined meets the skill deployment conditions. For example, if the isolation skill performance is better than or equal to the current skill performance, no new bad cases are introduced, and the historical regression rate, adjacent interference rate, and cost difference do not exceed the threshold, then the skill launch conditions are met; otherwise, they are not met.

[0152] According to a specific implementation method provided in this specification, an initial mapping relationship from user intent to skill invocation is established by acquiring the intent to be processed and the reference skill, and determining the initial execution link. Then, by generating a reference execution result and recording the link attribute information, not only is the output quality assessment of the current link obtained, but the complete execution context is also saved. When the reference execution result does not meet the preset verification conditions, failure attribution is performed using the link attribute information to determine the target failure type, thereby transforming the general execution failure into actionable, categorized failure information. Finally, the execution link is adjusted in a targeted manner according to the target failure type, generating the target execution link and producing the target execution result, thus realizing an automated closed loop from failure detection to failure attribution to link repair.

[0153] To facilitate understanding, this specification explains the above information processing method using the following example and in conjunction with two independent control surfaces: the proposal control surface and the access control surface.

[0154] The proposal control plane, based on failures and accumulated health evidence, determines what to modify and generates candidate changes that have not yet taken effect. The admission control plane compares candidate versions with the current version across multiple replay partitions with different functionalities, deciding to accept, reject, or postpone them. Both control planes work together on the currently maintained intent skill library, which is represented as... t represents the current version or evolution time of the intent skill library.

[0155] It should be noted that the construction method of the current intent skill library is detailed in the above content of this manual, and will not be repeated here.

[0156] Skill (S) can be represented as:

[0157] Where d represents the route description, π represents the execution method, C+ represents the positive constraint, and C... Here, M represents the test experience carried by the source, acting as a negative constraint. Given the intent to be processed, the router selects a skill, the executor retrieves evidence based on the skill's execution method and the positive constraint, and then combines the positive and negative constraints to generate an answer.

[0158] The following is in conjunction with the appendix Figure 3 and appendix Figure 4 Taking the application of the information processing method provided in this specification in the intelligent customer service scenario of cloud services as an example, the information processing method will be further explained.

[0159] in, Figure 3 This specification illustrates a flowchart of an information processing method for the proposal control plane, provided by an embodiment of this specification, which specifically includes the following steps.

[0160] Step 302: In response to detecting a query request initiated by a user, obtain the intent to be processed and at least one reference skill corresponding to the intent to be processed, and determine the initial execution link corresponding to the intent to be processed based on each reference skill.

[0161] In one specific implementation, in an intelligent customer service scenario, when a user inputs a query through the dialogue interface, the system monitors the query request in real time, performs semantic parsing on the query request, and determines the intent to be processed. It then matches reference skills related to the intent to be processed from the current intent skill library. An initial execution chain is constructed based on the routing description and execution order of each reference skill.

[0162] It should be noted that the construction method of the current intent skill library is detailed in the above content of this manual, and will not be repeated here.

[0163] Step 304: Determine the intent semantics corresponding to the intent to be processed, and obtain the skill semantics corresponding to each intent skill in the current intent skill library.

[0164] In one specific implementation, the intent to be processed is semantically encoded to generate an intent semantic vector. Simultaneously, the skill semantic vector corresponding to each intent skill in the current intent skill library is obtained. The skill semantic vector is generated based on the skill's routing description, applicable scenarios, and functional specifications.

[0165] Step 306: Based on the intent semantics and the skill semantics corresponding to each intent skill, determine at least one initial reference skill corresponding to the intent to be processed, and determine the initial routing score corresponding to each initial reference skill.

[0166] Step 308: Determine the initial reference skill corresponding to the initial route score that meets the preset route threshold as the reference skill corresponding to the intent to be processed.

[0167] Step 310: Generate a reference execution result corresponding to the intent to be processed based on the initial execution link, and record the link attribute information corresponding to the initial execution link.

[0168] In one specific implementation, the reference skills in the initial execution chain are executed sequentially according to the order in which they are invoked. The execution results of each skill are integrated into a reference execution result, generating a response text containing configuration step descriptions. Simultaneously, chain attribute information is recorded, including the current version information of the current intent skill library, the intent identifier of the intent to be processed, each reference skill and its corresponding routing score, the execution attribute information corresponding to the reference execution result, and the initial chain quality indicators.

[0169] Step 312: Does the reference execution result meet the preset verification conditions?

[0170] In one specific implementation, the reference execution results are verified in multiple dimensions through an evaluation model. The verification dimensions include the accuracy, completeness, relevance, and sufficiency of the answers, as well as whether the initial link quality indicators have reached the corresponding quality indicator thresholds.

[0171] If the reference execution result meets the preset verification conditions, then step 322 will directly use the reference execution result as the final answer; if the reference execution result does not meet the preset verification conditions, then step 314 will be executed.

[0172] Step 314: If the reference execution result does not meet the preset verification conditions, parse the link attribute information to determine the execution failure description information, and determine the target failure type among at least one failure type based on the execution failure description information.

[0173] In one specific implementation, the reasons for execution failure are analyzed based on the link attribute information. For example, based on the execution attribute information corresponding to the reference execution result, if it is determined that the verification found the answer lacks key configuration steps and references an outdated interface version, the link attribute information is parsed to extract the execution failure description information. This execution failure description information includes the failure phenomenon, the location of occurrence, and related skill identifiers. The location of occurrence can be understood as the time node corresponding to the failure and the reference skill.

[0174] The execution failure description information is semantically matched with predefined failure types to determine the target failure type as "retrieval failure - outdated evidence," and an attribution output is generated that includes the failure type, confidence level, affected skill components, and supporting evidence. Failure types include at least one of routing failure, retrieval failure, generation failure, presentation failure, and noise failure. Low-confidence failures may retain multiple candidate attributions simultaneously, or proceed to manual review; immediate modification is not mandatory.

[0175] Step 316: Based on the target failure type, obtain at least one historical failure type within a preset historical query time period, wherein the semantic similarity between each historical failure type and the target failure type meets a similarity threshold.

[0176] For example, three similar historical failure records were found, involving data export queries for different products, but all were of the outdated evidence type, and their semantic similarity all exceeded the similarity threshold of 0.8. By introducing historical failure data, the risk of overfitting caused by adjusting based solely on a single failure was avoided.

[0177] Step 318: If the target failure type and each historical failure type meet the preset adjustment conditions, determine the target constraint conditions.

[0178] In one specific implementation, the preset adjustment condition includes the cumulative number of similar failures exceeding a preset threshold. Specifically, based on the target failure type and historical failure types, the proposal strategy outputs lifecycle operations:

[0179] It is important to clarify that lifecycle operations refer to the lifecycle defined in the initial execution chain for the failure type of the target. These can include updates, creations, splits, merges, prunes, retires, and defers.

[0180] For example, determine the lifecycle operation as an update ( In the case of an update operation, one or more candidate increments are generated based on attribution and evaluation evidence:

[0181] Positive constraints can be written into fields such as official target pages, web search keywords, site document retrieval keywords, supplementary search paths, answer coverage dimensions, evidence verification requirements, and recommended presentation formats. Negative constraints can be written into fields such as product obfuscation, region obfuscation, identifier obfuscation, computational concept obfuscation, operation path obfuscation, prohibition of hard-coding unverified facts, prohibition of outputting links without sources, and prohibition of disclosing internal reasoning. The system performs semantic duplication merging, constraint rejection for excessive modification or only memorizing case answers, positive and negative constraint conflict detection, length or entry budget checks for each type of field, constraint and source failure, evidence evaluation, generator version binding, and encoding of operations for adding, deleting, or replacing existing constraints. The incremental modification of external skill state does not require gradient updates or reinforcement learning to be performed on the base model.

[0182] Step 320: Adjust the initial execution chain according to the target constraints.

[0183] In one specific implementation, during the adjustment process, a skill interference graph is constructed and updated:

[0184] Each node corresponds to one skill. For skill S i and S j The edge weights are calculated as follows:

[0185] in, This is a statistic representing the bidirectional confusion between two skills in historical routing. Similarity between route descriptions, source clusters, or experiential memories; Update skills S in history i This later led to it belonging to S j The frequency of sample degradation; α, β, γ are fixed, learned, or cost-adaptive weights.

[0186] Initial candidate edges can be established based on adjacency relationships from source query clustering, and then dynamically updated by real route obfuscation and observed regression. (This is for modifying skill S.) i The top h skills with the highest edge weights, or skills with edge weights exceeding a threshold τ, are selected to form candidate relevant interference domains. (i). In this specification, candidate interference skills are determined based on the interference map, and the initial execution link is adjusted according to each candidate interference skill.

[0187] in, Figure 4The present specification illustrates a process flowchart of an information processing method for an access control plane according to an embodiment of this specification, which specifically includes the following steps.

[0188] Step 402: Determine at least one skill to be determined, and generate the target execution link if each of the skills to be determined meets the skill launch conditions.

[0189] In one specific implementation, the adjusted skill is used as a skill to be determined, and its compliance with the skill deployment criteria is verified. The skill deployment criteria include: the adjusted skill's pass rate on the replay test set exceeds a pass threshold, it does not affect the quality of historical correct answers, and it does not violate the boundary constraints of adjacent skills.

[0190] Specifically, in the access control plane, the following steps are performed: First, an isolated candidate version of the skill library is constructed. The system copies the necessary state of the current skill library to a temporary directory, versioned database, container, branch, or immutable object store, and applies candidate incremental or structural operations to obtain:

[0191] The candidate version does not receive official online traffic and will not be covered. Candidate records must include at least: candidate identifier, parent version, operation type, affected skills, content differences, source failure, supported semantic families, and generation time.

[0192] Secondly, construct a replay list with separate functions. The replay list includes target partitions, unseen partitions, history partitions, adjacent partitions, general partitions, and coverage partitions. These are used to confirm whether the candidate solves the target problem, detect whether the fix can generalize rather than be a memory case, detect historical capability regression, detect cross-skill misabsorption and route boundary offset, detect whether the specialized skill has become too broad, and detect joint coverage or unique coverage loss. The same semantic family cannot be used as both a proposal construction sample and an unseen partition sample simultaneously. Deduplication and time isolation are preferably performed using query identifiers, semantic family identifiers, and event times. Construct a replay set for candidate changes:

[0193] The target partition is used to trigger failed queries for changes, to confirm whether the candidate resolves the target issue. The unseen partition is used for queries that belong to the same semantic family as the triggering query but did not participate in the proposal construction, to check whether the repair can generalize rather than remember the case; The history partition is used for queries that have been successfully processed by the current skill, and to detect regressions in historical capabilities. The adjacent partition is used to interfere with positive samples of other skills in the domain, and to modify the boundary negative samples of skills for adjacent intentions, to detect cross-skill misabsorption and routing boundary offset. The general partition is used for general or out-of-scope queries that should not be absorbed by any specific skill, and to detect when a specific skill becomes too broad. Coverage partitioning is used to store representative samples with independent and valid behavior before structural operations such as splitting, merging, pruning, and decommissioning, and to detect joint coverage or unique coverage loss.

[0194] The same semantic family cannot be used as both a proposal construction sample and an unseen partition sample. It is preferable to use query identifiers, semantic family identifiers, and event timestamps to perform deduplication and time isolation.

[0195] Then, under the same operating conditions, perform paired replays of candidate and current versions. The system executes the current intent skill set and candidate skill set separately under the same router, generative model, retrieval backend, re-ranking threshold, search budget, and evaluator version. Each replay output retains the original answer, evidence, citations, cost, and partition identifier; it should not only save the summary score. Calculate the following metrics:

[0196]

[0197]

[0198] It can also calculate the same family unseen pass rate, general false absorption rate, unfounded generation rate, reference error rate, over-constraint rate, unique coverage loss, number of constraint contradictions, search call difference, token difference, and latency difference.

[0199] Implement a change admission process that first imposes hard constraints and then optimizes the objective. Candidate k is only admitted to the acceptable set if it meets at least the following feasibility conditions:

[0200] Furthermore, the number of constraint conflicts, the general misabsorption, and the structural coverage loss do not exceed their respective upper limits, where (δ, ε) h， ε n B represents the threshold for the target partition, historical partition, adjacent partition, and resource consumption, respectively.

[0201] For retirement, downsizing, or consolidation operations, a unique validation coverage loss of zero may be required. The threshold can be a fixed value, a value determined by skill risk level, a value determined by confidence intervals, or a dynamic value based on historical distribution. Among feasible candidates, the one with the largest objective function is selected:

[0202] Where Ω(C) represents the penalty term for constraint repetition, contradiction, or unbounded growth. If no candidate satisfies the hard constraint, the output is rejection or postponement, rather than forcing submission to fix the target sample.

[0203] When the gating result is "accept," the system generates a commit record containing the parent version, candidate hash, replay list, evaluation results, affected areas, and decision reasons, and completes the following within a single transaction: writing candidate skill content and dependency status; writing new skill library version metadata; updating the effective version pointer; saving recoverable references to the previous version; and updating the interference graph and future replay experience. If any step in the transaction fails, the effective version pointer remains unchanged. If the gating result is "reject," the candidate content can be retained for auditing, but the current version continues to be effective. If the gating result is "deferred," the system writes the failure and missing evidence to the observation queue. If post-deployment monitoring detects regressions exceeding a threshold, the version pointer can be atomically switched back to the previous version.

[0204] In addition, specific access conditions must be met for structural operations: The Create requirement is that new skills should improve the coverage of uncovered semantic families without increasing the general false absorption rate or adjacent route conflicts above the threshold. Split requires that each sub-skill after splitting should cover a stable sub-family, and the success coverage of the joint history after splitting should not be less than that before splitting; Merge requires that the cost of bidirectional obfuscation or maintenance be reduced after the merger, while retaining the joint unique coverage of the two original skills; Prune requires that after deleting constraints or experiences, no unique validation sample should change from pass to fail, and deleted content and restored references should be retained; Retirement requires that the retired skill has no irreplaceable validation coverage and that its representative sample can be correctly processed by the remaining skill or general path. Defer requires that the current skill set should not be changed when the number of supported semantic families, attribution confidence, number of replay samples, or statistical confidence is insufficient.

[0205] When all skills to be determined meet the skill launch conditions, the system will determine the adjusted execution link as the target execution link.

[0206] Step 404: Generate the target execution result corresponding to the intent to be processed based on the target execution chain.

[0207] Step 406: Return the target execution result to the user.

[0208] Based on the two independent control planes described above, production failures are transformed into controllable and traceable candidate changes without updating the base model weights. Admission is determined by replay evidence related to these candidates. The proposal control plane generates candidate changes that have not yet taken effect based on failure and cumulative library health evidence. The admission control plane does not rely solely on the repair results that triggered the failure for evidence; instead, it compares candidate versions with the current version across replay partitions with separate functions such as target, unseen, historical, adjacent, general, and coverage, deciding whether to accept, reject, or postpone. Through versioning, isolated replays, hard-constraint admission, atomic commits, and rollbacks, the stability, controllability, and traceability of the online skill library are ensured, thereby balancing repair effectiveness, historical stability, cross-skill interference control, testing costs, and version recoverability.

[0209] Corresponding to the above method embodiments, this specification also provides an information processing method applied to cloud-side equipment, including: The receiving end-side device sends a pending request, parses the pending request, obtains the pending intent and at least one reference skill corresponding to the pending intent, and determines the initial execution link corresponding to the pending intent based on at least one reference skill. Generate a reference execution result corresponding to the intent to be processed based on the initial execution link, and record the link attribute information corresponding to the initial execution link; If the reference execution result does not meet the preset verification conditions, the target failure type is determined based on the link attribute information; Based on the target failure type, the initial execution chain is adjusted to generate the target execution chain, and the target execution result corresponding to the intent to be processed is generated based on the target execution chain; The target execution result is sent to the end device.

[0210] Corresponding to the above method embodiments, this specification also provides embodiments of an information processing apparatus. Figure 5 A schematic diagram of the structure of an information processing apparatus according to one embodiment of this specification is shown. Figure 5 As shown, the device includes: The acquisition unit 502 is configured to acquire the intent to be processed and at least one reference skill corresponding to the intent to be processed, and determine the initial execution link corresponding to the intent to be processed based on each reference skill. The recording unit 504 is configured to generate a reference execution result corresponding to the intent to be processed based on the initial execution link, and to record the link attribute information corresponding to the initial execution link; The determining unit 506 is configured to determine the target failure type based on the link attribute information when the reference execution result does not meet the preset verification conditions. The generation unit 508 is configured to adjust the initial execution link according to the target failure type, generate the target execution link, and generate the target execution result corresponding to the intent to be processed based on the target execution link.

[0211] Furthermore, the acquisition unit 502 is further configured as follows: Obtain the current intent skill library, determine the intent to be processed in the current intent skill library, and determine at least one reference skill corresponding to the intent to be processed, wherein the current intent skill library is constructed according to the correspondence between intents and intent skills; or, In response to the detection of a target query request, the intent to be processed is determined based on the target query request, and at least one reference skill corresponding to the intent to be processed is determined.

[0212] Furthermore, the current intent skill library is constructed in the following manner: Obtain multiple historical query requests and at least one intent skill corresponding to each historical query request; Based on the request semantics of each historical query request, generate at least one intent cluster; Based on at least one intent skill corresponding to each historical query request in each intent cluster, construct the correspondence between intents and intent skills; Generate the current intention skill library based on the correspondence between intentions and intention skills.

[0213] Furthermore, the acquisition unit 502 is further configured as follows: Determine the intent semantics corresponding to the intent to be processed, and obtain the skill semantics corresponding to each intent skill in the current intent skill library; based on the intent semantics and the skill semantics corresponding to each intent skill, determine at least one initial reference skill corresponding to the intent to be processed, and determine the initial route score corresponding to each initial reference skill; determine the initial reference skill corresponding to the initial route score that satisfies the preset route threshold as the reference skill corresponding to the intent to be processed.

[0214] Furthermore, the generating unit 508 is further configured as follows: Based on the target failure type, at least one historical failure type is obtained within a preset historical query time period, wherein the semantic similarity between each historical failure type and the target failure type satisfies a similarity threshold. If the target failure type and each historical failure type meet the preset adjustment conditions, the target constraint conditions are determined based on the target failure type. The initial execution chain is adjusted according to the target constraints.

[0215] Furthermore, the generation unit 508 is also configured as follows: Based on the target failure type, identify the skills to be adjusted in the initial execution chain; Based on the skill to be adjusted, at least one candidate interference skill is determined in the skill interference graph, and the initial execution link is adjusted according to each candidate interference skill, wherein the skill interference graph is constructed based on the intention skill.

[0216] Furthermore, the skill interference graph is constructed in the following manner: In the current intent skill library, the intent skills corresponding to each intent are obtained, and the skill attribute information corresponding to each intent skill is determined. The current intent skill library is constructed based on the correspondence between intents and intent skills. Based on skill attribute information, determine the relationships between various intentional skills; The intention skills are defined as nodes, and the relationships between each intention skill are defined as edges between nodes, thus constructing a skill interference graph.

[0217] Furthermore, the generating unit 508 is further configured as follows: Based on the historical execution path of each intent skill, determine the bidirectional obfuscation parameters and historical regression impact parameters between each intent skill; Based on the intent corresponding to each intent skill and the skill semantics corresponding to each intent skill, determine the similarity parameters between each intent skill; The relationships between intention skills are determined based on the bidirectional confusion parameter, historical regression influence parameter, and similarity parameter among the intention skills.

[0218] Furthermore, the determining unit 506 is further configured as follows: Parse the link attribute information to determine the execution failure description information; Based on the execution failure description information, determine the target failure type corresponding to the execution failure description information from at least one failure type.

[0219] Furthermore, the generating unit 508 is further configured as follows: Based on the adjusted initial execution chain, at least one skill to be determined is identified; if each skill to be determined meets the skill launch conditions, a target execution chain is generated based on each skill to be determined.

[0220] Furthermore, the conditions for the skill to be available are determined in the following manner: Obtain the current skill performance information for each skill to be determined; Configure the isolation skills corresponding to each skill to be determined in the isolation environment, and determine the performance information of each isolation skill. Compare the isolated skill performance information with the current skill performance information to determine the performance comparison result; Based on the performance comparison results, the conditions for skill activation are determined.

[0221] Furthermore, the link attribute information includes at least one of the following: The current version information corresponding to the current intent skill library, wherein the current intent skill library is constructed according to the correspondence between intents and intent skills; the intent identifier corresponding to the intent to be processed; each reference skill and the routing score corresponding to each reference skill; the execution attribute information corresponding to the reference execution result; and the initial link quality index.

[0222] Furthermore, the determination unit 506 is also configured as follows: Determine whether the reference execution result meets the preset verification conditions; if the reference execution result meets the preset verification conditions, record the intent to be processed and each reference skill; update the current intent skill library according to the intent to be processed and each reference skill, wherein the current intent skill library is constructed according to the correspondence between intents and intent skills.

[0223] The above is an illustrative scheme of an information processing device according to this embodiment. It should be noted that the technical solution of this information processing device and the technical solution of the information processing method described above belong to the same concept. For details not described in detail in the technical solution of the information processing device, please refer to the description of the technical solution of the information processing method described above.

[0224] Figure 6 A structural block diagram of a computing device provided according to one embodiment of this specification is shown.

[0225] The computing device 600 includes: Memory 610 and processor 620; The memory 610 is used to store computer programs / instructions, and the processor 620 is used to execute the computer programs / instructions, which, when executed by the processor 620, implement the steps of the information processing method.

[0226] In one or more embodiments of this specification, the computing device can be understood as an integrated smart terminal, including but not limited to a server, desktop computer, PC (Personal Computer), all-in-one model machine, mobile phone, tablet computer or other portable smart terminal, etc., and the computing device may have the model described in the above embodiments of this application pre-installed.

[0227] Specifically, this computing device can pre-install various types of models, including but not limited to models in natural language processing, visual processing, speech processing, code processing, and multimodal task processing, thus providing diverse model selection. In different product forms, this computing device can support one or more model usage methods, including but not limited to model training, model invocation, model fine-tuning, model deployment, model inference, and application. In some product forms, this computing device also supports model management, including but not limited to multi-type model management (supporting the management of discriminative, generative, and other model types), model version control (supporting the control of different model versions), and model evaluation (evaluating model performance and effectiveness based on model evaluation tools). In other product forms, this computing device can also create applications based on models, providing API (Application Programming Interface) calling capabilities. Users can call models into created applications through the API interface, and application management tools are also provided to manage and monitor the applications.

[0228] Furthermore, the computing device can also include data management (supporting the creation and management of model tuning datasets), a training center (providing abundant training resources to help users learn and master AI (Artificial Intelligence) technology), and basic control capabilities (providing enterprise-level basic control capabilities to ensure the security and efficient operation of the system). Through the above functions, it provides a comprehensive and integrated device for AI development, training, deployment, and application.

[0229] Figure 7 A structural block diagram of an electronic device provided according to one embodiment of this specification is shown.

[0230] A memory 710 and a processor 720 are connected via a bus 730; The memory 710 is used to store computer programs / instructions, and the processor 720 is used to execute the computer programs / instructions, which, when executed by the processor 720, implement the steps of the method.

[0231] Specifically, the components of the electronic device 700 include, but are not limited to, a memory 710 and a processor 720. The processor 720 is connected to the memory 710 via a bus 730, and the database 750 is used to store data.

[0232] Electronic device 700 also includes access device 740, which enables electronic device 700 to communicate via one or more networks 760. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. Access device 740 may include one or more of any type of wired or wireless network interface (e.g., network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, Wi-MAX (Worldwide Interoperability for Microwave Access) interface, Ethernet interface, Universal Serial Bus (USB) interface, cellular network interface, Bluetooth interface, Near Field Communication (NFC) interface, and so on.

[0233] In one embodiment of this specification, the above-described components of the electronic device 700 and Figure 7 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 7 The block diagram of the electronic device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0234] Electronic device 700 can be any type of stationary or mobile electronic device, including mobile computers or mobile electronic devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable electronic devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary electronic devices such as desktop computers or personal computers (PCs). Electronic device 700 can also be a mobile or stationary server.

[0235] The above is an illustrative scheme of an electronic device according to this embodiment. It should be noted that the technical solution of this electronic device and the technical solution of the information processing method described above belong to the same concept. For details not described in detail in the technical solution of the electronic device, please refer to the description of the technical solution of the information processing method described above.

[0236] An embodiment of this specification also provides a computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the information processing method described above.

[0237] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium and the technical solution of the information processing method described above belong to the same concept. For details not described in detail in the technical solution of the storage medium, please refer to the description of the technical solution of the information processing method described above.

[0238] An embodiment of this specification also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described information processing method.

[0239] The above is an illustrative scheme of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product and the technical solution of the information processing method described above belong to the same concept. For details not described in detail in the technical solution of the computer program product, please refer to the description of the technical solution of the information processing method described above.

[0240] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0241] The computer program / instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0242] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0243] In the above embodiments, 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.

[0244] The preferred embodiments disclosed above are merely illustrative of this specification. Optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. An information processing method, comprising: Obtain the intent to be processed and at least one reference skill corresponding to the intent to be processed, and determine the initial execution link corresponding to the intent to be processed based on at least one reference skill; Generate a reference execution result corresponding to the intent to be processed based on the initial execution link, and record the link attribute information corresponding to the initial execution link; If the reference execution result does not meet the preset verification conditions, the target failure type is determined based on the link attribute information; Based on the target failure type, the initial execution chain is adjusted to generate the target execution chain, and the target execution result corresponding to the intent to be processed is generated based on the target execution chain.

2. The method as described in claim 1, wherein obtaining the intent to be processed and at least one reference skill corresponding to the intent to be processed includes: Obtain the current intent skill library, determine the intent to be processed in the current intent skill library, and determine at least one reference skill corresponding to the intent to be processed, wherein the current intent skill library is constructed according to the correspondence between intents and intent skills; or, In response to the detection of a target query request, the intent to be processed is determined based on the target query request, and at least one reference skill corresponding to the intent to be processed is determined.

3. The method of claim 2, wherein the current intent skill library is constructed in the following manner: Obtain multiple historical query requests and at least one intent skill corresponding to each historical query request; Based on the request semantics of each historical query request, generate at least one intent cluster; Based on at least one intent skill corresponding to each historical query request in each intent cluster, construct the correspondence between intents and intent skills; Generate the current intention skill library based on the correspondence between intentions and intention skills.

4. The method of claim 2, wherein determining at least one reference skill corresponding to the intent to be processed includes: Determine the intent semantics corresponding to the intent to be processed, and obtain the skill semantics corresponding to each intent skill in the current intent skill library; Based on the intent semantics and the skill semantics corresponding to each intent skill, at least one initial reference skill corresponding to the intent to be processed is determined, and the initial routing score corresponding to each initial reference skill is determined. The initial reference skill corresponding to the initial route score that meets the preset route threshold is determined as the reference skill corresponding to the intent to be processed.

5. The method of claim 1, wherein adjusting the initial execution chain according to the target failure type includes: Based on the target failure type, at least one historical failure type is obtained within a preset historical query time period, wherein the semantic similarity between each historical failure type and the target failure type satisfies a similarity threshold. If the target failure type and each historical failure type meet the preset adjustment conditions, the target constraint conditions are determined based on the target failure type. The initial execution chain is adjusted according to the target constraints.

6. The method of claim 1, further comprising: Based on the target failure type, identify the skills to be adjusted in the initial execution chain; Based on the skill to be adjusted, at least one candidate interference skill is determined in the skill interference graph, and the initial execution link is adjusted according to each candidate interference skill, wherein the skill interference graph is constructed based on the intention skill.

7. The method of claim 6, wherein the skill interference graph is constructed in the following manner: In the current intent skill library, retrieve the intent skills corresponding to each intent, and determine the skill attribute information corresponding to each intent skill. The current intent skill library is constructed based on the correspondence between intents and intent skills; Based on skill attribute information, determine the relationships between various intentional skills; The intention skills are defined as nodes, and the relationships between each intention skill are defined as edges between nodes, thus constructing a skill interference graph.

8. The method as described in claim 7, wherein determining the association between each intentional skill based on skill attribute information includes: Based on the historical execution path of each intent skill, determine the bidirectional obfuscation parameters and historical regression impact parameters between each intent skill; Based on the intent corresponding to each intent skill and the skill semantics corresponding to each intent skill, determine the similarity parameters between each intent skill; The relationships between intention skills are determined based on the bidirectional confusion parameter, historical regression influence parameter, and similarity parameter among the intention skills.

9. The method as described in claim 1, wherein determining the target failure type based on the link attribute information includes: Parse the link attribute information to determine the execution failure description information; Based on the execution failure description information, determine the target failure type corresponding to the execution failure description information from at least one failure type.

10. The method of claim 1, generating the target execution chain, comprising: Based on the adjusted initial execution chain, at least one skill to be determined is identified; If each skill to be determined meets the skill launch conditions, a target execution link is generated based on each skill to be determined.

11. The method of claim 10, wherein the skill activation conditions are determined in the following manner: Obtain the current skill performance information for each skill to be determined; Configure the isolation skills corresponding to each skill to be determined in the isolation environment, and determine the performance information of each isolation skill. Compare the isolated skill performance information with the current skill performance information to determine the performance comparison result; Based on the performance comparison results, the conditions for skill activation are determined.

12. The method of claim 1, wherein the link attribute information includes at least one of the following: The current version information corresponding to the current intent skill library, among which, The current intent skill library is constructed based on the correspondence between intents and intent skills; The intent identifier corresponding to the intent to be processed; Each reference skill and its corresponding routing score; The execution attribute information corresponding to the reference execution result; Initial link quality metrics.

13. The method of any one of claims 1 to 12, further comprising: Determine whether the reference execution result meets the preset verification conditions; If the reference execution result meets the preset verification conditions, the intent to be processed and each reference skill are recorded. The current intent skill library is updated based on the intent to be processed and each reference skill, wherein the current intent skill library is constructed based on the correspondence between intents and intent skills.

14. An information processing apparatus, comprising: The acquisition unit is configured to acquire the intent to be processed and at least one reference skill corresponding to the intent to be processed, and determine the initial execution link corresponding to the intent to be processed based on the at least one reference skill. The recording unit is configured to generate a reference execution result corresponding to the intent to be processed based on the initial execution link, and to record the link attribute information corresponding to the initial execution link; The determining unit is configured to determine the target failure type based on the link attribute information when the reference execution result does not meet the preset verification conditions. The generation unit is configured to adjust the initial execution link according to the target failure type, generate the target execution link, and generate the target execution result corresponding to the intent to be processed based on the target execution link.

15. An information processing method applied to cloud-side equipment, comprising: The receiving end-side device sends a pending request, parses the pending request, obtains the pending intent and at least one reference skill corresponding to the pending intent, and determines the initial execution link corresponding to the pending intent based on at least one reference skill. Generate a reference execution result corresponding to the intent to be processed based on the initial execution link, and record the link attribute information corresponding to the initial execution link; If the reference execution result does not meet the preset verification conditions, the target failure type is determined based on the link attribute information; Based on the target failure type, the initial execution chain is adjusted to generate the target execution chain, and the target execution result corresponding to the intent to be processed is generated based on the target execution chain; The target execution result is sent to the end device.

16. A computing device, comprising: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 13.

17. An electronic device comprising: A memory and a processor, the memory and the processor being connected via a bus; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 13.

18. A computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 13.

19. A computer program product comprising a computer program / instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 13.

Images