RAG Scientific Paper QA
A retrieval-augmented assistant for grounded question answering over scientific papers.
Overview
This project explores whether retrieval-augmented generation can make answers about scientific papers easier to verify and less dependent on a language model’s general knowledge. It uses a small, reproducible sample from the QASPER dataset and compares direct generation with retrieval-grounded generation.
Technical approach
- Converts papers into overlapping text chunks and embeds them with
BAAI/bge-m3. - Builds a FAISS index and retrieves supporting passages with cosine similarity.
- Generates answers with Qwen2.5-7B-Instruct or Mistral-7B-Instruct-v0.3.
- Compares four settings: direct and RAG generation with each model.
- Supports quantitative evaluation, manual scoring, and evidence-hit analysis.
- Provides a Streamlit interface that displays answers, retrieved evidence, and similarity scores.
- Includes a mock mode for lightweight testing without downloading large models.
Technologies: Python, Hugging Face Transformers, FAISS, Streamlit, QASPER, Qwen, and Mistral.