AdaptiveRAG-Project

AdaptiveRAG is a Retrieval-Augmented Generation (RAG) application designed to answer questions from company policy documents. Unlike a basic RAG system, AdaptiveRAG evaluates the quality of the initial retrieval before generating an answer. If relevant information is not found, the system automatically reformulates the user's query and performs a second retrieval attempt. If the required information is still unavailable, the system provides a safe fallback response instead of allowing the LLM to generate unsupported information. The project demonstrates how retrieval evaluation and adaptive query processing can improve the reliability of RAG applications.

The system uses TF-IDF based document retrieval and cosine similarity scoring to identify relevant document chunks. It then performs keyword-based relevance checking and retrieval quality evaluation using heuristic similarity thresholds. If the first retrieval does not produce relevant results, the system reformulates the query and performs a second retrieval. The retrieved context is passed to a Groq-hosted LLM through LangChain for final answer generation, with the LLM instructed to answer only from the retrieved context. The application includes an interactive Streamlit chat interface and displays supporting documents and similarity scores.