Retrieval enhancement generation system and method for self-defined multi-level intention recognition

Abstract

The invention discloses a retrieval enhancement generation system and method for customizing multilevel intention recognition. The retrieval enhancement generation system comprises a multilevel intention label and document management module, a multilevel intention recognition module and a retrieval enhancement generation module. The multi-level intention label and document management module can self-define and build a corresponding relationship among a multi-level intention label, an intention label and a document level by level to form a label document corresponding relationship form; the multi-level intention recognition module recognizes intentions of the user level by level according to the multi-level intention labels and the multi-level intention labels established by the document management module, and gives corresponding intention labels; and the retrieval enhancement generation module performs vector similarity search in a document range delineated by the intention tag, and transmits a vector similarity search result to the large language model for result generation. Through step-by-step subdivision of the user intention, the document retrieval range is gradually reduced to the corresponding intention label, so that the answer generated by the large language model in the generation stage is more accurate.

Description

Technical Field

[0001] The present invention belongs to the field of natural language processing (NLP) and artificial intelligence technology, and specifically is a system for customizing multi-level intent recognition RAG (retrieval enhancement generation) technology and its implementation method. Background Art

[0002] The following content will introduce and explain the terms involved in the present invention and the existing technical problems.

[0003] (1) Intent Recognition: Intent recognition is a key technology in natural language processing (NLP). It aims to determine the specific purpose or needs of users in conversations or interactions by analyzing the semantics and context of user statements. It usually relies on classification algorithms or deep learning models, combining semantic features, keywords, contextual dependencies, and other information to output one or more intent categories. In some professional scenarios, the types and levels of intent can be more complex, and usually a higher-level intent will be further subdivided.

[0004] (2) Retrieval-Augmented Generation (RAG): Retrieval-Augmented Generation (RAG) is a method that combines generative language models and information retrieval technology, aiming to enhance the accuracy and relevance of generative models through external knowledge. In the RAG framework, the retrieval module first selects the documents most relevant to the input question from the external knowledge base, and then the retrieved content is input as context to the generation module to generate answers or content. This method overcomes the knowledge cutoff problem of the generative model and improves the ability to handle complex questions or long-tail information. RAG has significant advantages in tasks such as question-answering systems, dialogue generation, and content summarization.

[0005] (3) Large Language Model (LLM): Large Language Model (LLM) refers to a large-scale language model based on deep learning technology, which is usually composed of billions or even hundreds of billions of parameters and is designed to process and generate natural language text. This type of model is based on the Transformer architecture and masters multi-dimensional information such as grammar, semantics, context, and world knowledge through unsupervised training on massive text data. Typical large language models include OpenAI's GPT series, Google's BERT, and Meta's LLaMa, which are widely used in tasks such as natural language understanding (NLU), generation (NLG), dialogue systems, and machine translation. Its core advantage is that it can efficiently capture complex language patterns and contexts through large-scale parameters and training data. An important feature of large language models is their token length limit, which is determined by the complexity of computing resources and the self-attention mechanism in the Transformer architecture. The computational complexity of the self-attention mechanism grows quadratically with the input length (O(n 2 Therefore, to balance performance and efficiency, large language models often set an upper limit on the number of tokens they can process when processing input. For example, GPT-4's token length limit may be several thousand, which determines the scope of context that the model can handle in a single interaction.

[0006] (4) Large language model assisted by retrieval enhancement: Using retrieval enhancement to enhance large language models is a technical strategy that combines external knowledge bases with generative language models. Although large language models have powerful natural language generation and understanding capabilities, their knowledge scope is usually limited by the time and coverage of training data, making it difficult to handle problems with high real-time requirements or complex domain knowledge. Retrieval-enhanced generation introduces a retrieval module into the generation process, obtaining input-related information from an external knowledge base in real time, and then inputs these retrieved contents as additional context into the language to generate results. This approach can significantly improve the accuracy of language generation and the scope of knowledge coverage, while reducing the limitations of the model parameter scale on the quality of the results.

[0007] In the prior art, a large language model assisted by retrieval enhancement is implemented in the following ways: (1) Retrieval stage: using vectorized representation and semantic search technology, the document or fragment that best matches the input question is extracted from the knowledge base. (2) Generation stage: the input question is combined with the retrieval results and passed to the large language model as an extended context, which generates the answer or content. This enhancement method is widely used in question-answering systems, enterprise knowledge management, legal interpretation, medical consultation and other fields. It solves the problems of knowledge loss or ambiguous answers that may occur when the large language model is used alone, and improves the practicality and domain adaptability of the model. In order to enable the large language model to give more accurate answers in the generation stage, it is required that the retrieval results in the limited context passed to the model are more accurate.

[0008] Currently, the conventional approach is to first segment document content and store it in a vector database through vectorization to manage external knowledge. Then, during the retrieval phase, the vectors are compared for similarity, and the results with the highest similarity are selected and passed to the large language model. However, this technical solution has the following drawbacks: due to its crude search method, it fails to prioritize similar words in different fields, resulting in inaccurate retrieval results and, in turn, incorrect results generated by the large language model. Summary of the Invention

[0009] To address the aforementioned issues in the background art, this paper proposes a customized multi-level intent recognition retrieval enhancement generation system and its implementation method. By segmenting user intent level by level, the document retrieval scope is gradually narrowed to the corresponding intent tags, thereby making the answers generated by the large language model during the generation phase more accurate.

[0010] To achieve the above-mentioned objectives, the present invention provides the following technical solutions: a retrieval enhancement generation system with customized multi-level intent recognition; comprising a multi-level intent label and document management module, a multi-level intent recognition module and a retrieval enhancement generation module; wherein the multi-level intent label and document management module can be customized, and according to the user's definition of the scene, multi-level intent labels are established step by step, and according to the user's settings, a correspondence between intent labels and documents is established to form a label-document correspondence table; the multi-level intent recognition module identifies the user's intent step by step based on the multi-level intent labels established by the multi-level intent label and document management module, and gives corresponding intent labels; the retrieval enhancement generation module performs a vector similarity search in the document range circled by the intent label formed in the multi-level intent recognition module, and passes the results of the vector similarity search to the large language model for result generation.

[0011] Furthermore, the implementation steps of the multi-level intent label and document management module are as follows: (1) the user adds intent labels step by step according to the needs of the scenario to form an intent label form, and each intent label has a corresponding number in the intent label form to form an intent label number; (2) the user prepares the corresponding knowledge base document according to the needs of the scenario, vectorizes the knowledge base document, obtains the document slice vector, stores the document slice vector in the document vector database, and each document slice vector has a corresponding document vector number; (3) the user marks each knowledge base document with the corresponding intent label according to the added intent label, and the intent label number and the document vector number correspond one to one, and are added to the document to form a label-document correspondence form.

[0012] Furthermore, the multi-level intent label recognition module includes a prompt word template database and a multi-level intent recognition system. The prompt word template database maintains the prompt word templates of users in different intent scenarios. The multi-level intent recognition system uses a large language model as the base model for intent recognition. Through a predetermined prompt word template database, the contents of the multi-level intent labels and the intent labels in the document management module are spliced, and the spliced ​​contents and user requests are input together as a splicing problem into the large language model for intent recognition.

[0013] Furthermore, the steps of the intent recognition process are as follows: (1) the user inputs a document request; (2) the multi-level intent recognition system recognizes the intent tag of the document; (3) if the intent tag of the document is recognized and the intent does not have a sub-intent, the recognition result is the final intent of the document; (4) if the intent tag of the document is recognized, but the intent still has a sub-intent, then the sub-level intent recognition is continued; (5) if the obvious intent tag of the document is not recognized and there is an identified parent intent, then the parent intent is the final intent of the text; (6) if the obvious intent tag is not recognized and there is no identified parent intent, then the current intent is temporarily determined to have no obvious intent; when the intent tag recognition result of the current user request is temporarily determined to have no obvious intent, the historical intent recognition tag is considered; if there is a historical intent recognition tag, the intent recognition tag of the current document is the historical intent; if there is no historical intent recognition tag, then the current text is defined as having no obvious intent.

[0014] Furthermore, the implementation steps of the retrieval enhancement generation module are as follows: (1) extracting document slice vector data corresponding to the document slice from the knowledge base document database; (2) vectorizing the user's request; (3) comparing the cosine similarity between the document slice vector and the user request vectorized data, and selecting the document slice with the highest similarity as the most similar document slice; (4) splicing the most similar document slice and the user's request vector to form a spliced ​​question; (5) passing the spliced ​​question to the large language model, allowing the large language model to generate the corresponding answer result; (6) returning the user's request vector and the answer result generated by the large language model to the historical conversation record form, as a historical record of the current conversation, or as a subsequent splicing with the spliced ​​question input, the historical conversation record and the result of the large language model retrieval, together with the historical conversation record and the user's request vector to form the splicing of the input question, and calling the large language model to generate the answer.

[0015] Furthermore, the knowledge base documents are vectorized using Llama vectorization, and the document vector database uses the FAISS database.

[0016] In addition, the present invention also provides a method for implementing a retrieval enhancement generation system for customized multi-level intent recognition, including the steps described in the above modules.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] 1. The customized multi-level intent recognition retrieval enhancement generation system provided by the present invention can adapt to the needs of different professional fields, such as medical care, education, finance, law, etc., through the combination of multi-level intent recognition and RAG, and has a broad application base.

[0019] 2. The customized multi-level intent recognition retrieval enhancement generation system provided by the present invention combines multi-level intent recognition with RAG, has efficient retrieval speed, accurate content generation, can provide accurate professional support, and solves the complicated problems of complex information query and knowledge transfer.

[0020] 3. The customized multi-level intent recognition retrieval enhancement generation system provided by the present invention can serve as the core component of many functions, including: intelligent customer service (the system can handle complex dialogue scenarios with multiple intentions and multiple tasks, support FAQs, personalized suggestions, problem escalation and other functions, provide users with a smooth interactive experience, and significantly reduce the workload of manual customer service), intelligent assistants in professional fields (the system can combine professional knowledge and dynamic data to play the role of an intelligent assistant in a specific field, providing users with auxiliary decision support, in-depth data analysis, and real-time knowledge update services to help users quickly obtain valuable information), enterprise knowledge base management and consulting tools (the system integrates the internal knowledge base of the enterprise with external data resources to provide accurate information retrieval and consulting services. It can optimize the knowledge management process, improve employees' query efficiency and work decision-making ability, and is suitable for cross-departmental collaboration and enterprise knowledge sharing scenarios).

[0021] 4. Compared with traditional large models and simple RAG, the customized multi-level intent recognition retrieval enhancement generation system provided by the present invention has been improved and enhanced on the basis of the RAG's own capabilities by introducing a customizable multi-level intent recognition module: (1) Ability to adapt to complex interactions - the system can hierarchically parse the intentions of user input, accurately handle complex requirements, and is suitable for multi-task, multi-level interaction scenarios; (2) Configurable and customizable functions - users can flexibly configure intent classification, retrieval strategy and generation style according to specific scenarios. The system has strong adaptability and great expansion potential; (3) Cross-domain versatility - the system has cross-industry adaptability and can be integrated into different business processes at low cost to meet diverse usage needs; (4) Efficient real-time response - through the mechanism of multi-level intent recognition, the scope of retrieval enhancement documents is narrowed, the system can quickly process user requests, support high concurrent interactions, and ensure service stability; (5) Continuous optimization and learning capabilities - based on usage feedback, users can easily adjust the configuration of intent labels and documents, which can quickly improve and enhance the accuracy of intent recognition and response generation and user satisfaction.

[0022] Therefore, the customized multi-level intent recognition retrieval enhancement generation system provided by the present invention has strong scenario application capabilities and potential market competitiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A schematic diagram of the system of the present invention;

[0024] Figure 2 This is a flowchart of the multi-level intent tagging and document management module of the present invention;

[0025] Figure 3 This is a flow chart of the multi-level intent tag recognition module of the present invention;

[0026] Figure 4 for Figure 3 Flowchart of the multi-level intent recognition system implementation method of the multi-level intent label recognition module;

[0027] Figure 5 This is a flow chart of the search enhancement generation module of the present invention. DETAILED DESCRIPTION

[0028] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0029] like Figure 1 As shown in Figure 1, the customized multi-level intent recognition retrieval enhancement generation system consists of a multi-level intent label and document management module, a multi-level intent recognition module, and a retrieval enhancement generation module. Figure 2 The multi-level intent label and document management module shown can be customized. It establishes multi-level intent labels step by step according to the user's definition of the scene, and establishes the correspondence between intent labels and documents according to user settings to form a label-document correspondence table. Figure 3 The multi-level intent recognition module shown is a multi-level intent recognition model, which can establish multi-level intent labels based on the multi-level intent labels and document management modules, such as Figure 4 The multi-level intention recognition system shown in Figure 1 identifies the user's intentions step by step and gives corresponding intention labels. Figure 5 The retrieval enhancement generation module shown performs a vector similarity search within the document range defined by the intent label formed in the multi-level intent recognition module, and passes the result of the vector similarity search to the large language model for result generation.

[0030] like Figure 2The flowchart of the multi-level intent label and document management module is shown. The specific implementation method is as follows: First, the user adds intent labels step by step according to the needs of the scenario to form an intent label form, and each intent label has a corresponding number in the intent label form to form an intent label number. Second, the user prepares the corresponding knowledge base document according to the needs of the scenario, vectorizes the knowledge base document, obtains the document slice vector, stores the document slice vector in the document vector database, and each document slice vector has a corresponding document vector number; in the best implementation method, the knowledge base document is vectorized by Llama vectorization, and the document vector database adopts the FAISS database. Third, the user marks each knowledge base document with the corresponding intent label one by one according to the added intent label, wherein the intent label number and the document vector number correspond one to one, and are added and formed in the label document correspondence table.

[0031] The multi-level intent tag and document management module maintains multi-level intent tags, document slice vectorization data, and the correspondence between intent tags and knowledge base documents in a database. This module creates an intent tag form and a tag-document correspondence form. These two forms are maintained in a database format and support addition, deletion, modification, and querying. This allows users to adjust the design of multi-level intent tags, the composition of knowledge base documents, and the correspondence between the two based on scenario requirements.

[0032] The multi-level intent tag and document management module allows users to customize multi-level intent tags and knowledge base document configurations based on business needs, and to add, delete, modify, and query the database. Through tag management, users can define a hierarchy of intent classifications (e.g., first-level intents, second-level intents, etc.) and dynamically configure the document knowledge base. This maintains the intent tag table and the tag-document mapping table.

[0033] Define the intent label form as follows:

[0034] serial number Tag content Tag Hierarchy Parent tag number

[0035] The header contents are as follows:

[0036] Number: Primary key, unified number (UID) for all tags, which increases naturally as tags are added;

[0037] Tag content: The content of the tag itself, usually a phrase describing the intent, such as "application document";

[0038] Tag level: The corresponding tag level number, starting from 0 for the highest level and the lowest level is unlimited and depends on the user's definition;

[0039] Parent tag number: Indicates the tag's parent tag number. The parent tag number must be lower than the tag's own number, and the parent tag's level must also be lower than the current tag itself. When a parent tag is deleted, all its descendant tags are deleted as well.

[0040] The label document correspondence form is defined as follows:

[0041] Document Number Document Overview Intent tag number Vector tile number

[0042] The header contents are as follows:

[0043] Document ID: Primary key, unique ID (UID) of the original document in the knowledge base, which increases naturally as documents are added;

[0044] Document Overview: An overview of the document content, which can be the title of the document, mainly assists in verifying the search results;

[0045] Intent tag number: The intent tag number group corresponding to the document is maintained in a set form, representing the intent content that this document can answer. When the added tag is a subtag, its parent tag can be automatically added one by one;

[0046] Vector slice number: After the document slice is vectorized, the label stored in the vector database is used to extract the corresponding database for similarity matching after intent recognition.

[0047] In this embodiment, document slices are vectorized using the Llama model, and the vectorized documents are maintained in the FAISS database.

[0048] Since the multi-level intent tag and document management module provides flexible tag management functions, users can add, modify or delete intent tags to adapt to different business scenarios; it supports dynamic configuration of document sources and retrieval rules, and changes the document content and intent tag correspondence through the addition, deletion, modification and query of the tag-document correspondence database; in addition, the multi-level intent tag and document management module provides a UI interface for tag and document management, which can be used to add, delete, modify and query operations, more intuitively manage the relationship between multi-level tags and documents, and simplify complex database operation processes. Due to the flexibility and scalability of the multi-level intent tag and document management module, users can quickly update the system according to business changes and adapt to new task requirements; low threshold use: no deep technical background is required to complete configuration operations, reducing implementation costs; precise retrieval: through customized retrieval rules, ensure that highly relevant content is retrieved first, improving the quality of system response.

[0049] like Figure 3The multi-level intent tag recognition module flow diagram is shown in Figure 1. The multi-level intent recognition module is a multi-level intent recognition model that can identify the user's intent level by level based on the multi-level intent tags and multi-level intent tags established by the document management module and give the corresponding intent tags.

[0050] The multi-level intent label recognition module uses a large language model as the base model for intent recognition. It uses a pre-defined prompt word template database to splice the contents of the multi-level intent labels and the intent labels in the document management module. According to the user's input questions and historical intent record data, the module is divided into a prompt word template database and a document management module. Figure 4 The multi-level intent recognition system shown in the figure, in which the prompt word template database is used to maintain the user's prompt word templates in different intent scenarios, and is spliced ​​with the intent content and used together with the user request as a splicing question to be input into the large language model for intent recognition.

[0051] The prompt word template database serves as the prompt word template for the large language model for intent recognition. The table (header) of the database is defined as follows:

[0052] Template Number Template Contents Applicable tags

[0053] Template ID: Primary key, unique ID (UID) of the prompt word template, which increases naturally as prompt word templates are added. Template Content: The main content of the prompt word template, used to concatenate the content of the intent tag and pass it to the large language model for intent recognition. Applicable Label: The default value is empty, indicating that it applies to all unspecified intent tags. In addition, the prompt word template database should contain at least one template with an empty applicable label as the default template for all intent tag recognition to complete the intent recognition function.

[0054] The multi-level intent recognition system uses a large language model to identify the intent of user input requests. After the user enters a request, the multi-level intent recognition system will determine the intent label of the current request step by step until it reaches the lowest level intent label, or no obvious intent label.

[0055] At each level of judgment, the multi-level intent recognition system will extract the corresponding prompt word template from the prompt word template database based on the results of the intent label recognition of the previous level, and input it into the large language model for intent recognition; after obtaining the result of intent recognition, if the current result is "no obvious intent recognition", the intent recognition stops, otherwise it continues to the next level of recognition until the bottom-level label is recognized. When the recognition result stops at "no obvious intent recognition", it is necessary to check whether there is a higher-level intent recognition result for the current request: if it exists, the intent of the current request is the higher-level recognition result; if it does not exist, it means that the current request has no obvious intent. At this time, the system will introduce historical intent records for judgment: if there is a historical intent label record, the historical intent label will be used as the current intent recognition label; if there is no historical intent label record, the current request is ultimately determined to have no obvious intent.

[0056] Specifically, the identification process steps are as follows: Figure 4 As shown in the figure, (1) the user inputs a document request; (2) the multi-level intent recognition system recognizes the intent tag of the document; (3) if the intent tag of the document is recognized and the intent does not have any sub-intent, the recognition result is the final intent of the document; (4) if the intent tag of the document is recognized, but the intent still has sub-intent, then the sub-level intent recognition is continued; (5) if the obvious intent tag of the document is not recognized and there is an identified parent intent, then the parent intent is the final intent of the text; (6) if the obvious intent tag is not recognized and there is no identified parent intent, then the current text is temporarily determined to have no obvious intent; when the intent tag recognition result of the current user request is temporarily determined to have no obvious intent, the historical intent recognition tag is considered; if there is a historical intent recognition tag, then the intent recognition tag of the current document is the historical intent; if there is no historical intent recognition tag, then the current text is defined as having no obvious intent.

[0057] By identifying intent labels at each level, the user's input intent is analyzed hierarchically from macro to micro. The model uses semantic vector embedding (such as BERT and Sentence Transformers) and combines it with conversation history information to dynamically adjust classification results. This historical information is incorporated into intent label recognition, enabling understanding of multi-round conversation content and accurately identifying and assigning subtasks.

[0058] In addition, through the identification of step-by-step intent labels, complex user needs can be decomposed into multiple subtasks, ensuring a complete understanding of complex conversations and achieving the effect of complex intent decomposition; by combining context analysis and deep learning models, the misjudgment rate of intent recognition can be effectively reduced and the accuracy rate can be improved; it supports the simultaneous processing of multiple intents, providing a clear basis for task allocation and strategy execution, and adapting to multi-task interaction.

[0059] like Figure 5 The flowchart of the retrieval enhancement generation module shown in the figure forms an identification result after the multi-level intent tag recognition module completes the intent recognition of the user request; first, the document slice vector data of the corresponding document slice is extracted from the knowledge base document database; second, the user's request is vectorized; third, the cosine similarity between the document slice vector and the user request vectorized data is compared, and the document slice with the highest similarity is selected as the most similar document slice; fourth, the most similar document slice and the user's request vector are spliced ​​to form a spliced ​​question; fifth, the spliced ​​question is passed to the large language model, and the large language model generates the corresponding answer result; finally, the user's request vector and the answer result generated by the large language model are returned to the historical conversation record form and archived as the historical record of the current conversation, so as to facilitate the subsequent splicing of the question input, historical conversation record and the results of the large language model retrieval, and together with the historical conversation record and the user's request vector to form the splicing of the input question, and then call the large language model to generate the answer.

[0060] The retrieval-enhanced generation module uses semantic vector-based search technology to extract the most relevant information from knowledge base documents. It combines user input with search results using large language models (such as GPT, LLaMa, and Qwen) to generate contextually relevant and accurate answers. It supports multi-round conversations and dynamically integrates historical user input to improve the coherence of generated content.

[0061] Combining the dual capabilities of retrieval and generation, the generated responses are both knowledge-based and linguistically fluent. They excel in answering open-ended questions, providing insightful responses. Generated results are continuously optimized throughout multiple rounds of conversation, enhancing the user experience.

[0062] The retrieval enhancement generation module uses a large language model call function: it integrates mainstream large language models (such as GPT-4 and Claude) and enables dynamic model call through an API. It selects the appropriate model based on task requirements and uses a task decomposition mechanism to assign complex tasks to different models, achieving modular development and high concurrency support.

[0063] LLM has significant advantages in complex language understanding and generation, ensuring the accuracy and fluency of system-generated results. Selecting the appropriate model based on task requirements reduces unnecessary resource consumption.

[0064] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A customized multi-level intent recognition retrieval enhancement generation system, characterized by: It includes a multi-level intent label and document management module, a multi-level intent recognition module and a retrieval enhancement generation module; among them, the multi-level intent label and document management module can be customized, and according to the user's definition of the scenario, multi-level intent labels are established step by step. According to the user settings, the correspondence between intent labels and documents is established to form a label-document correspondence table; the multi-level intent recognition module identifies the user's intention step by step according to the multi-level intent labels established by the multi-level intent label and document management module, and gives the corresponding intent label; the retrieval enhancement generation module performs a vector similarity search in the document range circled by the intent label formed in the multi-level intent recognition module, and passes the results of the vector similarity search to the large language model for result generation.

2. The customized multi-level intent recognition retrieval enhancement generation system according to claim 1, characterized in that: The implementation steps of the multi-level intent label and document management module are as follows: (1) the user adds intent labels step by step according to the needs of the scenario to form an intent label form, and each intent label has a corresponding number in the intent label form to form an intent label number; (2) the user prepares the corresponding knowledge base document according to the needs of the scenario, vectorizes the knowledge base document, obtains the document slice vector, stores the document slice vector in the document vector database, and each document slice vector has a corresponding document vector number; (3) the user marks each knowledge base document with the corresponding intent label according to the added intent label, and the intent label number and the document vector number correspond one to one, and are added to the document to form a label-document correspondence relationship form.

3. The customized multi-level intent recognition retrieval enhancement generation system according to claim 1 is characterized by: The multi-level intent label recognition module includes a prompt word template database and a multi-level intent recognition system. The prompt word template database maintains the prompt word templates of users in different intent scenarios. The multi-level intent recognition system uses a large language model as the base model for intent recognition. Through the predetermined prompt word template database, it splices the contents of the multi-level intent labels and the intent labels in the document management module, and inputs the spliced ​​content and user requests as a splicing problem into the large language model for intent recognition.

4. The customized multi-level intent recognition retrieval enhancement generation system according to claim 3, characterized in that: The steps of the intent recognition process are as follows: (1) The user inputs a document request; (2) The multi-level intent recognition system identifies the intent tag of the document; (3) If the intent tag of the document is identified and the intent does not have any sub-intents, the recognition result is the final intent of the document; (4) If the intent tag of the document is identified, but the intent still has sub-intents, the sub-level intent recognition is continued; (5) If the obvious intent tag of the document is not identified and there is an identified parent intent, the parent intent is the final intent of the document. (6) If no obvious intent label is identified and there is no identified parent intent, it is currently tentatively classified as having no obvious intent; When the intent tag recognition result of the current user request is tentatively determined to have no obvious intent, the historical intent recognition tags are considered; If there is a historical intent recognition tag, the intent recognition tag of the current document is the historical intent; If there is no historical intent identification label, the current text is defined as having no obvious intent.

5. The customized multi-level intent recognition retrieval enhancement generation system according to claim 1 is characterized by: The implementation steps of the retrieval enhancement generation module are as follows: (1) extracting the document slice vector data corresponding to the document slice from the knowledge base document database; (2) vectorizing the user's request; (3) comparing the cosine similarity between the document slice vector and the user request vectorized data, and selecting the document slice with the highest similarity as the most similar document slice; (4) splicing the most similar document slice and the user's request vector to form a spliced ​​question; (5) passing the spliced ​​question to the large language model, and letting the large language model generate the corresponding answer result; (6) returning the user's request vector and the answer result generated by the large language model to the historical conversation record form, as a historical record of the current conversation, or as a subsequent splicing with the spliced ​​question input, historical conversation record and large language model retrieval result, together with the historical conversation record and the user's request vector to form the splicing of the input question, and calling the large language model to generate the answer.

6. The customized multi-level intent recognition retrieval enhancement generation system according to claim 2, characterized in that: The knowledge base documents are vectorized using Llama vectorization, and the document vector database uses the FAISS database.

7. A method for generating a retrieval enhancement system comprising the customized multi-level intent recognition system described in any one of claims 1-6.

Images