Agentic workflow-based scientific research team information question and answer agent construction method

Abstract

The invention provides a scientific research team information question and answer agent construction method based on Agentic workflow, and relates to the technical field of intelligent question and answer. The method comprises the following steps: acquiring multi-source heterogeneous scientific research data, and constructing a plurality of vertical domain exclusive RAG knowledge bases through text partitioning, vectorization and clustering; a large model management and Agentic workflow platform is deployed locally, and a directed acyclic graph workflow including intention recognition, cue word engineering, knowledge retrieval, response generation and feedback records is constructed. The system utilizes lightweight text classification and a dynamic weight mechanism to carry out team and project routing, combines multi-strategy RAG retrieval and multiple models to cooperatively generate natural language answers, constructs dialogue logs based on user behaviors and satisfaction, dynamically updates the confidence of each knowledge base, and improves the user experience. Therefore, the self-adaptive high-quality question and answer service oriented to the scientific research team scene is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent question answering technology, and in particular to a method for constructing an intelligent agent for information question answering in research teams based on Agentic workflow. Background Technology

[0002] With the rapid development of artificial intelligence, especially generative large-scale models, AI technology is being applied more and more widely in scientific research and team management. As the core entities of scientific research activities, research teams often need to manage and retrieve large amounts of experimental data, technical documents, equipment records, and research results. However, traditional manual backup, classification, and retrieval methods are inefficient, easily leading to information omissions and knowledge silos. This makes it difficult to meet the demands of modern scientific research for efficient, intelligent, and highly dynamic information processing, and is detrimental to the continuity of research directions within research teams and the personalized training of graduate students.

[0003] In recent years, intelligent information retrieval methods based on the RAG (Retrieval-Augmented Generation) model have gradually become an effective means to improve knowledge management. This method transforms team technical documents and experimental data into a vectorized knowledge base, achieving high-precision semantic matching and content generation. Compared to traditional keyword retrieval, the RAG system combines semantic understanding and generation capabilities to provide researchers with a more accurate and natural knowledge question-and-answer experience.

[0004] However, most existing RAG systems still use a single model or a general knowledge base, which cannot provide targeted question answering based on team type or research direction. When faced with different research tasks, general RAGs often lack the ability to adapt to professional context and are static retrieval, resulting in large answer biases, low practicality, and an inability to adapt to the dynamic research process of a team. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a method for constructing an intelligent agent for information question answering in research teams based on Agentic workflow. This invention solves the problems of low retrieval accuracy and lack of intelligent collaboration in existing research team document management and knowledge question answering systems.

[0006] To achieve the above objectives, the present invention provides the following solution: A method for constructing an information question-answering intelligent agent for research teams based on Agentic workflow, comprising: Acquire multi-source heterogeneous data and construct multiple vertical domain-specific RAG knowledge bases based on the multi-source heterogeneous data; Construct a directed acyclic graph workflow with multiple functional nodes and associate it with the RAG knowledge bases specific to the multiple vertical domains; Based on the multi-node question-answering workflow, intent recognition is performed on the user's input query and historical dialogue context. The query is then input into a lightweight text classification model to determine the research team and research direction. Based on a dynamic weight allocation mechanism, the routing decision result of the target-specific RAG knowledge base corresponding to the query is obtained. Based on the routing decision results, multi-strategy RAG retrieval and multi-model collaborative generation are performed on the query to obtain the final natural language answer; Collect metadata and behavioral characteristics of user interactions, and quantify multi-dimensional indicators such as copy and use of answers, affirmative or negative actions, session depth and page dwell time to obtain user satisfaction evaluation results that reflect the current interaction quality. Based on the user satisfaction evaluation results, the user queries, enhanced prompt information, retrieved knowledge fragments, natural language answers, user satisfaction and explicit feedback are recorded in a structured manner and stored in a relational database to form a dialogue log dataset. Based on the dialogue log dataset and the corresponding user satisfaction evaluation results, dynamic confidence parameters are maintained for each dedicated RAG knowledge base. The confidence of each dedicated RAG knowledge base is updated according to the preset confidence update rules, so as to be used for the retrieval and sorting of dedicated RAG knowledge bases in different vertical fields during subsequent question answering.

[0007] Preferably, the multi-source heterogeneous data includes: Academic papers, experimental reports, technical manuals, project documents, and algorithm code.

[0008] Preferably, multi-source heterogeneous data is acquired, and multiple vertical domain-specific RAG knowledge bases are constructed based on the multi-source heterogeneous data, including: The multi-source heterogeneous data is preprocessed to obtain preprocessed multi-source heterogeneous data; Text extraction and segmentation are performed on the preprocessed multi-source heterogeneous data to uniformly convert data from different sources and in different formats into several text fragments; Each of the text segments is vectorized and encoded to obtain a corresponding high-dimensional vector representation, and the high-dimensional vector and the corresponding associated original text segment are stored in a vector database; Dimensionality reduction and clustering are performed on the high-dimensional vectors stored in the vector database. Based on the clustering results, multiple semantically similar vector clusters are determined, and representative classification labels are extracted for each vector cluster. Based on each vector cluster and its corresponding classification label, the associated text fragments are organized and categorized to construct vertical domain-specific RAG knowledge bases for different research teams and research topics.

[0009] Preferably, the plurality of functional nodes include: Intent recognition, prompt word engineering, knowledge retrieval, response generation, and feedback recording.

[0010] Preferably, the lightweight text classification model is the ALBERT lightweight text classifier.

[0011] Preferably, the process of performing intent recognition on the user-input query and historical dialogue context based on the multi-node question-answering workflow, inputting the query into a lightweight text classification model to determine the research team and research direction, and obtaining the routing decision result of the target-specific RAG knowledge base corresponding to the query according to a dynamic weight allocation mechanism includes: Obtain the query currently entered by the user and combine it with the user's historical dialogue context in this session to form the input text of the intent to be identified; The input text is fed into a lightweight text classification model to classify and predict the research team and research direction to obtain candidate research team and research direction labels. Based on the candidate research teams and research direction tags, a number of candidate vertical domain-specific RAG knowledge base sets are determined. Based on the preset dynamic weight allocation mechanism, the similarity information between the current input text and the representative text vectors in each candidate vertical domain exclusive RAG knowledge base, as well as the historical interaction records or knowledge base confidence parameters, are comprehensively considered to calculate the weight score of each candidate vertical domain exclusive RAG knowledge base. The candidate vertical domain-specific RAG knowledge bases are sorted according to the weight scores, and one or more vertical domain-specific RAG knowledge bases whose weight scores meet the preset conditions are selected to obtain the routing decision result of the target specific RAG knowledge base corresponding to the query.

[0012] Preferably, the calculation expression for the user satisfaction evaluation result is as follows: ; in, These are the scores for the first indicator, the second indicator, and the third indicator, respectively. These are the weights of the first indicator, the second indicator, and the third indicator, respectively.

[0013] Preferably, the expression corresponding to the preset confidence update rule is: ; in, and The knowledge base confidence scores at time points t and t+1 are respectively. As the attenuation factor, It is a learning rate.

[0014] The present invention discloses the following technical effects: This invention provides a method for constructing an intelligent information question-answering agent for research teams based on Agentic workflow. This invention improves upon single RAG systems by intelligently recognizing user intent in response to user questions and dynamically selecting and activating the most relevant dedicated knowledge base, thus achieving an improvement from general question-answering to team-specific, task-oriented question-answering. A closed-loop, adaptive learning system is designed. By quantifying user satisfaction and feeding it back to the confidence level of the retrieval system, the system can continuously self-calibrate and optimize, solving the problem of traditional static RAG retrieval's difficulty in adaptive optimization. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating a method for constructing an intelligent agent for information question answering in a research team based on Agentic workflow, provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the dialogue workflow provided in an embodiment of the present invention. Detailed Implementation

[0017] 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 embodiments of the present invention, and not all embodiments. 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.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, this invention provides a method for constructing an information question-answering intelligent agent for research teams based on Agentic workflow, including: Step 100: Obtain multi-source heterogeneous data and construct multiple vertical domain-specific RAG knowledge bases based on the multi-source heterogeneous data; Step 200: Construct a directed acyclic graph workflow for multiple functional nodes and associate it with the RAG knowledge bases specific to the multiple vertical domains; Step 300: Based on the multi-node question-answering workflow, perform intent recognition on the user-input query and historical dialogue context, input the query into a lightweight text classification model to determine the research team and research direction, and obtain the routing decision result of the target-specific RAG knowledge base corresponding to the query according to the dynamic weight allocation mechanism. Step 400: Based on the routing decision result, perform multi-strategy RAG retrieval and multi-model collaborative generation on the query to obtain the final natural language answer; Step 500: Collect metadata and behavioral characteristics of user interactions, and quantify multi-dimensional indicators such as copy and use of answers, affirmative or negative operations, session depth and page dwell time to obtain user satisfaction evaluation results that reflect the current interaction quality. Step 600: Based on the user satisfaction evaluation results, the user query, prompt enhancement information, retrieved knowledge fragments, natural language answers, user satisfaction and explicit feedback are recorded in a structured manner and stored in a relational database to form a dialogue log dataset; Step 700: Based on the dialogue log dataset and the corresponding user satisfaction evaluation results, maintain dynamic confidence parameters for each dedicated RAG knowledge base, and update the confidence of each dedicated RAG knowledge base according to the preset confidence update rules, so as to use it for retrieval and sorting of dedicated RAG knowledge bases in different vertical fields during subsequent question answering.

[0020] Specifically, such as Figure 2 As shown, the workflow is as follows: First, the user asks a question. The system identifies the user's intent and sends the identification result to the question classification module. The question classification module categorizes the question according to the intent and uses the classification result to update the database. Subsequently, the system generates corresponding prompt words based on the intent identification result, retrieves relevant knowledge from the local knowledge base, and adjusts the retrieval results based on the updated knowledge base confidence. After obtaining the prompt words and relevant knowledge, the system calls a generative large model to generate an answer and outputs the generated result. The interaction results between the user and the system are then fed back to update the knowledge base confidence, thereby continuously optimizing the retrieval quality and question-answering effect in subsequent local knowledge base retrievals, forming a closed-loop intelligent question-answering process.

[0021] Furthermore, on the Linux system, Docker containerization technology is used to deploy the Ollama local large model management framework and the Dify AI application development platform. Preferably, Ollama is used to deploy open-source large models with moderate parameter counts, such as Qwen2-7B or Llama3-8B, as a base to ensure both domain-specific understanding and hardware compatibility for local deployment. The Dify platform is used for visual building and orchestration of subsequent agentic workflows.

[0022] Furthermore, multi-source heterogeneous data is acquired, and multiple vertical domain-specific RAG knowledge bases are constructed based on the multi-source heterogeneous data, including: The multi-source heterogeneous data is preprocessed to obtain preprocessed multi-source heterogeneous data; Text extraction and segmentation are performed on the preprocessed multi-source heterogeneous data to uniformly convert data from different sources and in different formats into several text fragments; Each of the text segments is vectorized and encoded to obtain a corresponding high-dimensional vector representation, and the high-dimensional vector and the corresponding associated original text segment are stored in a vector database; Dimensionality reduction and clustering are performed on the high-dimensional vectors stored in the vector database. Based on the clustering results, multiple semantically similar vector clusters are determined, and representative classification labels are extracted for each vector cluster. Based on each vector cluster and its corresponding classification label, the associated text fragments are organized and categorized to construct vertical domain-specific RAG knowledge bases for different research teams and research topics.

[0023] Specifically, the team collected multi-source heterogeneous data, including but not limited to academic papers, experimental reports, technical manuals, project documents, and algorithm code. Using the bge-large-zh-v1.5 embedding model that supports both Chinese and English, the text was chunked and transformed into 1024-dimensional high-dimensional vectors, which were then stored in the Chroma vector database. UMAP was used to reduce the dimensionality of the high-dimensional vectors, and HDBSCAN was used for density clustering to automatically discover natural categories in the data (such as "data visualization methods," "lithium batteries," and "catalyst synthesis"). For each cluster, its core keywords were extracted as the category name for that knowledge base, thus constructing dedicated RAG knowledge bases for multiple vertical domains.

[0024] Furthermore, within the Dify platform, a directed acyclic graph workflow is created that includes multiple functional nodes such as intent recognition, prompt word engineering, knowledge retrieval, response generation, and feedback logging. Multiple dedicated RAG knowledge bases are associated with this workflow, and corresponding vector database connection parameters are configured.

[0025] Furthermore, the lightweight text classification model is the ALBERT lightweight text classifier.

[0026] Furthermore, based on the multi-node question-answering workflow, intent recognition is performed on the user-input query and historical dialogue context. The query is then input into a lightweight text classification model to determine the research team and research direction. Based on a dynamic weight allocation mechanism, a routing decision result for the target-specific RAG knowledge base corresponding to the query is obtained, including: Obtain the query currently entered by the user and combine it with the user's historical dialogue context in this session to form the input text of the intent to be identified; The input text is fed into a lightweight text classification model to classify and predict the research team and research direction to obtain candidate research team and research direction labels. Based on the candidate research teams and research direction tags, a number of candidate vertical domain-specific RAG knowledge base sets are determined. Based on the preset dynamic weight allocation mechanism, the similarity information between the current input text and the representative text vectors in each candidate vertical domain exclusive RAG knowledge base, as well as the historical interaction records or knowledge base confidence parameters, are comprehensively considered to calculate the weight score of each candidate vertical domain exclusive RAG knowledge base. The candidate vertical domain-specific RAG knowledge bases are sorted according to the weight scores, and one or more vertical domain-specific RAG knowledge bases whose weight scores meet the preset conditions are selected to obtain the routing decision result of the target specific RAG knowledge base corresponding to the query.

[0027] Specifically, the process begins by receiving user input (i.e., the query) from the start node. Based on the user input, the query is routed to the intent recognition node. This intent recognition node is implemented using the ALBERT lightweight text classifier and encapsulates the model as an independent API service using the FastAPI framework. Deployed in a Docker container, the intent recognition node categorizes the user input into predefined team or research topics (e.g., 'a query belonging to Team A regarding catalyst stability'). This node then integrates the current query with the historical dialogue context, generating dynamic weights through an attention weight network. It then determines one or more of the most likely relevant dedicated RAG knowledge bases for subsequent retrieval and makes a final decision.

[0028] Furthermore, the user input is routed to the corresponding prompt word generation node. This node uses a large model instance with fewer parameters, Qwen2-1.5B, to select a predefined prompt word template based on the domain characteristics of the target knowledge base (e.g., 'You are a materials science expert, please answer the question about catalyst stability based on the following knowledge fragments...'), thereby enhancing the user intent. The enhanced user input is then sent to the knowledge base retrieval node, which retrieves the Top-K most relevant knowledge fragments from the corresponding vector library. Finally, a large model instance with even more parameters, Qwen2-7B-Instruct, serves as the answer generation node, combining the retrieved knowledge fragments and the enhanced user input to generate the final natural language answer.

[0029] Furthermore, metadata is extracted from user input, including but not limited to the user's team affiliation, the research topic inquired about (for fine-grained knowledge base optimization), and user satisfaction (reflecting sentiment and used for dynamic updates to knowledge base confidence). To consistently obtain user satisfaction, an implicit feedback mechanism is employed. Considering that users rarely provide explicit satisfaction feedback after each interaction (i.e., clicking the "agree" or "disagree" button after an answer), implicit behavioral analysis is used to weight and determine the user's satisfaction at time point t based on the following formula. : The formula for calculating the user satisfaction evaluation result is as follows: ; in, These are the scores for the first indicator, the second indicator, and the third indicator, respectively. These are the weights of the first indicator, the second indicator, and the third indicator, respectively.

[0030] (1) Explicit adoption behavior ( =0, =0.5): Whether the user copied the answer, used affirmative words such as "okay" or "thank you" after it, or clicked the "agree" button under the answer. If any of the above behaviors is "yes", it is considered satisfactory. If "No", it is considered that there is no special feedback. If you click the "Oppose" button, it will be considered that you are not satisfied. ).

[0031] (2) Interaction depth ( : Does the user end the session after receiving a response? If yes, it is considered satisfactory. If there is no further interaction, it is considered as no special feedback. If follow-up questions or restatement of the question are asked, it is considered unsatisfactory. ).

[0032] (3) Duration of stay ( A very short stay on the results page may indicate that the answer is irrelevant; if the stay time is >= 60 seconds, then... If the dwell time is between 30 and 60 seconds, then If the dwell time is less than or equal to 30 seconds and the answer is not copied, then .

[0033] The system structurally records and stores the complete context of each user interaction (i.e., question and answer), including the user's question, the prompt words used, the knowledge fragments retrieved, the generated answer, and any subsequent explicit feedback that the user may provide, in a MySQL database to form a complete dialogue log.

[0034] This module is based on stored dialogue records and calculated user satisfaction scores. The confidence level of the knowledge base is dynamically optimized. The system maintains a knowledge base confidence table, in which each dedicated RAG knowledge base has a dynamic confidence level. The confidence level is updated according to the following formula: ; in, and The knowledge base confidence scores at time points t and t+1 are respectively. This is a decay factor, which needs to be set manually, and is used to control the degree to which historical confidence levels are retained. The closer it is to 1, the stronger the system's memory and the more stable the changes. To maintain the system's stability, it is set to 0.95. It is a learning rate, which needs to be manually set, and is used to control the degree of influence of new satisfaction feedback on the confidence level. The larger the value, the more sensitive the system is to new feedback. To avoid large fluctuations, it is set to 0.05.

[0035] At the knowledge retrieval node, when the system searches multiple dedicated RAG knowledge bases in parallel based on routing decisions, the vector database provides an original cosine similarity score for each returned knowledge fragment. To achieve adaptive optimization, the system introduces a confidence-based reordering mechanism: the final ranking score of each knowledge fragment is determined by its original similarity score and the current confidence score of the knowledge base to which the fragment belongs. It is decided jointly. The specific calculation formula is as follows: ; in, This is an amplification factor. The formula ensures that content from high-confidence knowledge bases receives a significant boost in its final ranking score, thus ranking higher in mixed search results and being more likely to be selected for generating the final answer. This mechanism allows the system's search quality to continuously optimize itself based on user feedback.

[0036] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0037] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for constructing an intelligent agent for information question answering in research teams based on Agentic workflow, characterized in that, include: Acquire multi-source heterogeneous data and construct multiple vertical domain-specific RAG knowledge bases based on the multi-source heterogeneous data; Construct a directed acyclic graph workflow with multiple functional nodes and associate it with the RAG knowledge bases specific to the multiple vertical domains; Based on the multi-node question-answering workflow, intent recognition is performed on the user's input query and historical dialogue context. The query is then input into a lightweight text classification model to determine the research team and research direction. Based on a dynamic weight allocation mechanism, the routing decision result of the target-specific RAG knowledge base corresponding to the query is obtained. Based on the routing decision results, multi-strategy RAG retrieval and multi-model collaborative generation are performed on the query to obtain the final natural language answer; Collect metadata and behavioral characteristics of user interactions, and quantify multi-dimensional indicators such as copy and use of answers, affirmative or negative actions, session depth and page dwell time to obtain user satisfaction evaluation results that reflect the current interaction quality. Based on the user satisfaction evaluation results, the user queries, enhanced prompt information, retrieved knowledge fragments, natural language answers, user satisfaction and explicit feedback are recorded in a structured manner and stored in a relational database to form a dialogue log dataset. Based on the dialogue log dataset and the corresponding user satisfaction evaluation results, dynamic confidence parameters are maintained for each dedicated RAG knowledge base. The confidence of each dedicated RAG knowledge base is updated according to the preset confidence update rules, so as to be used for the retrieval and sorting of dedicated RAG knowledge bases in different vertical fields during subsequent question answering.

2. The method for constructing an intelligent agent for information question answering in a research team based on Agentic workflow as described in claim 1, characterized in that, The multi-source heterogeneous data includes: Academic papers, experimental reports, technical manuals, project documents, and algorithm code.

3. The method for constructing an intelligent agent for information question answering in a research team based on Agentic workflow as described in claim 1, characterized in that, Acquire multi-source heterogeneous data, and construct multiple vertical domain-specific RAG knowledge bases based on the multi-source heterogeneous data, including: The multi-source heterogeneous data is preprocessed to obtain preprocessed multi-source heterogeneous data; Text extraction and segmentation are performed on the preprocessed multi-source heterogeneous data to uniformly convert data from different sources and in different formats into several text fragments; Each of the text segments is vectorized and encoded to obtain a corresponding high-dimensional vector representation, and the high-dimensional vector and the corresponding associated original text segment are stored in a vector database; Dimensionality reduction and clustering are performed on the high-dimensional vectors stored in the vector database. Based on the clustering results, multiple semantically similar vector clusters are determined, and representative classification labels are extracted for each vector cluster. Based on each vector cluster and its corresponding classification label, the associated text fragments are organized and categorized to construct vertical domain-specific RAG knowledge bases for different research teams and research topics.

4. The method for constructing an intelligent agent for information question answering in a research team based on Agentic workflow as described in claim 1, characterized in that, The plurality of functional nodes include: Intent recognition, prompt word engineering, knowledge retrieval, response generation, and feedback recording.

5. The method for constructing an intelligent agent for information question answering in a research team based on Agentic workflow according to claim 1, characterized in that, The lightweight text classification model is the ALBERT lightweight text classifier.

6. The method for constructing an intelligent agent for information question answering in a research team based on Agentic workflow as described in claim 1, characterized in that, The process, based on the multi-node question-answering workflow, involves intent recognition of the user's input query and historical dialogue context. The query is then input into a lightweight text classification model to determine the research team and research direction. Finally, based on a dynamic weight allocation mechanism, a routing decision result for the target-specific RAG knowledge base corresponding to the query is obtained, including: Obtain the query currently entered by the user and combine it with the user's historical dialogue context in this session to form the input text of the intent to be identified; The input text is fed into a lightweight text classification model to classify and predict the research team and research direction to obtain candidate research team and research direction labels. Based on the candidate research teams and research direction tags, a number of candidate vertical domain-specific RAG knowledge base sets are determined. Based on the preset dynamic weight allocation mechanism, the similarity information between the current input text and the representative text vectors in each candidate vertical domain exclusive RAG knowledge base, as well as the historical interaction records or knowledge base confidence parameters, are comprehensively considered to calculate the weight score of each candidate vertical domain exclusive RAG knowledge base. The candidate vertical domain-specific RAG knowledge bases are sorted according to the weight scores, and one or more vertical domain-specific RAG knowledge bases whose weight scores meet the preset conditions are selected to obtain the routing decision result of the target specific RAG knowledge base corresponding to the query.

7. The method for constructing an intelligent agent for information question answering in a research team based on Agentic workflow according to claim 1, characterized in that, The formula for calculating the user satisfaction evaluation result is as follows: ； in, These are the scores for the first indicator, the second indicator, and the third indicator, respectively. These are the weights of the first indicator, the second indicator, and the third indicator, respectively.

8. The method for constructing an intelligent agent for information question answering in a research team based on Agentic workflow according to claim 1, characterized in that, The expression corresponding to the preset confidence update rule is: ； in, and The knowledge base confidence scores at time points t and t+1 are respectively. As the attenuation factor, It is a learning rate.

Images