Research Paper RAG
rag
llm
fastapi
python
A hybrid retrieval pipeline for question answering over research papers, with citations and an automated evaluation suite.
Overview
A question-answering service over research papers. Answers come with citations back to the source papers, and the system declines to answer when retrieval doesn’t find enough relevant material.
Retrieval
Retrieval Recall@5 improved from 72% to 91% across 200 benchmark queries by combining four techniques:
- Dense vector search (Qdrant)
- BM25 keyword retrieval
- Metadata filtering
- Cross-encoder reranking
Grounding
Hallucinated responses fell by 45%, and grounded-answer accuracy reached 88%, after adding:
- Source-aware prompting
- Relevance thresholds that stop the model answering from weak context
- Citations back to the source papers
- An automated RAG evaluation suite to measure each change
Latency
Median end-to-end query latency fell by 38% after three changes: caching paper embeddings and retrieval results, batching embedding requests, and deploying the FastAPI service with PostgreSQL and Qdrant in Docker. Ingestion jobs are orchestrated with Inngest.
Stack
Python · FastAPI · Inngest · PostgreSQL · Qdrant · Docker