Project Aether

Anesthesia protocol intelligence — ingest, structure, search, and harmonize clinical guidelines with a local medical AI stack. Built for the PowerSync hackathon 2026.

---

What it does

Project Aether ingests anesthesia protocols from any format (scanned PDFs, images, DOCX, plain text), structures them with a local medical LLM (MedGemma), cross-references them against PubMed, and makes them instantly searchable via pgvector semantic search.

Feature	How
Document ingestion	Docling Serve (primary) → DeepSeek VL2 → pdf-parse / mammoth / Tesseract fallback chain
Medical structuring	MedGemma via Ollama — extracts title, category, drugs, steps, indications, contraindications
Confidence scoring	Auto-rates each guideline 0–1; flags low-quality protocols for peer review
Vector indexing	Full-section chunking → OpenAI embeddings → pgvector HNSW (LEANN in-memory layer for hot queries)
PubMed verification	Cross-references citations against NCBI E-utilities via Supabase Edge Function
Crowd-sourced tricks	Stack Overflow-style upvote system for anesthesia tips with auto-badges
Research gaps	Surfaces tips used at 5+ hospitals with zero published studies
Knowledge base health	Real-time ECG dashboard — confidence + flagged count drives heart rate and waveform

---

Quick start

Prerequisites

• Node.js 20+
• Ollama with MedGemma: ollama pull medgemma
• A Supabase project (free tier works)
• OpenAI API key (for embeddings)

1. Clone & install

git clone https://github.com/HarryBMa/easyRAG.git
cd easyRAG
npm install

2. Start Ollama

ollama serve
# in another terminal:
ollama pull medgemma

3. Configure environment

cp .env.example .env

Minimum required:

DATABASE_URL=postgresql://postgres:[password]@db.[ref].supabase.co:5432/postgres
OPENAI_API_KEY=sk-...
OLLAMA_MODEL=medgemma
SUPABASE_URL=https://[ref].supabase.co
SUPABASE_ANON_KEY=...

4. Run

npm run dev
# → http://localhost:3000

---

Architecture

Browser (TanStack Start + React 19)
  └── REST API
        ├── /api/upload ──── OCR → MedGemma → Postgres + pgvector chunks
        ├── /api/guidelines ─ CRUD + trending / newest sort
        ├── /api/query ────── semantic search (LEANN → pgvector fallback)
        ├── /api/tricks ───── crowd tips + voting
        ├── /api/vote ──────── upvote / downvote
        ├── /api/trends ────── knowledge base stats + research gaps
        └── /api/citation-graph ─ Semantic Scholar paper relationships

Upload pipeline:
  File → extractText() ──────── Docling Serve / DeepSeek VL2 / Tesseract
       → splitIntoSections() ── markdown headers or fixed 3 000-char chunks
       → structureGuideline() ─ MedGemma (full section, num_ctx 4096)
       → evaluateGuideline() ── confidence scoring + flag detection
       → INSERT guidelines ──── Postgres
       → embedBatch() ─────────  OpenAI text-embedding-3-small (per chunk)
       → upsertChunks() ──────── protocol_chunks (pgvector HNSW)

Lib:
  lib/ocr.ts        Docling Serve → DeepSeek VL2 → pdf-parse / Tesseract
  lib/medllm.ts     Ollama → MedGemma structuring + categorization
  lib/scoring.ts    Confidence scoring + flag detection + badge system
  lib/turso.ts      Postgres client + schema migrations (guidelines, tricks, sources, votes)
  lib/milvus.ts     pgvector store + LEANN in-memory HNSW acceleration
  lib/leann.ts      hnswlib-node HNSW index (falls back to pgvector if unavailable)
  lib/embed.ts      OpenAI embeddings with chunk splitting (350-word windows, 50-word overlap)
  lib/literature.ts Multi-database literature search (PubMed, Semantic Scholar, OpenAlex)

---

Tech stack

Layer	Technology
Frontend	TanStack Start · React 19 · Tailwind CSS v4
State / fetching	TanStack Query
Database	Supabase Postgres
Vector store	pgvector (HNSW cosine) + hnswlib-node in-memory layer
OCR	Docling Serve · DeepSeek VL2 (Ollama) · Tesseract.js 5
Medical LLM	MedGemma via Ollama (local, no cloud required)
Embeddings	OpenAI text-embedding-3-small
Literature	PubMed NCBI · Semantic Scholar · OpenAlex

---

Environment variables

# Required
DATABASE_URL=postgresql://...
OPENAI_API_KEY=sk-...
SUPABASE_URL=https://[ref].supabase.co
SUPABASE_ANON_KEY=eyJ...

# Ollama (defaults shown)
OLLAMA_HOST=http://localhost:11434
OLLAMA_MODEL=medgemma

# Optional — Docling Serve for best PDF/DOCX extraction
DOCLING_URL=http://localhost:5001

# Optional — embedding dimension (match your model)
EMBEDDING_MODEL=text-embedding-3-small
EMBEDDING_DIM=1536

Swapping models

• Medical LLM: set OLLAMA_MODEL — any Ollama model works (llama3.2, mistral, phi4)
• OCR: Docling Serve is the primary path; DeepSeek VL2 and Tesseract are automatic fallbacks
• Embeddings: set EMBEDDING_MODEL — any OpenAI-compatible embedding endpoint