AI service hosting method and device, computer equipment and storage medium

Abstract

The invention discloses an AI service hosting method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring various questions of a user; performing semantic purification and disambiguation on the question, identifying a generic intention and a key entity of the question, and generating a structured abstract; based on a single-round or multi-round dialogue history, accurately understanding and processing problems by using a large model technology to obtain a user intention; in combination with the structured abstract, refining the user intention, and outputting an intention label with confidence, a clear question expression and an entity list to obtain a processing result; processing results are further subdivided into specific secondary business intentions, and user demands are accurately positioned to obtain specific intentions; and automatically calling a corresponding interface to generate an answer according to a specific intention, and presenting the answer in a structured and visual manner. By implementing the method provided by the invention, the defects of the existing system can be effectively solved, and higher automation level, better user experience and stronger adaptability and expandability are provided.

Description

Technical Field

[0001] This invention relates to computers, and more specifically to methods, apparatus, computer equipment, and storage media for hosting AI services. Background Technology

[0002] Asset custody, as a crucial component of the modern financial system, plays a vital role in connecting the client and the trustee. Banks, as professional custodians, assume multiple responsibilities in this process, including but not limited to secure safekeeping of entrusted assets, clearing and settlement, asset supervision, financial accounting, investment regulation, and information disclosure. By providing these comprehensive services, asset custody not only opens up new sources of non-interest income for banks but also significantly enhances their overall service capabilities, expands their customer base, and strengthens their brand image and reputation in the financial market. To support these complex business needs, banks typically establish dedicated custody service platforms. These platforms aim to ensure the security of client assets while providing transparency, allowing clients to understand their asset status and disclosing necessary information to relevant regulatory agencies. An efficient and reliable custody service platform is fundamental to achieving these goals; it must be able to handle large volumes of transaction data, support various types of financial services, and guarantee a high level of information security.

[0003] Currently, custody service platforms in the banking industry are mainly divided into two categories: traditional service platforms and AI-assisted service platforms. Traditional custody service platforms typically rely on traditional front-end and back-end architectures and middleware technologies to achieve basic functions such as custody account management, risk management, compliance management, and customer service. While this approach is stable and reliable, its automation level is low, and a large number of operations still require manual intervention. This can lead to efficiency bottlenecks when handling large-scale data or complex transactions. Furthermore, due to the lack of flexibility, users need to access different custody services through different entry points, resulting in a poor user experience and high learning costs. More importantly, this model struggles to provide personalized services based on specific user needs and cannot quickly respond to changes in market and customer demands. In recent years, with the development of artificial intelligence technology, some banks have begun to adopt AI-enabled custody service platforms. These platforms utilize tools such as question-answering robots to attempt to improve the ability to respond quickly to specific user needs and help users find the services they require more quickly. However, existing AI solutions have certain limitations, such as high implementation costs and technical complexity. Especially when faced with vague user questions or questions lacking key information, existing systems often struggle to accurately understand the user's intent and may even provide misleading answers, further impacting service efficiency and user experience.

[0004] Therefore, it is necessary to design a new approach to effectively address the shortcomings of existing systems and provide higher levels of automation, better user experience, and greater adaptability and scalability. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, apparatus, computer equipment and storage medium for hosting AI services.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a managed AI service method, comprising:

[0007] Obtain various questions from users;

[0008] The problem is semantically cleaned and disambiguated, and its general intent and key entities are identified to generate a structured summary;

[0009] Based on single-turn or multi-turn dialogue history, large model technology is used to accurately understand and process the question in order to obtain user intent;

[0010] The user intent is refined by combining the structured summary, and the intent tags with confidence levels, clear question statements and entity lists are output to obtain the processing results;

[0011] The processing results are further subdivided into specific secondary business intents to accurately pinpoint user needs and obtain specific intents;

[0012] The system automatically invokes the corresponding interface to generate a response based on the specific intent, and presents it in a structured and visual manner.

[0013] Its further technical solution is: the acquisition of various user questions includes:

[0014] Users can submit questions through customized Tools card services or ask questions using voice recognition. Based on big data modeling technology, real-time semantic analysis is performed, and historical dialogue records are automatically linked to provide suggested questions to obtain various types of questions from users.

[0015] The further technical solution is as follows: The problem is semantically cleaned and disambiguated, and its general intent and key entities are identified to generate a structured summary, including:

[0016] The Qwen 72B large model was used to remove redundant information and sentiment words from the question, and the ambiguity was resolved based on the context to extract the core question expression in order to obtain the purified question text.

[0017] Based on the purified question text, the business scope to which the user's question belongs is identified and classified;

[0018] Key business entities and action elements were identified from the purified problem text;

[0019] The business scope, key business entities, and action elements are integrated to obtain a structured semantic summary that includes the core of the problem, general intent, and key elements.

[0020] Its further technical solution is as follows: Based on single-turn or multi-turn dialogue history, the large model technique is used to accurately understand and process the question in order to obtain the user intent, including:

[0021] Analyze the problem and determine whether it is a single-turn or multi-turn dialogue type through structured semantic summarization;

[0022] When the problem is a multi-turn dialogue type, the RoPE technique is used to maintain the dialogue order and sense of distance by combining the recent dialogue history. The self-attention and cross-attention mechanism of the large model is used to evaluate the semantic association between the current and historical dialogue words to achieve referential resolution. When encountering situations where the intent jumps or is ambiguous, a question is generated based on the existing information to clarify the user's intent in order to obtain the user's intent.

[0023] The further technical solution is as follows: The user intent is refined by combining the structured summary, and intent tags with confidence levels, clear question statements, and entity lists are output to obtain the processing results, including:

[0024] The user intent is compared with a predefined business intent vector to identify and determine the generic intent category that best meets the user's needs, thus obtaining the generic intent.

[0025] By combining the structured summary and user history behavior data, the general intent is adjusted to specific user needs, and a structured processing result is generated to obtain the processing outcome.

[0026] The further technical solution is as follows: combining the structured summary and user historical behavior data, adjusting the general intent to specific user needs, and generating a structured processing result to obtain the processing result, including:

[0027] By combining contextual information from single-turn and multi-turn dialogues, the general intent is integrated;

[0028] The general intent is processed using a large model to clarify the user's core needs;

[0029] By combining the structured summary, the consistency between the extracted entity information and the context is compared and verified, and the general intent is adjusted to a specific and precise intent.

[0030] Taking into account the influence of users' historical behavior patterns, the final user intent is output through weighted analysis to obtain the processing results.

[0031] The further technical solution is as follows: The processing result is further subdivided into specific secondary business intents to accurately locate user needs and obtain specific intents, including:

[0032] The processing results are reviewed, and the corresponding general intent is matched with the corresponding secondary intent. Natural language processing technology is used to conduct in-depth analysis and determine the specific secondary intent that meets the user's needs.

[0033] Based on the specific secondary intent, a structured expression containing specific business requirement types and necessary parameters is created to obtain the specific intent.

[0034] The present invention also provides a device for hosting AI services, comprising:

[0035] The problem acquisition unit is used to acquire various problems from users;

[0036] The preprocessing unit is used to perform semantic cleansing and disambiguation on the problem, identify its general intent and key entities, and generate a structured summary.

[0037] The intent processing unit is used to accurately understand and process the question based on the single-turn or multi-turn dialogue history using large model technology to obtain the user's intent;

[0038] The refinement unit is used to refine the user intent in conjunction with the structured summary, and output intent tags with confidence levels, clear question statements and entity lists to obtain the processing results;

[0039] The positioning unit is used to further subdivide the processing result into specific secondary business intents, accurately locate user needs, and obtain specific intents.

[0040] The generation unit is used to automatically call the corresponding interface to generate a response based on the specific intent, and present it in a structured and visual manner.

[0041] The present invention also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the above-described method.

[0042] The present invention also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0043] The advantages of this invention compared to existing technologies are as follows: By introducing advanced intelligent agents and large model technology, this invention achieves a fully automated processing mechanism from acquiring user questions to automatically calling the corresponding interfaces to generate answers. First, the user's question undergoes semantic cleansing and disambiguation, identifying its general intent and key entities to generate a structured summary. Next, based on single-turn or multi-turn dialogue history, large model technology is used to accurately understand and refine the user's intent, outputting intent labels with confidence levels, clear question statements, and entity lists as the processing result. Subsequently, this result is further subdivided into specific secondary business intents to accurately pinpoint user needs. Finally, based on the specific intent, the corresponding interfaces are automatically called to generate highly targeted answers, which are presented to the user in a structured and visual manner. This process not only significantly improves the system's automation level and reduces the need for manual intervention but also greatly optimizes the user experience through intelligent front-end interaction, while enhancing the system's adaptability and scalability in the face of complex and ever-changing business scenarios. This effectively solves the shortcomings of existing systems, such as low automation, poor user experience, and insufficient flexibility.

[0044] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a schematic diagram illustrating an application scenario of the managed AI service method provided in an embodiment of the present invention;

[0047] Figure 2 This is a flowchart illustrating the managed AI service method provided in an embodiment of the present invention.

[0048] Figure 3 A schematic diagram of a sub-process of the managed AI service method provided in an embodiment of the present invention;

[0049] Figure 4 A schematic diagram of a sub-process of the managed AI service method provided in an embodiment of the present invention;

[0050] Figure 5 A schematic diagram of a sub-process of the managed AI service method provided in an embodiment of the present invention;

[0051] Figure 6 A schematic diagram of a sub-process of the managed AI service method provided in an embodiment of the present invention;

[0052] Figure 7 A schematic diagram of a sub-process of the managed AI service method provided in an embodiment of the present invention;

[0053] Figure 8 A schematic block diagram of a managed AI service device provided in an embodiment of the present invention;

[0054] Figure 9 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0057] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0058] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0059] Please see Figure 1 and Figure 2 , Figure 1 This is a schematic diagram illustrating an application scenario of the managed AI service method provided in an embodiment of the present invention. Figure 2This is a schematic flowchart illustrating the managed AI service method provided in this embodiment of the invention. The managed AI service method is applied to a server. The server interacts with the terminal, and through the integration of large model technology and multi-turn dialogue history analysis, it achieves fully automated processing from user question acquisition, semantic cleansing and disambiguation, intent recognition to precise business matching. This method can not only automatically eliminate redundant information and emotional words in user questions, but also accurately understand and refine the user's general intent to specific secondary business needs based on context, thereby providing intent labels with confidence, clear question statements, and entity lists as processing results. By real-time association with historical dialogue records, utilizing RoPE technology to maintain dialogue order and distance, and combining user behavior data for personalized adjustments, the system greatly improves its ability to accurately locate user needs, ensuring the accuracy and relevance of answers. Furthermore, the structured and visual answer presentation further enhances the user experience, giving the system higher adaptability and scalability, effectively solving the shortcomings of existing systems in terms of semantic understanding depth, user intent recognition accuracy, and interaction naturalness.

[0060] Figure 2 This is a flowchart illustrating the managed AI service method provided in an embodiment of the present invention. Figure 2 As shown, the method includes the following steps S110 to S160.

[0061] S110, Obtain various questions from users.

[0062] In this embodiment, a "problem" refers to a specific inquiry or request submitted by a user through a customized Tools card service or speech recognition function. This is processed through real-time semantic analysis using big data modeling technology, automatically linking it to historical dialogue records, and generating structured information including a refined problem statement, high-level general intent, and key entities. In short, a problem refers to a specific business request or question that needs to be answered, obtained by the system after intelligently processing user input.

[0063] Specifically, users can submit questions through customized Tools card services or ask questions using voice recognition. Based on big data modeling technology, real-time semantic analysis is performed to automatically link historical dialogue records and provide suggested questions to obtain various types of questions from users.

[0064] To cater to different user preferences and usage scenarios, the system provides multiple access points, including PC and mobile devices (such as WeChat official accounts). This includes not only traditional text input methods but also integrated voice recognition, allowing users to ask questions via voice, greatly improving the convenience and flexibility of question submission.

[0065] The PC client of the managed assistant features a dedicated intelligent Q&A entry point and integrates a customized Tools card service system. These cards have preset common question templates based on different business scenarios, allowing users to quickly ask questions by directly selecting the relevant card without manual input.

[0066] WeChat Official Accounts utilize advanced voice recognition technology, enabling users to express their questions or needs in a more natural way—that is, by speaking. This design is particularly suitable for users in mobile environments, allowing them to easily initiate conversations anytime, anywhere.

[0067] Once a user submits a question, the system immediately employs powerful large-scale model technology to perform real-time semantic analysis on the input. The large-scale model mentioned here refers to Qwen2-72B, which boasts exceptional natural language understanding and processing capabilities.

[0068] The system can not only understand the information directly expressed by the user, but also automatically connect it to the user's past conversation history to provide contextual support for the current question. This means that even questions that are fragmented or lack complete background information can be correctly understood and responded to.

[0069] Based on the above analysis, the system will dynamically generate several possible suggested questions at the bottom of the chat interface. These suggestions are based on the user's historical behavior patterns and the content of the current conversation, helping users to quickly find the issues they truly care about, while also reducing misunderstandings caused by unclear expressions.

[0070] This interaction optimization mechanism significantly improves the user experience because it reduces the time cost for users to find answers while increasing satisfaction during the interaction process.

[0071] In summary, the S110 goes beyond simply collecting user input. By integrating multi-channel access, intelligent front-end interaction design, and back-end processing based on advanced AI technology, it constructs a comprehensive and efficient user question acquisition system. This approach effectively solves the problems of slow response speed and low understanding accuracy in traditional customer service systems, thus laying a solid foundation for subsequent intent recognition, question processing, and answer generation.

[0072] S120. Perform semantic cleansing and disambiguation on the problem, identify its general intent and key entities, and generate a structured summary.

[0073] In this embodiment, structured summary refers to the structured data generated by semantically cleansing and disambiguating user queries to identify general intents, key business entities, and action elements. This summary contains the core question statement, the relevant business scope, and information about key entities. This summary provides clear guidance for subsequent task processing, ensuring that the system can accurately understand and respond to user requests.

[0074] In one embodiment, please refer to Figure 3 The above-mentioned step S120 may include steps S121 to S124.

[0075] S121. Use the Qwen 72B large model to remove redundant information and sentiment words from the question, and resolve ambiguity based on context to extract the core question expression in order to obtain the purified question text.

[0076] In this embodiment, the purified question text refers to the result obtained after processing the original input question using large model technology. This includes removing non-critical information such as repetitive expressions, colloquial filler words, and emotional words, and resolving potential ambiguities based on the context of the current statement, thereby obtaining a more accurate and concise question expression.

[0077] S122. Based on the purified question text, identify and classify the business category to which the user's question belongs.

[0078] In this step, the system uses large-scale modeling capabilities to analyze the cleaned-up question text and identify the specific business area or category to which the question belongs (e.g., account inquiry, transaction operation, report retrieval, etc.). This step is fundamental to understanding user needs and helps to provide more accurate services subsequently.

[0079] S123. Identify key business entities and action elements from the purified problem text.

[0080] Key business entities refer to the specific business objects mentioned in the question, such as accounts, products, and amounts; while action elements refer to the operations that users want to perform, such as querying, modifying, or adding. This step aims to extract the core content of the question for subsequent processing.

[0081] S124. Integrate the business scope, key business entities, and action elements to obtain a structured semantic summary containing the core of the problem, general intent, and key elements.

[0082] Finally, the system integrates all the above information to form a structured summary. This summary not only includes the core description of the problem but also clarifies the business scope to which the problem belongs, as well as the key business entities and action elements involved. Such a structured summary provides an important basis for subsequent task routing and decision-making, enabling the system to respond to user requests more intelligently and efficiently.

[0083] In this embodiment, after obtaining the user's initial question in step S120, in order to ensure that the customer service system can accurately understand and respond to the user's needs, and at the same time avoid the problem of historical information pollution caused by multi-turn dialogue, this solution innovatively adopts a mode that combines single-turn and multi-turn dialogue to process questions in parallel.

[0084] Leveraging the powerful semantic understanding capabilities of the Qwen 72B large model, it automatically identifies and removes non-critical information such as repetitive expressions, sentiment words, and colloquial filler words from the current question. Furthermore, it resolves ambiguous words and phrases based on context, extracting a concise core question statement. This eliminates redundant information, sentiment words, and colloquial filler words from user input, resolving potential ambiguities.

[0085] Based on the purified text, the specific semantic extraction and understanding capabilities of the large model are used to identify the specific business domain or category to which the question belongs (e.g., account inquiry, transaction operation, report retrieval, etc.). It is important to note that this step is strictly limited to a single-turn context and only performs directional classification to avoid the influence of multi-turn dialogue history on the current question. Quickly determining the business category to which the user's question belongs lays the foundation for subsequent task routing.

[0086] Extract specific key business entities (such as account, product, amount, instruction number, date, etc.) and the actions the user wishes to perform (such as query, modify, add, etc.) from the cleaned question text. These key elements are crucial for understanding user needs. Identify the key business entities and action elements in the question.

[0087] Through the above three steps—semantic sanitization and disambiguation, single-round general intent recognition, and core element extraction—the system ultimately generates a structured semantic summary containing the following content:

[0088] Refined problem statement: The core description of the problem obtained after purification and disambiguation.

[0089] High-level general intent: The directional classification result of the business scope to which the problem belongs.

[0090] Key entities and actions: The specific business objects extracted from the problem and the expected operations to be performed.

[0091] This structured summary not only clearly reflects the user's actual needs but also provides a precise task input package for subsequent multi-agent decision trees. It ensures that each task is correctly assigned to the most suitable agent for in-depth processing, thereby achieving efficient and accurate service responses.

[0092] By combining single-turn and multi-turn conversations, this solution effectively avoids the interference of historical information caused by multi-turn conversations while maintaining the simplicity and accuracy of the dialogue. This approach significantly improves the understanding and response efficiency of the customer service system, providing users with a more intelligent, flexible, and efficient managed AI assistant service.

[0093] S130. Based on the history of single-turn or multi-turn dialogue, use large model technology to accurately understand and process the question in order to obtain the user's intent.

[0094] In this embodiment, user intent refers to identifying and understanding the user's true needs by analyzing the user's questions, combining contextual information (including single-turn and multi-turn dialogue history), and using advanced big data modeling techniques. This process is not limited to simple question-and-answer sessions, but also includes in-depth analysis and clarification of user intent, as well as precise matching to ensure that the services provided are accurate.

[0095] In one embodiment, please refer to Figure 4 The above-mentioned step S130 may include steps S131 to S132.

[0096] S131. Analyze the problem and determine whether the problem is a single-turn or multi-turn dialogue type through structured semantic summarization.

[0097] This step first requires a preliminary analysis of the user's input question, and based on the structured semantic summary generated in the preprocessing stage mentioned earlier, determines whether the current question is a single query or a complex query requiring multiple turns of dialogue. The key to this step is to quickly and accurately classify the question so that appropriate processing strategies can be adopted subsequently. Specifically:

[0098] Single query: refers to questions that can be understood and answered independently without relying on preceding or following context. For example, account balance inquiry.

[0099] Multi-turn dialogue types: These involve more complex interaction scenarios and may require combining previous conversations to fully understand the user's needs. For example, a detailed inquiry about a product may require multiple rounds of questioning to obtain complete information.

[0100] S132. When the question is a multi-turn dialogue type, the RoPE technique is used to maintain the dialogue order and sense of distance by combining the recent dialogue history. The self-attention and cross-attention mechanism of the large model is used to evaluate the semantic association between the current and historical dialogue words to achieve referential resolution. When encountering situations where the intent jumps or is ambiguous, questions are generated based on existing information to clarify the user's intent in order to obtain the user's intent.

[0101] For queries identified as multi-turn dialogues, the system will take the following measures for further processing:

[0102] Load recent conversation history: To maintain conversation continuity, the system will automatically load the user's most recent conversations as a reference background.

[0103] Using RoPE technology to maintain dialogue order and distance perception: Relative Position Encoding (RoPE) technology enables models to distinguish the position of each token in its sequence, thereby helping models understand the historical order of dialogue and the time intervals, which is crucial for maintaining the consistency and logic of long-term dialogue.

[0104] Application of self-attention and cross-attention mechanisms: Through these two mechanisms, the model can calculate the semantic similarity and relevance between words in the current query and words in the historical dialogue, thereby achieving effective referential resolution (e.g., the specific product name that "it" refers to). This helps improve the accuracy of understanding user intent.

[0105] Intent Clarification: If the system detects a jump in user intent (i.e., a sudden shift from one topic to another unrelated one) or ambiguity, it will automatically generate clarifying questions based on existing dialogue information to guide the user to provide more details or clarify their intent. This method effectively reduces misunderstandings and ensures that the final obtained user intent is clear and accurate.

[0106] Through the aforementioned sub-steps S131 and S132, the system can handle various user inquiries more intelligently, providing service responses that meet user expectations, whether it's a simple single query or a complex multi-turn interaction, greatly improving user experience and service efficiency. Furthermore, this approach demonstrates how modern AI technologies such as large models and agent architecture can be used to optimize the limitations of traditional customer service systems, achieving an overall improvement in both automation and service quality.

[0107] In this embodiment, the question processing module primarily employs a multi-turn conversation-based approach to further deepen the understanding of user queries. This process relies not only on the current user input but also on the historical records of previous conversations to ensure the accuracy and completeness of the understanding. First, based on the structured semantic summary generated in the preprocessing stage, the system categorizes the user's query. This is divided into two types:

[0108] Single query: Questions that can be understood and answered independently without referring to previous conversation history.

[0109] Context-enhanced queries: Questions that require a comprehensive understanding by considering the content of recent rounds of dialogue.

[0110] For single queries, the system proceeds directly to the next stage of problem confirmation. However, for context-enhanced queries, a more in-depth problem-solving process is required.

[0111] For queries identified as context-enhanced, the system will perform the following operations:

[0112] Load conversation history: The system automatically loads the user's most recent conversation records as background information for the current query.

[0113] Utilizing RoPE technology: Large models can automatically tag the positional information of each token in a sequence using Relative Position Encoding (RoPE). This means the model can distinguish the order and distance relationship between historical dialogues and the current query, thus gaining a better understanding of the entire dialogue flow.

[0114] Self-attention and cross-attention mechanisms: These mechanisms allow the model to calculate the semantic association weights between words in the current query and words in the historical dialogue. This helps to resolve referential inconsistencies, such as parsing words like "it" or "according to the preceding text," clarifying their specific references in the preceding context, and improving the accuracy of understanding user intent.

[0115] In actual conversations, users may exhibit jumps in intent (i.e., suddenly veer towards unrelated topics) or express themselves vaguely. To address this, the system takes the following measures:

[0116] Contextual Clarification: When the system detects that user input contains ambiguous or disjointed intent, it will automatically generate clarifying questions based on the current result and historical dialogue information. These questions aim to guide the user to provide more details or clarify their true intent, avoiding misunderstandings and misanswers.

[0117] Generating clarifying questions: If the current input is not clear enough or deviates significantly from previous discussions, the system will ask specific follow-up questions, requiring the user to provide more information or explain their exact needs. This step ensures that the final user intent is clear and accurate.

[0118] Through the steps described above, the system can maintain a high degree of accuracy and consistency in handling complex, multi-turn dialogues. This approach not only improves the accuracy of understanding user intent but also effectively addresses various complex interaction scenarios, including but not limited to issues such as intent jumps and ambiguous expressions. Furthermore, by leveraging modern AI technologies such as large models and intelligent agent architectures, traditional customer service systems can be optimized, achieving an overall improvement in both automation and service quality, significantly enhancing user experience and service efficiency.

[0119] S140. Refine the user intent based on the structured summary, and output intent tags with confidence levels, clear question statements, and entity lists to obtain the processing results.

[0120] In this embodiment, the processing result refers to adjusting the general intent to specific user needs by combining structured summaries and user historical behavior data, and outputting a final service response with a confidence level intent label, a clear question statement, and an entity list.

[0121] In one embodiment, please refer to Figure 5 The above-mentioned step S140 may include steps S141 to S142.

[0122] S141. Compare the user intent with the predefined business intent vector, identify and determine the generic intent category that best meets the user's needs, so as to obtain the generic intent.

[0123] In this embodiment, general intent refers to a preliminary intent judgment that identifies the general area of ​​user needs by performing high-level abstract classification of user questions without relying on specific business details.

[0124] A series of intent categories covering major business areas (such as balance inquiry, product information inquiry, etc.) are predefined, and a corresponding embedding vector is constructed for each intent.

[0125] The user's query is converted into an embedded vector, and then compared with the aforementioned set of intent vectors for similarity.

[0126] Based on similarity scores, the top few most matching intents are recalled as candidates, and then a large model is used to further filter out the most suitable general intent category for the current query.

[0127] S142. Combining the structured summary and user historical behavior data, adjust the general intent to specific user needs, and generate a structured processing result to obtain the processing result.

[0128] In one embodiment, please refer to Figure 6 The above step S142 may include steps S1421 to S1424.

[0129] S1421. Integrate the general intent by combining the contextual information from single-turn and multi-turn dialogues;

[0130] S1422. Process the general intent using a large model to clarify the user's core needs.

[0131] S1423. Based on the structured summary, compare and verify the consistency between the extracted entity information and the context, and adjust the general intent to a specific precise intent.

[0132] S1424. Considering the influence of users' historical behavior patterns, output the final user intent through weighted analysis to obtain the processing result.

[0133] By leveraging all available contextual information in single-turn and multi-turn dialogues, the initially identified general intent is integrated to ensure that it covers all aspects of the user's question.

[0134] Large-scale models conduct in-depth analysis of the integrated general intents to extract users' core needs, such as whether they want to check their account balance or want to obtain detailed information about a specific product.

[0135] Compare the key entities extracted from the question (such as account number, date, etc.) with the contextual information, and adjust the general intent as necessary to more accurately reflect the user's specific needs.

[0136] By considering users' historical behavioral habits, such as preferences and frequently used services, weighted analysis can be used to fine-tune the final user intent.

[0137] The output includes the final processing result, which includes intent labels (with confidence levels), a clear question statement, and a list of relevant entities.

[0138] This process enables managed AI assistants to understand user needs more intelligently, improving both interaction efficiency and user experience. Through refined intent recognition and personalized service delivery, this method significantly reduces the need for human intervention, improving overall business processing efficiency and service quality. Furthermore, this approach helps break free from the limitations of traditional customer service robots that rely on fixed logical paths, achieving more flexible and efficient customer service.

[0139] In this embodiment, step S140 involves initially classifying the user's question based on the output of the question processing stage and determining its corresponding generic intent category. The specific implementation steps are as follows:

[0140] First, a series of business intent categories are predefined, and a corresponding embedding vector is constructed for each intent.

[0141] Next, the user's query is also constructed as an embedded vector, and similarity calculation is performed with the intent vector set to recall the top 3 intent vectors.

[0142] Then, the top 3 intent vectors and the query are assembled, and intent recognition and selection are performed through a large model. The generic intent described in this method refers to a high-level abstract classification of user intent without relying on specific business details. Generic intents are mutually exclusive, meaning that each question can be assigned to one and only one type of generic intent, avoiding classification ambiguity. In this solution, generic intents mainly include four categories: balance-related, product and bond-related, transaction history-related, and instruction-related.

[0143] By combining the contextual information from single-turn and multi-turn conversations obtained in the previous steps, user intent is further integrated to improve the accuracy of intent recognition. Specifically, the question optimization phase first integrates the intents obtained from single-turn and multi-turn conversations, and, combined with the current dialogue information, uses a thought chain approach to guide the large model in analyzing the current intent, providing the final intent result, and returning a question matching the intent as input for subsequent stages. The specific steps are as follows:

[0144] The large model first anchors the basic intent of single-round or multi-round queries to confirm the user's core intent (e.g., the user wants to query an account, query various funds, etc.) and clarify their own goals.

[0145] Subsequently, entity consistency verification is performed, cross-validating the entity information extracted in a single round with the context information of multiple rounds of sessions. In case of conflict, the context information of multiple rounds of sessions is given priority.

[0146] Secondly, contextual intent correction is performed. The large model will integrate the features of multi-turn dialogues and refine the general intent into precise intent layer by layer.

[0147] Finally, the large model weights the user's historical behavior patterns and combines the user's historical behavior features as decision weights, ultimately outputting a three-element label with confidence (intent label, unambiguous explicit query, and entity list).

[0148] As can be seen from the detailed explanation of the two stages above, the general intent recognition stage is mainly responsible for the initial classification of user questions, while the question optimization stage further integrates information from single-turn and multi-turn conversations to ensure that the final intent recognition result is both accurate and meets the user's actual needs. This process not only improves the accuracy of intent recognition but also provides more reliable data support for subsequent services.

[0149] S150. Further subdivide the processing result into specific secondary business intents to accurately locate user needs and obtain specific intents.

[0150] In this embodiment, specific intent refers to further in-depth analysis and classification based on the initially identified primary general intents (such as balance-related, product and bond-related, transaction history-related, and instruction-related intents) to determine the more detailed secondary business intents corresponding to the user's actual needs. For example, under the primary intent of balance, it can be further subdivided into intents for querying the balance of a single account on the same day, intents for querying the balance of a single account within a certain range, and intents for querying the balance of multiple accounts.

[0151] In one embodiment, please refer to Figure 7 The above-mentioned step S150 may include steps S151 to S152.

[0152] S151. Review the processing results and match the corresponding general intent with the corresponding secondary intent. Use natural language processing technology to analyze in depth and determine the specific secondary intent that meets the user's needs.

[0153] This step primarily relies on natural language processing techniques to deeply analyze the structured semantic summary provided in the preceding steps. Leveraging the powerful text understanding and semantic parsing capabilities of large models, the system first examines and understands the core question statements, high-level general intents, and key entity information in the processing results. Next, based on this information, the system matches the user's request with a predefined library of secondary intents. Each secondary intent has its specific triggering conditions and feature descriptions; by comparing the user's actual needs with these features, the system can determine the specific secondary intent that best matches the user's requirements.

[0154] S152. Based on the specific secondary intent, create a structured expression containing specific business requirement types and necessary parameters to obtain the specific intent.

[0155] Once the specific secondary intent is determined, the next step is to construct a structured representation of this intent. This structured representation includes not only the user's specific business requirement type (e.g., the specific date range for querying account balances) but also all necessary parameters required to complete the task (e.g., account number, query start and end dates). This structured representation facilitates subsequent API calls, parameter validation, and final response generation, making the entire process more automated and efficient.

[0156] Through steps S151 and S152 described above, the system can not only accurately identify the user's actual needs, but also precisely represent these needs in a structured manner, providing clear direction and data support for subsequent services. This approach greatly improves the accuracy and efficiency of the service, while also enhancing the user experience and achieving intelligent and automated processing throughout the entire process from question acquisition to answer generation.

[0157] In step S150 of this embodiment, after confirming the user's question, this solution will further refine the intent based on the results of the general intent recognition and the optimized user question. Specifically, in this solution, the four categories of general intents are each divided into different secondary intents to ensure accurate identification of user needs and provision of targeted service responses.

[0158] The primary intent related to balances mainly involves requests to inquire about the status of account funds. These requests are further subdivided into the following specific secondary intents:

[0159] Intended purpose of single account daily balance inquiry: Users want to know the balance of funds in a specific account on a given day.

[0160] Single account range balance query intent: Users want to obtain information on the changes in funds and balance of a specific account within a specified date range.

[0161] Intended purpose of batch account balance query: Users need to query the balance status of multiple accounts at the same point in time or at different time periods.

[0162] The primary intent of this product category is to query information related to bank custody products. This type of request can be further categorized as:

[0163] Product-related account information intent: The user requests to query all account details related to a specific product, including information such as account holder and account type.

[0164] Product holdings pagination query intent: Users want to view details of the various products they hold in a paginated format, which is usually used to improve loading efficiency and user experience when dealing with large amounts of data.

[0165] The primary intent for bond-related requests is to provide users with specific information about bond investments. These requests can be further subdivided into three secondary intents:

[0166] Key Information Inquiry: Users wish to obtain key information about a specific bond, such as the issuer, maturity date, and coupon rate.

[0167] Related information search: In addition to basic key information, users may also be interested in the bond's market performance, rating changes, etc.

[0168] Default Data Supplement: Default Information Inquiry: Specifically for bonds with default risk or that have already defaulted, users who wish to obtain detailed default records and subsequent processing progress.

[0169] The first-level intent of the instruction class involves operation queries based on date or status. Specifically, it is divided into:

[0170] Date command query: Users can retrieve records of operations or transactions performed based on a specific date.

[0171] Status command query: Users are concerned with the status of certain operations or transactions (such as pending, completed, failed, etc.) and use this query to monitor the progress of their business processes.

[0172] By further subdividing the four main categories of general intents, the system can more accurately understand users' actual needs, thereby providing more precise service responses. This hierarchical structure not only helps improve the accuracy of intent recognition but also enhances the system's flexibility and adaptability, meeting the diverse needs of different users. Furthermore, this detailed classification provides clear direction for subsequent API calls, parameter validation, and response generation, promoting the automation and intelligence of the entire service process. In this way, service efficiency is improved, and the user experience is significantly enhanced.

[0173] S160. Automatically call the corresponding interface to generate a response based on the specific intent, and present it in a structured and visual manner.

[0174] This stage involves precise processing and response generation based on the identification results (including general intents and secondary intents). This stage aims to generate high-quality response content that meets business scenarios and compliance requirements through a series of methods and technologies, such as precise matching, parameter validation, API calls, and interaction optimization.

[0175] First, a detailed mapping mechanism is established to explicitly associate each secondary intent with the corresponding business processing tool interface. This ensures that once intent recognition is complete, the system can quickly locate the corresponding tool interface. For example, if a user's request is identified as "single account daily balance query intent," it directly corresponds to the specific interface responsible for that type of query.

[0176] Before calling the relevant API, the system automatically performs parameter integrity checks. If all required parameters are provided, the corresponding API is called directly to retrieve the data. If any parameters are missing, the system generates a unique tag based on the specific missing item and sends it to the user requesting supplementary information. Simultaneously, the system provides the user with a clear list of parameters, allowing them to choose between "user-supplemented" or "system-supplemented" options to complete the required information.

[0177] To improve user reading efficiency and comprehension, the front-end page supports displaying answers in a structured format. For questions involving details, such as account transaction records or product holdings, tables and charts are used for visualization, making complex data relationships clear at a glance.

[0178] When a problem involves multiple steps that require the user to perform, the system provides distributed operation guidance and displays prompts on the interface to help the user complete the task step by step. This approach not only simplifies complex processes but also reduces potential confusion and errors that users may encounter during operation.

[0179] To meet users' needs for documents and materials, the system supports online viewing and downloading of relevant files. Safe and reliable download links are provided to ensure users can easily and quickly access the documents they need, while also protecting the security and privacy of the data.

[0180] By optimizing the interactive experience through the two main aspects mentioned above—the intent-tool mapping mechanism and the multimodal display on the front end—the entire system not only achieves seamless integration from intent recognition to actual business processing, but also greatly enhances user experience and satisfaction. This comprehensive solution considers both the accuracy and efficiency of technical implementation and the convenience and comfort of users during use, representing an important trend in modern service system design.

[0181] This embodiment combines intelligent agent and large model technologies to develop a more flexible and efficient AI assistant for custody services, tailored to the specific business characteristics of banking custody service platforms. This solution not only enhances the platform's intelligence level but also significantly improves business processing efficiency and optimizes the user experience.

[0182] At the model level, Qwen2-72B was adopted as the basic large-scale model capability provider. Qwen2-72B is a highly versatile interactive language model that can support a wide range of application scenarios. Its powerful natural language processing capabilities provide a solid foundation for managed AI assistants, enabling them to understand and respond to various user needs and achieve more efficient and accurate communication.

[0183] At the agent level, this example innovatively employs a multi-agent architecture design. Each agent constructs its interaction pattern in the form of a decision tree. This hierarchical and progressive task processing approach allows the system to progressively refine the key points of the task, thereby achieving accurate and efficient processing of various specific tasks. This design not only enhances the system's flexibility but also enables the entire system to better adapt to complex and ever-changing real-world business environments.

[0184] To address random user behavior during question-and-answer sessions, such as missing question components or semantic ambiguity, this solution employs a variety of advanced large-scale modeling techniques, including but not limited to sampling decoding, intent judgment, intent rewriting, entity extraction, and function calls, for targeted processing. These technologies work together to ensure that the AI ​​assistant accurately understands the semantics of the customer's speech, combines this with information from the knowledge base, and provides precise answers, overcoming the limitations of traditional chatbots that exhaustively enumerate user dialogue intent.

[0185] Since the introduction of managed AI assistants, the method in this embodiment has significantly improved its intelligence level and business processing efficiency. Specifically:

[0186] Statistics show that the proportion of manual intervention has decreased by more than 45%, effectively reducing the workload of business personnel.

[0187] Overall operational efficiency improved by approximately 50%, providing strong support for business continuity.

[0188] The intelligent front-end interactive experience greatly reduces the barrier to entry for using the platform and improves the ease of operation and response speed.

[0189] In summary, this embodiment, based on intelligent agent and large model technologies and combined with actual business needs, realizes an innovative application of managed AI assistants. It not only solves the problem of low automation in existing managed operation platforms, but also greatly reduces manual intervention through a fully automated processing mechanism, improving service efficiency and user experience. Furthermore, by adopting a multi-layered intelligent agent design and advanced large model technology, this embodiment can flexibly respond to various complex business scenarios, providing users with a more intelligent and convenient service experience.

[0190] The aforementioned managed AI service method, by introducing advanced intelligent agents and large model technologies, achieves a fully automated processing mechanism from acquiring user questions to automatically calling the corresponding interfaces to generate answers. First, the user's question undergoes semantic cleansing and disambiguation, identifying its general intent and key entities to generate a structured summary. Next, based on single-turn or multi-turn dialogue history, large model technology is used to accurately understand and refine the user's intent, outputting a confidence-based intent label, a clear question statement, and an entity list as the processing result. Subsequently, this result is further subdivided into specific secondary business intents to accurately pinpoint user needs. Finally, based on the specific intent, the corresponding interfaces are automatically called to generate highly targeted answers, which are presented to the user in a structured and visual manner. This process not only significantly improves the system's automation level and reduces the need for manual intervention but also greatly optimizes the user experience through intelligent front-end interaction, while enhancing the system's adaptability and scalability in the face of complex and ever-changing business scenarios. This effectively solves the shortcomings of existing systems, such as low automation, poor user experience, and insufficient flexibility.

[0191] Figure 8 This is a schematic block diagram of a managed AI service device 300 provided in an embodiment of the present invention. Figure 8 As shown, corresponding to the above-described managed AI service method, the present invention also provides a managed AI service apparatus 300. This managed AI service apparatus 300 includes a unit for executing the above-described managed AI service method, and the apparatus can be configured in a server. Specifically, please refer to... Figure 8 The managed AI service device 300 includes a problem acquisition unit 301, a preprocessing unit 302, an intent processing unit 303, a refinement unit 304, a localization unit 305, and a generation unit 306.

[0192] The question acquisition unit 301 is used to acquire various questions from users; the preprocessing unit 302 is used to perform semantic cleansing and disambiguation on the questions, and identify their general intents and key entities to generate a structured summary; the intent processing unit 303 is used to accurately understand and process the questions based on single-turn or multi-turn dialogue history using large model technology to obtain user intents; the refinement unit 304 is used to refine user intents in combination with the structured summary, and output intent tags with confidence levels, clear question descriptions, and entity lists to obtain processing results; the positioning unit 305 is used to further subdivide the processing results into specific secondary business intents to accurately locate user needs and obtain specific intents; the generation unit 306 is used to automatically call the corresponding interface to generate answers based on the specific intents and present them in a structured and visual manner.

[0193] In one embodiment, the question acquisition unit 301 is used to submit questions through a customized Tools card service or to raise questions using speech recognition, and to perform real-time semantic analysis based on big data modeling technology, automatically associate historical dialogue records and provide suggested questions to obtain various types of questions from users.

[0194] In one embodiment, the preprocessing unit 302 includes:

[0195] The purification subunit is used to remove redundant information and sentiment words from the question using the Qwen 72B big model, and to resolve ambiguity based on context, extracting the core question expression to obtain the purified question text; the classification subunit is used to identify and classify the business category to which the user's question belongs based on the purified question text; the identification subunit is used to identify key business entities and action elements from the purified question text; and the integration subunit is used to integrate the business categories and key business entities and action elements to obtain a structured semantic summary containing the core of the question, general intent, and key elements.

[0196] In one embodiment, the intent recognition unit includes:

[0197] The analysis subunit is used to analyze the question and determine whether the question is a single-turn or multi-turn dialogue type through structured semantic summarization. The user intent recognition subunit is used to, when the question is a multi-turn dialogue type, combine the recent dialogue history, use RoPE technology to maintain the dialogue order and sense of distance, and use the self-attention and cross-attention mechanism of the large model to evaluate the semantic association between the current and historical dialogue words to achieve referential resolution. When encountering intent jumps or ambiguities, it generates questions based on existing information to clarify the user intent and obtain the user intent.

[0198] In one embodiment, the refinement unit 304 includes:

[0199] The comparison subunit is used to compare the user intent with the predefined business intent vector, identify and determine the general intent category that best meets the user's needs, so as to obtain the general intent; the adjustment subunit is used to combine the structured summary and user historical behavior data to adjust the general intent to specific user needs, and generate a structured processing result, so as to obtain the processing result.

[0200] In one embodiment, the adjustment subunit includes;

[0201] The integration module combines contextual information from single-turn and multi-turn dialogues to integrate the general intent. The general intent processing module uses a large-scale model to process the general intent and clarify the user's core needs. The specific intent adjustment module combines the structured summary, compares and verifies the consistency between the extracted entity information and the context, and adjusts the general intent to a specific, precise intent. The weighted analysis module considers the influence of the user's historical behavior patterns and outputs the final user intent through weighted analysis to obtain the processing result.

[0202] In one embodiment, the positioning unit 305 includes:

[0203] The review subunit is used to review the processing results and match the corresponding general intent with the corresponding secondary intent. Natural language processing technology is used to conduct in-depth analysis and determine the specific secondary intent that meets the user's needs. The creation subunit is used to create a structured expression containing specific business requirement types and necessary parameters based on the specific secondary intent to obtain the specific intent.

[0204] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the aforementioned managed AI service device 300 and each unit can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0205] The aforementioned hosted AI service device 300 can be implemented as a computer program, which can, for example... Figure 9 It runs on the computer device shown.

[0206] Please see Figure 9 , Figure 9 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0207] See Figure 9 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0208] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform a managed AI service method.

[0209] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0210] The internal memory 504 provides an environment for the execution of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a managed AI service method.

[0211] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0212] The processor 502 is used to run a computer program 5032 stored in the memory to perform the following steps:

[0213] The system acquires various user questions; performs semantic cleansing and disambiguation on the questions, identifies their general intents and key entities, and generates a structured summary; based on single-turn or multi-turn dialogue history, it uses large-scale modeling techniques to accurately understand and process the questions to obtain user intents; it refines user intents by combining the structured summary, outputting intent tags with confidence levels, clear question descriptions, and entity lists to obtain processing results; it further subdivides the processing results into specific secondary business intents to accurately locate user needs and obtain specific intents; it automatically calls the corresponding interfaces to generate answers based on the specific intents and presents them in a structured and visual manner.

[0214] In one embodiment, when implementing the steps for obtaining various user-related questions, the processor 502 specifically implements the following steps:

[0215] Users can submit questions through customized Tools card services or ask questions using voice recognition. Based on big data modeling technology, real-time semantic analysis is performed, and historical dialogue records are automatically linked to provide suggested questions to obtain various types of questions from users.

[0216] In one embodiment, when the processor 502 performs semantic cleansing, disambiguation, identification of general intents and key entities, and generation of structured summaries on the problem, it specifically implements the following steps:

[0217] The Qwen 72B large model is used to remove redundant information and sentiment words from the question, and the core question expression is extracted based on the context to obtain the purified question text. Based on the purified question text, the business scope to which the user's question belongs is identified and classified. Key business entities and action elements are identified from the purified question text. The business scope, key business entities, and action elements are integrated to obtain a structured semantic summary containing the core of the question, general intent, and key elements.

[0218] In one embodiment, when the processor 502 implements the step of accurately understanding and processing the problem based on single-turn or multi-turn dialogue history using large model technology to obtain user intent, it specifically implements the following steps:

[0219] The problem is analyzed, and the structured semantic summary is used to determine whether the problem is a single-turn or multi-turn dialogue type. If the problem is a multi-turn dialogue type, the RoPE technique is used to maintain the dialogue order and sense of distance by combining the recent dialogue history. The self-attention and cross-attention mechanism of the large model is used to evaluate the semantic association between the current and historical dialogue words to achieve referential resolution. When encountering cases of intention jump or ambiguity, a question is generated based on the existing information to clarify the user's intention in order to obtain the user's intention.

[0220] In one embodiment, when the processor 502 refines the user intent by combining the structured summary and outputs intent tags with confidence levels, explicit question statements, and entity lists to obtain the processing result, it specifically implements the following steps:

[0221] The user intent is compared with a predefined business intent vector to identify and determine the generic intent category that best matches the user's needs, thus obtaining a generic intent. The generic intent is then adjusted to specific user needs by combining the structured summary and the user's historical behavior data, and a structured processing result is generated, thus obtaining the processing result.

[0222] In one embodiment, when the processor 502 implements the step of combining the structured summary and user historical behavior data to adjust the general intent to specific user needs and generate a structured processing result to obtain the processing result, the processor 502 specifically implements the following steps:

[0223] By combining contextual information from single-turn and multi-turn dialogues, the general intent is integrated; a large-scale model is used to process the general intent to clarify the user's core needs. Based on the structured summary, the consistency between the extracted entity information and the context is compared and verified, adjusting the general intent to a specific, precise intent; considering the influence of the user's historical behavior patterns, a weighted analysis is used to output the final user intent, thus obtaining the processing result.

[0224] In one embodiment, when the processor 502 further subdivides the processing result into specific secondary business intents to accurately locate user needs and obtain specific intents, it specifically implements the following steps:

[0225] The processing results are reviewed, and the corresponding general intent is matched with the corresponding secondary intent. Natural language processing technology is used to conduct in-depth analysis and determine the specific secondary intent that meets the user's needs. Based on the specific secondary intent, a structured expression containing specific business requirement types and necessary parameters is created to obtain the specific intent.

[0226] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0227] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0228] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform the following steps:

[0229] The system acquires various user questions; performs semantic cleansing and disambiguation on the questions, identifies their general intents and key entities, and generates a structured summary; based on single-turn or multi-turn dialogue history, it uses large-scale modeling techniques to accurately understand and process the questions to obtain user intents; it refines user intents by combining the structured summary, outputting intent tags with confidence levels, clear question descriptions, and entity lists to obtain processing results; it further subdivides the processing results into specific secondary business intents to accurately locate user needs and obtain specific intents; it automatically calls the corresponding interfaces to generate answers based on the specific intents and presents them in a structured and visual manner.

[0230] In one embodiment, when the processor executes the computer program to implement the steps of obtaining various user-related questions, it specifically implements the following steps:

[0231] Users can submit questions through customized Tools card services or ask questions using voice recognition. Based on big data modeling technology, real-time semantic analysis is performed, and historical dialogue records are automatically linked to provide suggested questions to obtain various types of questions from users.

[0232] In one embodiment, when the processor executes the computer program to perform semantic cleansing, disambiguation, identification of general intents and key entities, and generation of structured summaries for the problem, it specifically implements the following steps:

[0233] The Qwen 72B large model is used to remove redundant information and sentiment words from the question, and the core question expression is extracted based on the context to obtain the purified question text. Based on the purified question text, the business scope to which the user's question belongs is identified and classified. Key business entities and action elements are identified from the purified question text. The business scope, key business entities, and action elements are integrated to obtain a structured semantic summary containing the core of the question, general intent, and key elements.

[0234] In one embodiment, when the processor executes the computer program to implement the step of accurately understanding and processing the problem based on single-turn or multi-turn dialogue history using large model techniques to obtain user intent, it specifically implements the following steps:

[0235] The problem is analyzed, and the structured semantic summary is used to determine whether the problem is a single-turn or multi-turn dialogue type. If the problem is a multi-turn dialogue type, the RoPE technique is used to maintain the dialogue order and sense of distance by combining the recent dialogue history. The self-attention and cross-attention mechanism of the large model is used to evaluate the semantic association between the current and historical dialogue words to achieve referential resolution. When encountering cases of intention jump or ambiguity, a question is generated based on the existing information to clarify the user's intention in order to obtain the user's intention.

[0236] In one embodiment, when the processor executes the computer program to refine the user intent by combining the structured digest and outputting intent labels with confidence levels, explicit question statements, and entity lists to obtain the processing result, the processor specifically implements the following steps:

[0237] The user intent is compared with a predefined business intent vector to identify and determine the generic intent category that best matches the user's needs, thus obtaining a generic intent. The generic intent is then adjusted to specific user needs by combining the structured summary and the user's historical behavior data, and a structured processing result is generated, thus obtaining the processing result.

[0238] In one embodiment, when the processor executes the computer program to implement the step of combining the structured summary and user historical behavior data to adjust the general intent to specific user needs and generate a structured processing result to obtain the processing result, the processor specifically implements the following steps:

[0239] By combining contextual information from single-turn and multi-turn dialogues, the general intent is integrated; a large-scale model is used to process the general intent to clarify the user's core needs; the structured summary is combined to compare and verify the consistency between the extracted entity information and the context, and the general intent is adjusted to a specific and precise intent; the influence of the user's historical behavior patterns is considered, and the final user intent is output through weighted analysis to obtain the processing result.

[0240] In one embodiment, when the processor executes the computer program to further subdivide the processing result into specific secondary business intents, accurately locate user needs, and obtain specific intents, the processor specifically implements the following steps:

[0241] The processing results are reviewed, and the corresponding general intent is matched with the corresponding secondary intent. Natural language processing technology is used to conduct in-depth analysis and determine the specific secondary intent that meets the user's needs. Based on the specific secondary intent, a structured expression containing specific business requirement types and necessary parameters is created to obtain the specific intent.

[0242] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0243] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0244] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0245] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0246] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0247] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for managing AI services, characterized in that, include: Obtain various questions from users; The problem is semantically cleaned and disambiguated, and its general intent and key entities are identified to generate a structured summary; Based on single-turn or multi-turn dialogue history, large model technology is used to accurately understand and process the question in order to obtain user intent; The user intent is refined by combining the structured summary, and the intent tags with confidence levels, clear question statements and entity lists are output to obtain the processing results; The processing results are further subdivided into specific secondary business intents to accurately pinpoint user needs and obtain specific intents; The system automatically invokes the corresponding interface to generate a response based on the specific intent, and presents it in a structured and visual manner.

2. The managed AI service method according to claim 1, characterized in that, The various questions regarding user acquisition include: Users can submit questions through customized Tools card services or ask questions using voice recognition. Based on big data modeling technology, real-time semantic analysis is performed, and historical dialogue records are automatically linked to provide suggested questions to obtain various types of questions from users.

3. The managed AI service method according to claim 1, characterized in that, The process of semantic cleansing and disambiguation of the problem, identifying its general intent and key entities, and generating a structured summary includes: The Qwen 72B large model was used to remove redundant information and sentiment words from the question, and the ambiguity was resolved based on the context to extract the core question expression in order to obtain the purified question text. Based on the purified question text, the business scope to which the user's question belongs is identified and classified; Key business entities and action elements were identified from the purified problem text; The business scope, key business entities, and action elements are integrated to obtain a structured semantic summary that includes the core of the problem, general intent, and key elements.

4. The managed AI service method according to claim 1, characterized in that, The process of accurately understanding and processing the question based on single-turn or multi-turn dialogue history using large model techniques to obtain user intent includes: Analyze the problem and determine whether it is a single-turn or multi-turn dialogue type through structured semantic summarization; When the problem is a multi-turn dialogue type, the RoPE technique is used to maintain the dialogue order and sense of distance by combining the recent dialogue history. The self-attention and cross-attention mechanism of the large model is used to evaluate the semantic association between the current and historical dialogue words to achieve referential resolution. When encountering cases of intention jump or ambiguity, a question is generated based on the existing information to clarify the user's intention in order to obtain the user's intention.

5. The managed AI service method according to claim 1, characterized in that, The process involves refining the user intent using the structured summary, outputting intent tags with confidence levels, explicit question statements, and entity lists to obtain the processing results, including: The user intent is compared with a predefined business intent vector to identify and determine the generic intent category that best meets the user's needs, thus obtaining the generic intent. By combining the structured summary and user history behavior data, the general intent is adjusted to specific user needs, and a structured processing result is generated to obtain the processing outcome.

6. The managed AI service method according to claim 5, characterized in that, The process involves combining the structured summary and user historical behavior data to adjust general intents to specific user needs and generate structured processing results, including: By combining contextual information from single-turn and multi-turn dialogues, the general intent is integrated; The general intent is processed using a large model to clarify the user's core needs; By combining the structured summary, the consistency between the extracted entity information and the context is compared and verified, and the general intent is adjusted to a specific and precise intent. Taking into account the influence of users' historical behavior patterns, the final user intent is output through weighted analysis to obtain the processing results.

7. The managed AI service method according to claim 1, characterized in that, The process further subdivides the processing result into specific secondary business intents to accurately pinpoint user needs and obtain specific intents, including: The processing results are reviewed, and the corresponding general intent is matched with the corresponding secondary intent. Natural language processing technology is used to conduct in-depth analysis and determine the specific secondary intent that meets the user's needs. Based on the specific secondary intent, a structured expression containing specific business requirement types and necessary parameters is created to obtain the specific intent.

8. A device for hosting AI services, characterized in that, include: The problem acquisition unit is used to acquire various problems from users; The preprocessing unit is used to perform semantic cleansing and disambiguation on the problem, identify its general intent and key entities, and generate a structured summary. The intent processing unit is used to accurately understand and process the question based on the single-turn or multi-turn dialogue history using large model technology to obtain the user's intent; The refinement unit is used to refine the user intent in conjunction with the structured summary, and output intent tags with confidence levels, clear question statements and entity lists to obtain the processing results; The positioning unit is used to further subdivide the processing result into specific secondary business intents, accurately locate user needs, and obtain specific intents. The generation unit is used to automatically call the corresponding interface to generate a response based on the specific intent, and present it in a structured and visual manner.

9. A computer device, characterized in that, The computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images