JanaAlzobidi/bioasq-rag

🧬 BioASQ RAG Demo

A biomedical question-answering system built on Retrieval-Augmented Generation (RAG):

• —Corpus: 40,221 biomedical passages from the BioASQ benchmark
• —Retriever: `pritamdeka/S-PubMedBert-MS-MARCO` sentence encoder + FAISS index
• —Generator: LLaMA-3.3-70B-Versatile (hosted on Groq)
• —UI: Gradio with collapsible retrieved-evidence panel

The system retrieves the top 5 most semantically similar biomedical passages for each query, then generates an answer that is strictly grounded in those passages.

How it works

question → S-PubMedBert encoder → query vector
                                       ↓
                              FAISS (40,221 vectors)
                                       ↓
                          top-5 relevant passages
                                       ↓
                       LLaMA-3.3-70B (Groq) + strict prompt
                                       ↓
                            grounded biomedical answer

Disclaimer

This system is provided for informational and research purposes only. It is not a substitute for professional medical advice, diagnosis, or treatment. Always consult a qualified healthcare professional for clinical questions.

Setup notes for forks

If you fork this Space, you'll need to:

1. 1.Provide your own GROQ_API_KEY under Settings → Repository secrets.
2. 2.Ensure the data/ folder contains bioasq_passages.csv and bioasq_faiss.index (use Git LFS for files larger than 10 MB).
3. 3.The first launch takes 1–2 minutes to download the embedding model (~440 MB) and build the runtime. Subsequent launches are faster.

Evaluation summary

See the accompanying project report for full evaluation details (RAGAS, LLM comparison, error analysis).