Imported from Adriandecode/HIPAA-Compliant-AI-Healthcare-System (backend/AGENTS.md). Install upstream with npx skills add Adriandecode/HIPAA-Compliant-AI-Healthcare-System --skill backend. Copyright stays with the author.

AGENTS.md - Backend

Scoped rules for backend/. Read root AGENTS.md first.

Key Module Ownership

For a deep reference of all 50+ backend modules, check CONTEXT.md. Here is a high-level summary:

• AI Layer: core_ai.py (Central entry point for all LLM/Ollama inference), prompt_registry.py (Prompt templates), agent.py (Supervisor agent), chat_context.py (RAG builder).
• Data & DB: models.py (SQLAlchemy models), database.py (Session engine), schemas/ (Pydantic schemas), rag.py (Vector store).
• ML & Predictions: prediction.py (Initializes and runs disease prediction models diabetes/heart/liver/kidney/lungs), explainability.py (SHAP).
• Hospital Workflows: hospital_operations.py, monitoring.py, diagnostics.py, pharmacy.py, billing.py, discharge.py, nursing.py, care_events.py.
• Interoperability & Compliance: abdm.py (India ABDM API), fhir.py (FHIR R4 helpers), interoperability.py (Bundle exports), audit.py (PHI-safe auditing).

Rules

• AI Provider Abstraction: Never import google.generativeai or httpx for AI calls outside core_ai.py. Use core_ai.generate(), core_ai.chat(), core_ai.chat_stream(), core_ai.embed_text(), or the relevant core_ai provider helper.
• Prompt Management: Never inline system prompts in route handlers. Register them in prompt_registry.py and retrieve via get_prompt("name").
• Database Sessions: Route handlers must use backend.database.get_db, usually as Depends(database.get_db). Never create SessionLocal() manually in route handlers.
• Schema Changes: Update models.py, schemas.py, and main.py migration/startup logic together when changing persisted fields.
• Error Handling: Log errors with logger.error(), never expose stack traces to clients. Return structured {"detail": "..."} errors.
• HIPAA Awareness: Health data (predictions, records, chat logs) must be scoped to current_user.id. Never return another user's data.
• ML Models: Model files (*.pkl) currently live in backend/ and models/. Loading is centralized in prediction.initialize_models().

Recommended Tests

python -m pytest tests/ -v -k "test_chat or test_auth or test_prediction"