RAG Question Answering with Diverse Retrieval and Answer Ranking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing question answering systems using Large Language Models (LLMs) often generate biased text responses towards retrieved information, particularly when utilizing external data sources.
Innovation Solution
Implement a Retrieval Augmented Generation (RAG) model that retrieves information from diverse data sources, including unstructured and structured data, using multiple retrievers and a cross-encoder model to rank and consolidate data, reducing reliance on LLM reasoning and optimizing computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single retriever is used to fetch information from external data sources, then the system structure remains simple, but the response accuracy and diversity of information are limited
Solution Approach 1:
The patent divides the retrieval system into multiple specialized retrievers (e.g., dense retriever, sparse retriever, hybrid retriever) that each handle different types of queries and data sources. This segmentation allows each retriever to be optimized for specific tasks, improving overall response accuracy while maintaining manageable complexity through modular architecture
Solution Approach 2:
The patent implements a universal retrieval system where multiple retrievers work together to handle diverse query types and data sources. The system can adaptively select and combine different retrievers based on the specific query requirements, providing multi-functional capability that improves response accuracy across various scenarios
2Measurement precision
If multiple diverse data sources are retrieved to improve answer accuracy, then the information quality improves, but the computational resources and memory requirements increase
Solution Approach 1:
The patent retrieves only the necessary portion of information from diverse data sources by using query understanding to identify relevant sources and limiting the amount of data fetched. This partial action approach maintains high answer accuracy by focusing on essential information while reducing unnecessary computational overhead and memory usage
Solution Approach 2:
The patent dynamically adjusts retrieval parameters such as the number of results to fetch, confidence thresholds, and data source selection based on query characteristics. This allows the system to optimize the balance between information quality and computational resource consumption for different types of questions
3Device complexity
If the LLM relies heavily on its own reasoning capabilities to answer questions, then the system requires fewer external components, but the computational complexity and time consumption increase
Solution Approach 1:
The patent performs preliminary information retrieval and processing before the LLM generates its response. By pre-fetching and preprocessing relevant information from external sources, the system reduces the reasoning burden on the LLM and enables faster response generation without sacrificing answer quality
Data Source
AI summary
Embodiments described herein provide systems and methods for retrieval augmented generation. A neural network based language model may be provided a question as a user input. Based on the user input, semantically diverse queries may be generated for retrieval from diverse data sources. For example, a structured data source (e.g., database or knowledge base) and unstructured data (e.g., text articles) may be used to retrieve information relevant to the user input. The retrieve information may be ranked so that the most relevant information is used by the language model in generating an answer to the question in the user input. A non-retrieval based answer generated by the language model may be utilized in some embodiments in generating the final answer.