pubmed-local-rag

Query PubMed abstracts using a fully local LLM (Ollama) and a RAG pipeline — no API keys required.

Requirements

• Ollama installed and running
• llama3.2 model pulled (ollama pull llama3.2)

Setup

pip install -r requirements.txt

Usage

1. Build vector DB (run once per topic)

python build_db.py

Prompts for:

• Search term (default: single cell RNA-seq cancer)
• Number of papers to fetch (default: 10, max: 10000)

DB is saved to ./chroma_db/<search_term>/ — each search term gets its own folder.

2. Ask questions

python ask.py

Prompts for the DB path to load, then enters an interactive Q&A loop. Type q to quit.

3. Manage databases

python manage_db.py

Lists all stored databases with size and date. Supports deleting individual or all databases.

Example Output

$ python ask.py
Enter DB path to use (default: ./chroma_db): ./chroma_db/single_cell_RNA_seq_cancer
Loading DB from: ./chroma_db/single_cell_RNA_seq_cancer

Enter your question (quit: q): What are the key findings about tumor microenvironment?

Answer: Several studies identified distinct immune cell populations within the tumor
microenvironment using scRNA-seq, including exhausted T cells, tumor-associated
macrophages (TAMs), and regulatory T cells. These findings suggest that the
composition of the TME is strongly associated with patient prognosis and response
to immunotherapy.
[14.2s]

Enter your question (quit: q): q

Benchmark

Run benchmark.py to compare embedding models and LLMs across a fixed question set. Results are saved to benchmark_results.csv.

ollama pull mistral  # download additional model (~4GB)
python benchmark.py

Models tested:

Embedding Model	LLM	Avg Response Time
all-MiniLM-L6-v2	llama3.2	~14s
all-MiniLM-L6-v2	mistral	~18s
all-mpnet-base-v2	llama3.2	~16s
all-mpnet-base-v2	mistral	~21s

Project Structure

├── build_db.py       # Fetch PubMed abstracts and build vector DB
├── ask.py            # Interactive Q&A against stored DB
├── manage_db.py      # List and delete stored databases
├── benchmark.py      # Compare embedding models and LLMs
├── requirements.txt
└── README.md

Stack

• PubMed API (Biopython)
• LangChain LCEL
• ChromaDB
• HuggingFace Embeddings (all-MiniLM-L6-v2, all-mpnet-base-v2)
• Ollama (llama3.2, mistral)