MedDiscover

MedDiscover is an AI-powered tool designed to assist biomedical researchers using RAG-LLM models fine-tuned on PubMed literature.

CLI evaluation (headless)

Install the package (or use it in editable mode), set your OPENAI_API_KEY, and run the built-in evaluator:

pip install .
export OPENAI_API_KEY=...
# optional: ALLOW_MEDCPT_CPU=1 to force MedCPT on CPU

meddiscover-eval \
  --pdfs med_discover_ai/eval_samples/sample_pdfs/fmed-11-1345659.pdf med_discover_ai/eval_samples/sample_pdfs/s10549-023-07033-8.pdf \
  --qa_csv med_discover_ai/eval_samples/sample_qa.csv \
  --embedding_model "MedCPT (GPU Recommended)" \
  --llm_models gpt-4.1-mini \
  --k 3 \
  --max_tokens 64 \
  --out_dir ./eval_outputs_demo

# Evaluate both decoders in one run (example)
meddiscover-eval \
  --pdfs med_discover_ai/eval_samples/sample_pdfs/fmed-11-1345659.pdf med_discover_ai/eval_samples/sample_pdfs/s10549-023-07033-8.pdf \
  --qa_csv med_discover_ai/eval_samples/sample_qa.csv \
  --embedding_model "MedCPT (GPU Recommended)" \
  --llm_models gpt-4.1-mini,gpt-4.1-nano \
  --k 3 --max_tokens 64 --out_dir ./eval_outputs_all

• For Ada-based retrieval, switch --embedding_model to OpenAI Ada-002 (CPU/Cloud).
• RAGAS metrics are optional; if dependencies are missing or the QA CSV lacks a reference column, they fall back to None.
• Re-ranking stays disabled on CPU; enable --rerank only when a GPU and cross-encoder are available.
• Ollama models are available in both the UI and CLI (prefix with ollama:), e.g. --llm_models ollama:gemma3:4b.