PubMed RAG — Gut-Brain Axis Assistant

A lightweight, all-local pipeline that scrapes PubMed, fine-tunes a compact LLM, and layers Retrieval-Augmented Generation (RAG) to answer microbiome ↔ brain questions.

---

1 · Data Acquisition & Curation

Decision point	Selected choice & rationale
Access method	NCBI E-utilities (ESearch → ESummary/Efetch) — official API, structured, ToS-compliant.
Search syntax	(microbiome OR "gut microbiota") AND ("gut-brain axis" OR "brain-gut axis") — high recall, good precision.
Date / citation filters	Keep every paper; then require ≥ 5 citations if > 3 years old.
Batching & rate-limit	retmax=10000; set ENTREZ_API_KEY + email for 10 req/s (fallback 0.35 s sleep).
Fields stored	PMID, Title, Abstract, Year, Journal, MeSH, CitationCount.
Storage format	Raw XML cache → cleaned JSONL in data/.

---

2 · Local Pre-Processing

• Strip HTML, normalise Unicode, merge line breaks, lower-case (except gene symbols).
• Idempotent scripts skip already-cached PMIDs.
• Dependencies frozen in requirements.txt.

---

3 · Fine-Tuning on a MacBook Pro (M3)

Aspect	Locked-in choice
Base model	phi-2 (1.3 B) — small, high quality, quantisable.
Inference backend	mlc-llm — Metal-accelerated on Apple Silicon.
Tuning strategy	QLoRA (4-bit) adapters with HF peft; training done once on a free Colab GPU, adapters (< 100 MB) loaded locally.
Training data	Auto-generated instruction pairs from cleaned abstracts.
Evaluation	Perplexity + spot-check QA versus vanilla phi-2.

---

4 · Retrieval-Augmented Generation (RAG)

Component	Selected design
Embeddings	sentence-transformers/all-MiniLM-L6-v2 — 80 MB, CPU-friendly.
Vector store	FAISS (flat L2) — ~1 GB for 10 k abstracts.
Retriever	Top-k = 5 similarity search; abstracts passed into prompt with numeric tags.
Prompt template	You are a biomedical assistant. **Context:** {docs} **Question:** {query} **Answer:**
Grounding	Model footnotes sources [1]-[k]; UI shows abstracts on hover.
Frontend	Gradio single-file app; runs locally, deployable to HF Spaces.

---

5 · Demo & Deployment

• Packaging: Dockerfile (CPU-only) + Makefile; Homebrew path for macOS users.
• Docs: this README + architecture PNG + week-by-week plan.
• Future: swap in Mistral-7B-Instruct (GGUF) when more GPU available; add PubMedQA benchmark.

---

Architecture (high-level)