Research Paper AI Assistant
A production-grade RAG platform for indexing, searching, and querying scientific papers with citation-accurate answers.
Problem
Researchers wade through long PDFs to find relevant findings, and generic LLM chat tools hallucinate citations or lose page-level context.
Solution
Built an async ingestion pipeline (Celery) with layout-aware PDF parsing (PyMuPDF), recursive semantic chunking, and hybrid retrieval combining Qdrant dense vectors with BM25 sparse search, fused via Reciprocal Rank Fusion. A Jina cross-encoder reranks results before Gemini 2.5 Flash synthesizes answers with inline citations tied to exact page coordinates, surfaced through a React dashboard.
Technology
- Python
- FastAPI
- Celery
- Qdrant
- BM25
- Jina Reranker
- Gemini 2.5 Flash
- React
- Docker
Result
Delivers citation-accurate answers grounded in the source PDFs instead of generic LLM recall, with a Docker Compose quick start for one-command local deployment.