PubMed RAG Researcher is a Streamlit app for rapid literature review. It searches PubMed abstracts, summarizes findings with Gemini, and supports follow-up Q&A grounded in retrieved papers.

• PubMed search with configurable date range and target paper count.
• Iterative PMID batching to keep searching until enough valid papers are collected.
• Optional free-access filtering behavior through the sidebar toggle.
• Real-time retrieval progress UI (progress bar + per-paper status text).
• Gemini-generated structured summary with:
  • findings synthesis,
  • most relevant paper recommendation,
  • IEEE-style references.
• Vector retrieval for chat using ChromaDB + Google Generative AI embeddings.
• Automatic fallback to stored abstracts if vector retrieval is unavailable.
• Retrieval-mode transparency in UI (shows whether answer used vector search or fallback).
• Session-state chat history and conversation continuity.
• PDF export options:
  • summary only,
  • summary with full follow-up chat log.
• Graceful degradation when ChromaDB is unavailable.

1. Search PubMed with your topic and date range.
2. Build a local abstract corpus for the current run.
3. Store documents in ChromaDB (when available).
4. For follow-up questions, retrieve top-k semantically similar chunks.
5. Fall back to full stored abstracts if DB retrieval is unavailable.

• Python
• Streamlit
• Google Gemini API (google-generativeai)
• Metapub (PubMed access)
• ChromaDB (vector store)
• FPDF (PDF generation)
• python-dotenv (env configuration)

1. Clone this repository.
2. Install dependencies:

pip install -r requirements.txt

3. Create a .env file in the project root:

GEMINI_API_KEY=your_key_here
NCBI_API_KEY=your_ncbi_key_here

NCBI_API_KEY is optional but recommended for higher PubMed request throughput.

4. Run the app:

streamlit run app.py

This repo currently includes tests for:

• pdf_generator.py (smoke + regression coverage)
• database_manager.py (ResearchDB behavior and fallbacks)

Recommended (runs both unittest-style and pytest-style tests):

pytest -v tests

If pytest is not installed:

pip install pytest

• Summaries and Q&A are bounded by abstract availability/quality from PubMed records.
• If embeddings or vector DB initialization fails, chat falls back to raw abstract context.
• Export currently targets PDF only.

• Add citation-level source snippets in chat responses.
• Improve ranking with metadata-aware re-ranking (year, study type, relevance score).
• Add trend visualizations (publication count over time by topic).
• Add optional full-text ingestion for open-access papers.