Samarth Chachra

Conversational RAG Research Assistant

Built an adaptive, citation-grounded Research Assistant that enables researchers to ask natural language questions across 500+ research papers while receiving verifiable answers with source citations. Instead of relying on dense retrieval alone, I designed a hybrid retrieval pipeline combining BGE vector embeddings stored in Qdrant with BM25 lexical search, followed by cross-encoder reranking to improve retrieval quality. I implemented multi-turn conversational memory with PostgreSQL state persistence and an LLM query rewriter that converts follow-up questions into standalone retrieval queries. To prevent incomplete answers, I added a knowledge-gap detector that asynchronously indexes unseen papers without blocking user queries, reducing retrieval failures by 35% and improving end-to-end latency by 40%.