.env.example

# MedArmor Demo – all environment variables (single source of truth)
# Copy to .env and set values. Used by: docker-compose, backend (from demo/backend), frontend build.
# No values should be hardcoded in docker-compose or in app config – everything comes from this file.

# ---------- Database (MedArmor PostgreSQL) ----------
POSTGRES_USER=postgres
POSTGRES_PASSWORD=postgres
POSTGRES_DB=medaudit
DATABASE_URL=postgresql://postgres:postgres@medaudit_db:5432/medaudit

# ---------- Database (EHR PostgreSQL) ----------
EHR_POSTGRES_USER=ehr
EHR_POSTGRES_PASSWORD=ehr
EHR_POSTGRES_DB=ehr
EHR_DATABASE_URL=postgresql://ehr:ehr@medaudit_db:5432/ehr

# ---------- Redis / Celery ----------
REDIS_URL=redis://redis:6379/0
CELERY_BROKER_URL=redis://redis:6379/0
CELERY_RESULT_BACKEND=redis://redis:6379/0

# ---------- Backend server (REQUIRED) ----------
HOST=0.0.0.0
PORT=8000

# ---------- Backend (API, worker, beat) ----------
EHR_SERVICE_URL=http://ehr:8000
EHR_DATA_ROOT=
CHROMA_PERSIST_DIR=/app/chroma_data
MEDICAL_KB_PATH=/app/docs/Medical_KB

ENVIRONMENT=development
DEBUG=true

JWT_SECRET_KEY=change_me_jwt_secret_min_32_chars
JWT_ALGORITHM=HS256
JWT_EXPIRE_MINUTES=60

ALLOWED_ORIGINS=http://localhost:3000,http://localhost:5173,http://localhost,http://127.0.0.1

# ---------- RAG embeddings (REQUIRED) ----------
# Options: fastembed (default, no API key) | openai | medsiglip
# medsiglip: uses google/medsiglip-448 text encoder — 64-token limit, needs torch + HF_TOKEN.
#   Better suited for image+text retrieval. For text-only RAG, fastembed or openai are stronger.
EMBEDDING_PROVIDER=fastembed
OPENAI_API_KEY=
OPENAI_EMBEDDING_MODEL=text-embedding-3-small
# MedSigLIP (when EMBEDDING_PROVIDER=medsiglip)
MEDSIGLIP_MODEL=google/medsiglip-448
MEDSIGLIP_DEVICE=cpu

# ---------- Audit AI: one of medgemma | gemini | openai ----------
AUDIT_AI_PROVIDER=medgemma
# Extraction mode: one_pass | gemini_extract | medgemma_extract
# gemini_extract requires GOOGLE_API_KEY; medgemma_extract requires MedGemma (HF or Vertex)
AUDIT_EXTRACTION_MODE=gemini_extract
# Sensitivity: 0.0 (only critical) to 1.0 (surface everything). Default 0.5.
AUDIT_SENSITIVITY=0.5
# Guidelines mode: rag (vector retrieval via ChromaDB, requires embeddings) | full (read entire KB into prompt, no embeddings)
# Use full when you have a single guidelines file — avoids FastEmbed/ONNX dependency entirely.
AUDIT_GUIDELINES_MODE=full
# MedGemma backend: hf (Hugging Face) | vertex (Vertex AI). Default hf.
MEDGEMMA_BACKEND=hf
# Hugging Face (when MEDGEMMA_BACKEND=hf)
HF_TOKEN=
HF_MEDGEMMA_ENDPOINT=
# Vertex AI (when MEDGEMMA_BACKEND=vertex) – MedGemma hosted on Vertex AI
# Requires GOOGLE_APPLICATION_CREDENTIALS or gcloud auth application-default login
VERTEX_AI_PROJECT=
VERTEX_AI_LOCATION=
VERTEX_AI_MEDGEMMA_ENDPOINT_ID=
GOOGLE_API_KEY=
GEMINI_MODEL=gemini-2.5-flash
OPENAI_AUDIT_MODEL=gpt-4o-mini

# ---------- Voice Dictation / ASR ----------
# Provider: medasr | medasr_local | deepgram
ASR_PROVIDER=medasr_local
# medasr  – remote HuggingFace Inference API; reuses HF_TOKEN from MedGemma config
# e.g. https://api-inference.huggingface.co/models/google/medasr
MEDASR_HF_ENDPOINT=
# medasr_local – run google/medasr in-process via transformers>=5.0.0
# Model card: https://huggingface.co/google/medasr (must accept HF terms)
MEDASR_LOCAL_MODEL=google/medasr
# Device for local inference: cpu | cuda:0 | mps
MEDASR_LOCAL_DEVICE=cpu
# deepgram – Deepgram Nova-3-Medical; get a key at https://console.deepgram.com
DEEPGRAM_API_KEY=

# ---------- Optional (legacy / overrides) ----------
QDRANT_URL=http://localhost:6333
QDRANT_API_KEY=

# ---------- Frontend (Vite build; only VITE_* are exposed to client) ----------
VITE_API_URL=/api

# ---------- EHR service (ehr container) ----------
EHR_DATA_PATH=/data
HF_TOKEN=