2 ‐ Methodology overview
VerifAI is a Retrieval-Augmented Generation-based system designed to perform referenced QA in the biomedical domain. It consists of three main components.
- The Information Retrieval component, based on hybrid semantic and lexical search, retrieves relevant documents and provides a context for the generative LLM.
- Generative component, based on large language model, either self-hosted, or called using available API (e.g. OpenAI).
- Verification system, that uses encoder transformer model to cross-check generated answer with sources based on which answer was given and verify and flag any potential hallucinations
Schematically system can be explained using the following image:
The primary component of this toolbox is the information retrieval engine based on indexed documents (Verifai BioMed has indexed abstract dataset from PubMed database). The question-answering system utilizes either a standalone fine-tuned LLM, or LLM API, such as OpenAI to generate answers using retrieved documents in context (depending on the configuration settings). A fact-checking or verification engine examines the generated answer within the toolbox, identifying any potential hallucinations in the system.