GeoNjunge/MedicalRAG

Data extraction and summarization in Low-Resource Medical NLP Pipelines # MedicalRAG A medical document analysis system that turns PDFs into structured clinical data diseases, lab results, ICD-10 codes, and a short summary with a focus on **reliability and reduced token consumption**. Built for low-resource environments: CPU-only hardware, small local models, and cloud deployment without loading gigabytes of NLP weights onto the server. --- ## What it does Upload a medical PDF and the system returns: - **Diseases** with confidence scores and ICD-10 codes - **Lab results** with values, units, and normal/abnormal flags - **A plain-text summary** written from the extracted data (not from a raw document dump) The key idea: extract facts first with specialized tools, then let the LLM summarize only what was already found. This cuts down on made-up details. --- ## Two ways to run it MedicalRAG supports two modes, controlled by `APP_ENV`: | Mode | Best for | LLM | NLP stack | |------|----------|-----|-----------| | **Local (dev)** | Research, offline use, tuning | Qwen 2.5 via llama.cpp on your machine | Full pipeline: PubMedBERT, medSpaCy, Docling | | **Production** | Hosted deployment (Render, Vercel, etc.) | Groq cloud API | Lightweight: PyMuPDF + Groq prompts | Read the architecture guides for details: - **Production architecture** cloud pipeline, Groq, deployment - **Local LLM architecture** offline models, Redis worker, llama.cpp --- ## Quick start ### Production (minimal setup) 1. Clone the repo and install production dependencies: ```bash pip install -r requirements-prod.txt ``` 2. Copy `.env.example` to `apps/api/.env` and set `APP_ENV=production` plus your `GROQ_API_KEY`. 3. Start the API from `apps/api`: ```bash uvicorn app.main:app --reload ``` 4. For the frontend: ```bash cd apps/web/frontend npm install npm start ``` ### Local development (full pipeline) You need Redis, Ollama (for model files), and the dev dependencies. See the full walkthrough in **docs/SETUP.md**. Short version three terminals: ```b …