Context-Guided Answer Verification for LLM Information Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large language models (LLMs) like GPT and PaLM2 can generate improper answers due to 'hallucination', leading users to believe in untrue responses.
Innovation Solution
An information retrieval system that includes a question receiving unit, context information retrieving unit, first and second prompt generating units, and an answer verifying unit to generate and verify answers using a large language model, ensuring confidence levels are met before outputting responses.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a large language model generates answers directly from input questions, then the response speed is fast and the system is simple, but the accuracy deteriorates due to hallucination and improper answers
Solution Approach 1:
The system segments the answer generation process into multiple distinct units: a question receiving unit, a context information retrieving unit, a first prompt generating unit, an answer acquiring unit, a second prompt generating unit, and an answer verifying unit. Each unit performs a specific function, transforming a single complex generation task into a sequence of manageable steps that collectively improve reliability while maintaining reasonable system complexity.
Solution Approach 2:
The system introduces context information from a database as an intermediary between the input question and the LLM. This context information acts as a mediator that guides the LLM to generate more accurate answers by providing relevant background knowledge, thereby reducing hallucination without requiring fundamental changes to the LLM itself.
Solution Approach 3:
The system implements a feedback mechanism where the generated answer is fed back into the system through a second prompt that includes the context information and the first answer. The LLM processes this feedback to verify the answer's accuracy and provide confidence levels, creating a self-correcting loop that improves reliability.
2Reliability
If context information retrieval and answer verification are added to the system, then the answer reliability is improved, but the processing time increases
Solution Approach 1:
The system performs preliminary action by retrieving context information from the database before the LLM generates the answer. This pre-retrieval of relevant context ensures that when the LLM generates the answer, it has access to accurate reference information, reducing the need for multiple verification iterations and thereby minimizing the time penalty associated with added reliability measures.
3Measurement precision
If multiple prompts and verification steps are implemented, then the confidence level of answers is improved, but the computational resources consumed increase
Solution Approach 1:
The system applies partial action by implementing verification only when necessary - the second prompt with confidence measurement is generated based on the first answer, allowing the system to achieve sufficient confidence levels without always performing full verification. This selective approach provides adequate confidence measurement while avoiding excessive computational energy consumption for every single query.
Data Source
AI summary
A context information retrieving unit retrieves a document database with a characteristic vector of a question and thereby acquires as context information a text that a similarity level between the characteristic vector and a characteristic vector of the text satisfies a predetermined condition. A first prompt generating unit generates a first prompt that includes the question and the context information. An answer acquiring unit acquires as a first answer an answer corresponding to the first prompt using a large language model. A second prompt generating unit generates a second prompt that includes the context information and the first answer and causes to provide a confidence of the first answer. An answer verifying unit acquires as a second answer an answer corresponding to the second prompt using a large language model, and determines the confidence with the second answer.