Classified Document Vector Retrieval for More Accurate LLM Answers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Generative AI using large language models often outputs inaccurate answers to questions, necessitating a method to improve the selection of information added to prompts for more accurate responses.
Innovation Solution
An information processing system that includes a vector database and a large language model, which classifies content and users, converts content into document vectors, and generates prompts based on user and content classifications to enhance the relevance of answers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a versatile large language model is used to answer questions, then the model can handle various types of queries, but the accuracy of the answers decreases
Solution Approach 1:
The system segments the vast document collection into smaller, classified portions using a classification tree structure. Documents are organized into categories and subcategories based on their content, allowing the system to retrieve only relevant segments rather than searching the entire corpus. This segmentation enables the versatile LLM to focus on specific, relevant information, thereby improving answer accuracy.
Solution Approach 2:
The system applies local quality by assigning different classification characteristics to different portions of the document collection. Each document or section is tagged with specific classification items corresponding to its content type, user target, or other attributes. This allows the retrieval process to select documents with the appropriate local qualities (classifications) that match the question's requirements, enhancing the LLM's ability to provide accurate answers.
2Reliability
If more information is added to prompts to improve answer accuracy, then the relevance of answers improves, but the complexity of information selection and processing increases
Solution Approach 1:
The system performs preliminary classification and organization of documents before the question-answering process. Documents are pre-tagged with classification items and structured in a classification tree, so that when a question is received, the system can quickly navigate to relevant portions without complex real-time analysis. This preliminary action simplifies the retrieval process while maintaining high relevance.
Solution Approach 2:
The classification tree structure acts as an intermediary between the question and the document collection. Instead of directly searching through all documents, the system uses the classification tree as a mediator to guide the retrieval process. The tree structure translates user questions into appropriate classification paths, simplifying the information selection process while ensuring relevant documents are retrieved.
3Reliability
If classification-based filtering of document vectors is implemented, then answer accuracy improves, but the processing time and computational resources increase
Solution Approach 1:
The system segments the document collection into classified portions using a tree structure, allowing the retrieval process to focus on specific branches rather than scanning the entire corpus. This segmentation significantly reduces the number of documents that need to be processed and compared, thereby reducing processing time while maintaining accuracy through classification-based filtering.
Solution Approach 2:
Documents are pre-classified and organized into the classification tree structure before retrieval operations. This preliminary organization allows the system to quickly navigate to relevant document portions using the tree structure, avoiding the need to process and analyze all documents during the retrieval phase. The pre-computed classification information speeds up the filtering process while ensuring accurate document selection.
Data Source
AI summary
A method for processing information is executed by one or more computers. The method includes obtaining a user classification item related to a question, the user classification item being any of user classification items based on user classification, the user classification corresponding to content classification; converting the question into a question vector; searching document vectors assigned with a content classification item that corresponds to the user classification item related to the question vector for a specified number of the document vectors having a high degree of relevance to the question vector; generating a prompt for input to a large language model, the prompt including content of the specified number of the document vectors prior to conversion, serving as context, and the question; inputting the prompt to the large language model; and outputting the answer generated based on text output by the large language model.