How allu555/medical-chatbot works
Sorin-AI is a medical conversational assistant whose primary workflow takes a patient query through emergency triage, then—if routine—retrieves medical references and generates a cited response in a Gradio interface. Its supporting branches build the retrieval index, train/load a local model, and evaluate response quality and safety. The README documents the overall pipeline; source excerpts and the SOURCE INDEX substantiate only some implementation calls, so other documented relationships are cited to the README or omitted.
Main entry points: app.py Gradio consultation UI · generate.py Clinical response generation · retriever.py Medical retrieval · safety.py Emergency pattern gate
Components 18
Patient conversation
- Gradio consultation UI
app.py - Emergency pattern gate
src/inference/safety.py - Semantic triage
src/inference/gemini_client.py - Emergency guidance
app.py - Clinical response generation
src/inference/generate.py
Medical knowledge RAG
- Curated medical documents
knowledge_base/documents - Document loading and chunking
src/rag/knowledge_base.py - Embedding index builder
src/rag/build_index.py - FAISS index and metadata
artifacts/faiss - Medical retrieval
src/rag/retriever.py
Model inference and training
- Local model loader
src/inference/model_loader.py - Dataset preparation
src/data/prepare_dataset.py - QLoRA training
src/training/train_qlora.py
Safety and evaluation
- Model response evaluation
src/evaluation/evaluate_model.py - Retrieval evaluation
src/evaluation/evaluate_rag.py - Held-out triage validation
evaluate_heldout.py
Other components
- Patient
- Google Gemini API
Connections 20
- Patient → Gradio consultation UI: submits query
README.md - Gradio consultation UI → Emergency pattern gate: checks symptoms
README.md - Gradio consultation UI → Semantic triage: classifies urgency
app.py - Semantic triage → Google Gemini API: requests triage
- Gradio consultation UI → Emergency guidance: shows escalation
app.py - Gradio consultation UI → Medical retrieval: retrieves context
README.md - Gradio consultation UI → Clinical response generation: requests response
README.md - Clinical response generation → Local model loader: loads model
src/inference/generate.py - Clinical response generation → Emergency pattern gate: validates output
src/inference/generate.py - Clinical response generation → Medical retrieval: gets references
- Clinical response generation → Gradio consultation UI: returns answer
app.py - Curated medical documents → Document loading and chunking: loads documents
- Embedding index builder → Document loading and chunking: loads and chunks
src/rag/build_index.py - Embedding index builder → FAISS index and metadata: writes index
- Medical retrieval → FAISS index and metadata: searches index
- Dataset preparation → QLoRA training: prepares training data
- Model response evaluation → Clinical response generation: evaluates responses
src/evaluation/evaluate_model.py - Model response evaluation → Medical retrieval: uses retrieval
src/evaluation/evaluate_model.py - Retrieval evaluation → Medical retrieval: measures retrieval
src/evaluation/evaluate_rag.py - Held-out triage validation → Emergency pattern gate: tests detection
evaluate_heldout.py