Knowledge retrieval enhancement-based special agent implementation method and system

Abstract

The invention relates to the technical field of information, in particular to an implementation method and system of a specialized agent based on knowledge retrieval enhancement, accurate knowledge retrieval is achieved by using a BGE-ES-Reranker model, and the BGE-ES-Reranker model is used for reordering retrieved documents in an RAG system; the method has the beneficial effects that the knowledge is divided into the knowledge bases of different levels, and different weights are configured for each knowledge base, so that the retrieval strategy can be automatically adjusted according to the query content, and relevant information can be quickly and accurately recalled in mass data.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method and system for implementing a specialist intelligent agent based on knowledge retrieval enhancement. Background Art

[0002] In today's information age, people are faced with vast and complex amounts of data and information. Quickly and accurately obtaining the information they need has become crucial. The presence of redundant information and information noise has intensified people's demand for personalized, highly accurate information. Therefore, efficient and intelligent information acquisition tools play a vital role in this era, helping people quickly filter out truly valuable content. Intelligent agents based on knowledge-enhanced search utilize natural language interaction. Compared to the keyword matching of traditional search engines, these systems can dynamically adjust answers during conversations, more accurately capturing user needs and satisfying complex query intent.

[0003] Existing document retrieval and question-answering systems often rely on search engine natural language processing technology. Such dialogue system frameworks, such as Elasticsearch (ES), can store and retrieve both structured and unstructured data, and retrieve information through methods such as keywords. For example, Elasticsearch-based search engines efficiently search massive amounts of text data using an inverted index, supporting full-text search and multiple query syntaxes. However, ES cannot deeply understand the natural language questions asked in conversations based on keyword matching, making it difficult to adapt intelligent search strategies to meet complex and diverse query requirements. Another example is the OpenAI ChatGPT text generation model, which excels in language modeling, providing flexible capabilities and the ability to generate relevant answers based on different user questions. However, this language model cannot provide information about recent events or identify user queries in specific fields. Therefore, it struggles to effectively address specific problems in certain fields and cannot achieve incremental knowledge updates. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for implementing a specialist agent based on knowledge retrieval enhancement to solve the problems raised in the above background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for implementing a specialist agent based on knowledge retrieval enhancement, utilizing the BGE-ES-Reranker model to achieve accurate knowledge retrieval. The BGE-ES-Reranker model is used to rerank retrieved documents in a RAG system, specifically comprising the following steps:

[0006] The document is processed and segmented into segments suitable for processing, which are converted into vector form through the embedding model and stored in the vector database;

[0007] When a user asks a question, the question is converted into a vector and matched with the document vectors in the database to find the most relevant document set;

[0008] The initially retrieved documents are fed into the Reranker model for further evaluation and ranking.

[0009] Preferably, different knowledge is divided into knowledge bases at different levels. The knowledge base is divided into three different levels: company level, project level and session level, and is assigned different weights. When querying, relevant knowledge information is matched according to the weight in the knowledge base, specifically: Company-level knowledge base: contains all the company's core documents, policies, and technical documents, with a lower weight, and is called first when the problem involves general knowledge at the company level; Project-level knowledge base: contains documents, progress reports, and technical solutions related to specific projects, with a relatively high weight. For queries related to the current or specific project, information is obtained from the knowledge base at this level first; Session-level knowledge base: contains documents or content involved in the current session, with the highest weight, and is closely related to the context of the current session, providing the most accurate answer to the current specific problem.

[0010] Preferably, knowledge bases at different levels reflect weight information by configuring their own Reranker parameters and the number of knowledge fragments obtained from each knowledge base. Specifically: according to the role of knowledge bases at different levels in the retrieval task, their weights are dynamically adjusted, and different levels of knowledge bases are configured with different Reranker parameters according to their importance in the current query. These parameters affect the behavior of the retrieval and reranking process; for session-level and project-level knowledge bases with higher weights, more knowledge fragments are set to be obtained to more comprehensively cover the query content; for company-level knowledge bases with low weights, the number of knowledge fragments obtained is reduced to avoid excessive impact on the final retrieval results.

[0011] Preferably, before the results of knowledge retrieval are handed over to the large model for understanding, file deduplication operations are performed. Specifically, the following steps are performed: determine whether the files are duplicated by comparing the file names and the path information in MinIO; first check whether the name of each file is duplicated with the retrieved file. If the file names are the same, further compare them by the path in MinIO; if the file path information and file name are consistent, it is determined that the file already exists and no repeated processing is required, ensuring that the data set input to the large model is concise and accurate.

[0012] Preferably, before the results of knowledge retrieval are submitted to the large model for understanding, a relevance filtering operation is performed. Specifically, after retrieving a series of documents, these documents are matched with the user's query by constructing a specific prompt to evaluate the relevance of the documents; the retrieved content and the assembled prompt are input into the large model, and the model outputs the final answer.

[0013] A system for implementing a method for a specialist agent based on knowledge retrieval enhancement includes a knowledge retrieval module that uses a BGE-ES-Reranker model to implement accurate knowledge retrieval. The BGE-ES-Reranker model is used to rerank retrieved documents in a RAG system.

[0014] The knowledge retrieval module further includes:

[0015] The document preprocessing submodule is used to process and segment documents into segments suitable for processing, convert the segmented segments into vector form through the embedding model, and store them in the vector database;

[0016] Question vector conversion submodule: When a user asks a question, this submodule converts the question into a vector;

[0017] The document matching submodule is used to match the question vector with the document vector in the vector database to find the most relevant document set;

[0018] The reranking submodule sends the initially retrieved documents into the Reranker model for further evaluation and ranking.

[0019] Preferably, a knowledge retrieval module is included, which utilizes a BGE-ES-Reranker model to achieve accurate knowledge retrieval, wherein the BGE-ES-Reranker model is used to rerank retrieved documents in the RAG system;

[0020] The knowledge retrieval module further includes:

[0021] The document preprocessing submodule is used to process and segment documents into segments suitable for processing, convert the segmented segments into vector form through the embedding model, and store them in the vector database;

[0022] Question vector conversion submodule: When a user asks a question, this submodule converts the question into a vector;

[0023] The document matching submodule is used to match the question vector with the document vector in the vector database to find the most relevant document set;

[0024] The reranking submodule sends the initially retrieved documents into the Reranker model for further evaluation and ranking.

[0025] Preferably, the system further includes a knowledge base division and management module, which divides different knowledge into different levels of knowledge bases. The knowledge base is divided into three different levels: company level, project level, and session level, and is assigned different weights. When querying, relevant knowledge information is matched according to the weights in the knowledge base.

[0026] The knowledge base division and management module specifically includes:

[0027] The company-level knowledge base submodule contains all the company's core documents, policies, and technical documents. It has a lower weight and is called first when the problem involves general knowledge at the company level;

[0028] The project-level knowledge base submodule contains documents, progress reports, and technical solutions related to a specific project. It has a relatively high weight. For queries related to the current or specific project, information is obtained from this level of knowledge base first.

[0029] The session-level knowledge base submodule contains the documents or content involved in the current session. It has the highest weight and is closely related to the context of the current session, providing the most accurate answers to specific questions.

[0030] The weight configuration submodule is used to dynamically configure the weight of each knowledge base according to query requirements.

[0031] Preferably, the knowledge base division and management module further includes a weight configuration submodule, which is used to reflect weight information by configuring the Reranker parameters of different levels of knowledge bases and the number of knowledge fragments obtained from each knowledge base, specifically including:

[0032] The weight dynamic adjustment unit is used to dynamically adjust the weights of different levels of knowledge bases according to their roles in the retrieval task;

[0033] Reranker parameter configuration unit, which is used to configure different Reranker parameters for different levels of knowledge bases according to their importance in the current query. These parameters affect the behavior of the retrieval and reranking process;

[0034] The knowledge fragment number setting unit is used to set the acquisition of more knowledge fragments for higher-weighted session-level and project-level knowledge bases to more comprehensively cover the query content; for low-weighted company-level knowledge bases, the number of acquired knowledge fragments is reduced to avoid excessive impact on the final retrieval results.

[0035] Preferably, a data preprocessing module is also included. The data preprocessing module performs a file deduplication operation before the results of the knowledge retrieval are passed to the large model for understanding. The data preprocessing module specifically includes:

[0036] The file name check submodule is used to check whether the name of each file is repeated with the retrieved file;

[0037] The file path comparison submodule is used to further compare the paths in MinIO when the file names are the same;

[0038] The duplicate file determination submodule is used to determine that a file already exists if the path information and file name are consistent, and no repeated processing is required, thereby ensuring that the data set input into the large model is concise and accurate.

[0039] Preferably, the data preprocessing module further includes a relevance filtering submodule, which performs a relevance filtering operation before the results of the knowledge retrieval are submitted to the large model for understanding, specifically including:

[0040] The prompt construction unit is used to construct a specific prompt after retrieving a series of documents;

[0041] A relevance evaluation unit, configured to match the documents with the user's query through the prompt and evaluate the relevance of the documents;

[0042] The model input and output unit is used to input the retrieved content and the assembled prompt into the large model, and the model outputs the final answer.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] The implementation method and system of the specialist intelligent agent based on knowledge retrieval enhancement proposed in the present invention can automatically adjust the retrieval strategy according to the query content by dividing knowledge into knowledge bases of different levels and configuring different weights for each knowledge base, thereby quickly and accurately recalling relevant information from massive data; combined with large models and sophisticated prompt engineering, it can not only improve the relevance of retrieval results, but also fully understand and combine contextual information when generating answers to meet complex query requirements; for different industries and application scenarios, the present invention can flexibly respond to changing user needs by adjusting the configuration of the knowledge base and reranking model; through effective file deduplication and relevance filtering mechanisms, it ensures that the input data provided to the large model is more streamlined and accurate, thereby improving the quality and reliability of the results generated by the intelligent agent. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0046] In order to clearly and completely describe the objectives and technical solutions of the present invention and make the advantages more clearly understood, the embodiments of the present invention are further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are part of the embodiments of the present invention, not all of them, and are only used to explain the embodiments of the present invention, not to limit 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.

[0047] For example 1, please refer to Figure 1 This invention provides a technical solution: a method for implementing a specialist agent based on enhanced knowledge retrieval, which divides different knowledge into different knowledge bases. The knowledge bases are divided into three different levels: company-level, project-level, and session-level, and each is assigned different weights. During queries, relevant knowledge information is matched based on the weights in the knowledge bases, thereby improving efficiency and accuracy.

[0048] The three types of knowledge bases are as follows:

[0049] 1. Company-level knowledge base: This is the broadest knowledge base, encompassing all core company documents, policies, technical documentation, and more. It carries a lower weighting because its broad scope applies to general or unspecific questions. This level of knowledge base is prioritized only when questions involve general, company-wide knowledge.

[0050] 2. Project-level knowledge base: This level contains documents, progress reports, technical solutions, and other information related to a specific project. This level is weighted higher because it focuses on a specific project and provides more detailed and precise information. For queries related to the current or specific project, information from this level of knowledge base is prioritized.

[0051] 3. Session-level knowledge base: This knowledge base contains documents or content related to the current session and is typically the most immediate and relevant. It carries the highest weight because it is directly tied to the context of the current session. The content in this knowledge base is typically the most flexible and dynamic, providing the most accurate answers to specific questions.

[0052] Knowledge bases at different levels reflect the weight information of each knowledge base by configuring their own Reranker parameters and the number of knowledge fragments obtained from each knowledge base. The core purpose of this configuration strategy is to dynamically adjust the weights of knowledge bases at different levels according to their role in the retrieval task, thereby optimizing the retrieval results. Specifically, knowledge bases at different levels (such as company-level, project-level, and session-level knowledge bases) will configure different Reranker parameters based on their importance in the current query. These parameters affect the behavior of the retrieval and reranking process. In addition to adjusting the Reranker parameters, the number of knowledge fragments obtained from each knowledge base is also an important way to adjust the weight. For knowledge bases with higher weights (such as session-level and project-level knowledge bases), more knowledge fragments can be set to be obtained so that the query content can be covered more comprehensively. On the contrary, for company-level knowledge bases with low weights, the number of knowledge fragments obtained may be reduced accordingly to avoid excessive impact on the final retrieval results.

[0053] Before the results of knowledge retrieval are handed over to the large model for understanding, a series of preprocessing operations must be performed to ensure that the knowledge content finally input is unique and of high quality. First of all, file deduplication is an essential step, which is mainly determined by comparing the file name and the path information in MinIO to determine whether the file is duplicated. Specifically, the system will first check whether the name of each file is duplicated with the retrieved file. If the file name is the same, it will further compare the path in MinIO to confirm whether the file is duplicated in storage. If the path information and file name of the file are consistent, it can be determined that the file already exists and does not need to be processed repeatedly, thereby avoiding the introduction of redundant data in subsequent processing and ensuring that the data set input to the large model is more streamlined and accurate.

[0054] Relevance filtering, achieved through the use of carefully designed prompts, is a crucial step in the large-scale model's understanding phase. Specifically, after retrieving a series of documents, the system constructs specific prompts to match these documents to the user's query, thereby assessing their relevance. Ultimately, the retrieved content and assembled prompts are fed into the large-scale model, which then outputs the final answer.

[0055] Example 2. Based on Example 1, a system for implementing a method of a specialist agent based on knowledge retrieval enhancement is proposed, including a knowledge retrieval module, which uses the BGE-ES-Reranker model to realize accurate knowledge retrieval, and the BGE-ES-Reranker model is used to rerank the retrieved documents in the RAG system; the knowledge retrieval module further includes: a document preprocessing submodule, which is used to process and split the document into fragments suitable for processing, convert the split fragments into vector form through an embedding model, and store them in a vector database; a question vector conversion submodule, when a user asks a question, this submodule converts the question into a vector; a document matching submodule, which is used to match the question vector with the document vector in the vector database to find the most relevant document set; a reranking submodule, which sends the initially retrieved documents into the Reranker model for further evaluation and ranking.

[0056] It also includes a knowledge base division and management module, which divides different knowledge into knowledge bases of different levels. The knowledge base is divided into three different levels: company level, project level and session level, and is assigned different weights. When querying, relevant knowledge information is matched according to the weights in the knowledge base; the knowledge base division and management module specifically includes: a company-level knowledge base sub-module, which contains all the company's core documents, policies, and technical documents, has a lower weight, and is called first when the problem involves general knowledge at the company level; a project-level knowledge base sub-module, which contains documents, progress reports, and technical solutions related to specific projects, has a relatively high weight, and for queries related to the current or specific project, information is obtained from the knowledge base at this level first; a session-level knowledge base sub-module, which contains documents or content involved in the current session, has the highest weight, is closely related to the context of the current session, and provides the most accurate answer to the current specific problem; a weight configuration sub-module, which is used to dynamically configure the weights of each knowledge base according to query requirements.

[0057] The knowledge base division and management module also includes a weight configuration submodule, which is used to reflect the weight information by configuring the Reranker parameters of knowledge bases at different levels and the number of knowledge fragments obtained from each knowledge base. Specifically, it includes: a weight dynamic adjustment unit, which is used to dynamically adjust the weights of knowledge bases at different levels according to their role in the retrieval task; a Reranker parameter configuration unit, which is used to configure different Reranker parameters for knowledge bases at different levels according to their importance in the current query. These parameters affect the behavior of the retrieval and reranking process; a knowledge fragment number setting unit, which is used to set more knowledge fragments for session-level and project-level knowledge bases with higher weights to cover the query content more comprehensively; for company-level knowledge bases with low weights, the number of knowledge fragments obtained is reduced to avoid excessive impact on the final retrieval results.

[0058] It also includes a data preprocessing module, which performs file deduplication before the results of knowledge retrieval are handed over to the large model for understanding. The data preprocessing module specifically includes: a file name check submodule, which is used to check whether the name of each file is repeated with the retrieved file; a file path comparison submodule, which is used to further compare the paths in MinIO when the file names are the same; a duplicate file judgment submodule, which is used to determine that the file already exists if the path information and file name are consistent, and no repeated processing is required, ensuring that the data set input to the large model is concise and accurate.

[0059] The data preprocessing module also includes a relevance filtering submodule, which performs a relevance filtering operation before the results of the knowledge retrieval are handed over to the large model for understanding. Specifically, it includes: a prompt construction unit, which is used to construct a specific prompt after retrieving a series of documents; a relevance evaluation unit, which is used to match these documents with the user's query through the prompt and evaluate the relevance of the documents; a model input and output unit, which is used to input the retrieved content and the assembled prompt into the large model, and the model outputs the final answer.

[0060] 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 method for implementing a specialist agent based on knowledge retrieval enhancement, characterized by: The BGE-ES-Reranker model is used to realize accurate knowledge retrieval. The BGE-ES-Reranker model is used to rerank the retrieved documents in the RAG system. The following steps are involved: The document is processed and segmented into segments suitable for processing, which are converted into vector form through the embedding model and stored in the vector database; When a user asks a question, the question is converted into a vector and matched with the document vectors in the database to find the most relevant document set; The initially retrieved documents are fed into the Reranker model for further evaluation and ranking.

2. The method for implementing a specialist agent based on knowledge retrieval enhancement according to claim 1, characterized in that: Different knowledge is divided into knowledge bases at different levels. The knowledge base is divided into three different levels: company level, project level and session level, and is assigned different weights. When querying, relevant knowledge information is matched according to the weight in the knowledge base. Specifically: Company-level knowledge base: contains all the company's core documents, policies, and technical documents, with a lower weight. It is called first when the problem involves general knowledge at the company level; Project-level knowledge base: contains documents, progress reports, and technical solutions related to specific projects, with a relatively high weight. For queries related to the current or specific project, information is obtained from the knowledge base at this level first; Session-level knowledge base: contains documents or content involved in the current session, with the highest weight, and is closely related to the context of the current session, providing the most accurate answers to current specific questions.

3. The method for implementing a specialist agent based on knowledge retrieval enhancement according to claim 2, characterized in that: Knowledge bases at different levels reflect weight information by configuring their own Reranker parameters and the number of knowledge fragments obtained from each knowledge base. Specifically: according to the role of knowledge bases at different levels in the retrieval task, their weights are dynamically adjusted. Knowledge bases at different levels are configured with different Reranker parameters based on their importance in the current query. These parameters affect the behavior of the retrieval and reranking processes; for session-level and project-level knowledge bases with higher weights, more knowledge fragments are set to be obtained to more comprehensively cover the query content; for company-level knowledge bases with low weights, the number of knowledge fragments obtained is reduced to avoid excessive impact on the final retrieval results.

4. The method for implementing a specialist agent based on knowledge retrieval enhancement according to claim 3, characterized in that: Before the results of knowledge retrieval are passed to the large model for understanding, file deduplication is performed. Specifically, the file name is compared with the path information in MinIO to determine whether the file is duplicated. First, the name of each file is checked to see if it is duplicated with the retrieved file. If the file name is the same, the path in MinIO is further compared. If the file path information and file name are consistent, it is determined that the file already exists and no repeated processing is required, ensuring that the dataset input to the large model is concise and accurate.

5. The method for implementing a specialist agent based on knowledge retrieval enhancement according to claim 4, characterized in that: Before the results of knowledge retrieval are passed to the large model for understanding, a relevance filtering operation is performed. Specifically, after retrieving a series of documents, these documents are matched with the user's query by constructing specific prompts to evaluate the relevance of the documents; the retrieved content and the assembled prompts are input into the large model, and the model outputs the final answer.

6. A system for implementing the method for a specialist agent based on enhanced knowledge retrieval according to claim 5, characterized in that: The invention comprises a knowledge retrieval module, wherein the knowledge retrieval module realizes accurate knowledge retrieval by using a BGE-ES-Reranker model, wherein the BGE-ES-Reranker model is used to rerank retrieved documents in a RAG system; The knowledge retrieval module further includes: The document preprocessing submodule is used to process and segment documents into segments suitable for processing, convert the segmented segments into vector form through the embedding model, and store them in the vector database; Question vector conversion submodule: When a user asks a question, this submodule converts the question into a vector; The document matching submodule is used to match the question vector with the document vector in the vector database to find the most relevant document set; The reranking submodule sends the initially retrieved documents into the Reranker model for further evaluation and ranking.

7. A system according to claim 6, characterized in that: It also includes a knowledge base division and management module, which divides different knowledge into different levels of knowledge bases. The knowledge base is divided into three different levels: company level, project level, and session level, and assigned different weights. When querying, relevant knowledge information is matched according to the weights in the knowledge base; The knowledge base division and management module specifically includes: The company-level knowledge base submodule contains all the company's core documents, policies, and technical documents. It has a lower weight and is called first when the problem involves general knowledge at the company level; The project-level knowledge base submodule contains documents, progress reports, and technical solutions related to a specific project. It has a relatively high weight. For queries related to the current or specific project, information is obtained from this level of knowledge base first. The session-level knowledge base submodule contains the documents or content involved in the current session. It has the highest weight and is closely related to the context of the current session, providing the most accurate answers to specific questions. The weight configuration submodule is used to dynamically configure the weight of each knowledge base according to query requirements.

8. A system according to claim 7, characterized in that: The knowledge base partitioning and management module also includes a weight configuration submodule, which is used to reflect weight information by configuring the Reranker parameters of different levels of knowledge bases and the number of knowledge fragments obtained from each knowledge base. Specifically, it includes: The weight dynamic adjustment unit is used to dynamically adjust the weights of different levels of knowledge bases according to their roles in the retrieval task; Reranker parameter configuration unit, which is used to configure different Reranker parameters for different levels of knowledge bases according to their importance in the current query. These parameters affect the behavior of the retrieval and reranking process; The knowledge fragment number setting unit is used to set the acquisition of more knowledge fragments for higher-weighted session-level and project-level knowledge bases to more comprehensively cover the query content; for low-weighted company-level knowledge bases, the number of acquired knowledge fragments is reduced to avoid excessive impact on the final retrieval results.

9. A system according to claim 8, characterized in that: It also includes a data preprocessing module, which performs file deduplication before the results of knowledge retrieval are passed to the large model for understanding. The data preprocessing module specifically includes: The file name check submodule is used to check whether the name of each file is repeated with the retrieved file; The file path comparison submodule is used to further compare the paths in MinIO when the file names are the same; The duplicate file determination submodule is used to determine that a file already exists if the path information and file name are consistent, and no repeated processing is required, thereby ensuring that the data set input into the large model is concise and accurate.

10. A system according to claim 9, characterized in that: The data preprocessing module also includes a relevance filtering submodule. The relevance filtering submodule performs relevance filtering operations before the results of knowledge retrieval are passed to the large model for understanding. Specifically, it includes: The prompt construction unit is used to construct a specific prompt after retrieving a series of documents; A relevance evaluation unit, configured to match the documents with the user's query through the prompt and evaluate the relevance of the documents; The model input and output unit is used to input the retrieved content and the assembled prompt into the large model, and the model outputs the final answer.

Images