pediatric-ai-scribe-v3/.env.example
Daniel 93426ac089
Some checks failed
Forgejo Docker Build / Root app tests (push) Failing after 58s
Forgejo Docker Build / Build Docker image (push) Has been skipped
Forgejo Docker Build / End-to-end (browser) (push) Has been skipped
feat: the library index reconciles when the admin asks — status and Run indexing now in the Clinical Assistant card
The clinical-assist indexer no longer polls Nextcloud every five minutes; it
scans once at start and then on POST /api/v1/vector-sync/scan with a bearer
token. The admin panel shows what it reports and carries the button; the
address and token are settings (or the environment). The button stays usable
under lockdown, the fields do not.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
2026-09-13 16:02:35 +02:00

303 lines
15 KiB
Text

# ============================================================
# OPENBAO (optional — recommended for production)
# ============================================================
# When these three are set, the container fetches everything else below
# from OpenBao at kv/ped-ai/prod and ignores the equivalent .env values.
# Leave them unset (or blank) to fall back to .env-only (local dev, e2e).
#
# OPENBAO_ADDR=https://app.danvics.com
# OPENBAO_ROLE_ID=<from: bao read auth/approle/role/ped-ai/role-id>
# OPENBAO_SECRET_ID=<from: bao write -f auth/approle/role/ped-ai/secret-id>
# OPENBAO_KV_PATH=kv/ped-ai/prod # override path if needed
# ============================================================
# Everything below is sourced from OpenBao when OPENBAO_ADDR is set.
# Only fill these in for local dev / e2e / when running without vault.
# ============================================================
# ============================================================
# AI PROVIDER (choose one)
# ============================================================
# Option 1: OpenRouter (default, cheapest, NOT HIPAA)
AI_PROVIDER=openrouter
OPENROUTER_API_KEY=sk-or-v1-your-key
# Option 2: AWS Bedrock (HIPAA compliant with BAA)
# AI_PROVIDER=bedrock
# AWS_BEDROCK_REGION=us-east-1
# AWS_ACCESS_KEY_ID=AKIA...
# AWS_SECRET_ACCESS_KEY=...
# (Or use IAM role — no keys needed if running on EC2/ECS)
# Option 3: Azure OpenAI (HIPAA compliant with BAA)
# AI_PROVIDER=azure
# AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com
# AZURE_OPENAI_API_KEY=your-key
# AZURE_DEPLOYMENT_NAME=gpt-4o-mini
# AZURE_OPENAI_API_VERSION=2024-02-01
# Option 4: Google Vertex AI (HIPAA compliant with BAA)
# AI_PROVIDER=vertex
# Google models (Gemini and friends) are reached through LiteLLM, not the
# Google SDK — that dependency was removed. Configure them on the gateway
# and select them by their LiteLLM model id.
# GOOGLE_APPLICATION_CREDENTIALS=/path/to/service-account.json
# (Or use default credentials if running on GCE/GKE/Cloud Run)
#
# Google STT — Gemini inline audio (auto-detected when GOOGLE_VERTEX_PROJECT set)
# TRANSCRIBE_PROVIDER=google
# GOOGLE_STT_MODEL=gemini-2.0-flash # or gemini-2.5-flash for better accuracy
#
# Google TTS — Google Cloud Text-to-Speech (auto-detected when GOOGLE_VERTEX_PROJECT set)
# TTS_PROVIDER=google
# GOOGLE_TTS_VOICE=en-US-Journey-F # female | en-US-Journey-D = male
# Other options: en-US-Studio-O, en-US-Neural2-C, en-US-Neural2-J
# Option 5: LiteLLM Proxy (self-hosted, routes to any provider)
# AI_PROVIDER=litellm
# LITELLM_API_BASE=http://localhost:4000
# LITELLM_API_KEY=sk-litellm-your-key
# Admin can discover available models via the admin panel
#
# LiteLLM Speech-to-Text
# TRANSCRIBE_PROVIDER=litellm
# LITELLM_STT_MODEL=whisper-1 # Use the model name from your LiteLLM model_list
# If your LiteLLM config uses full paths as model names, use the full path:
# LITELLM_STT_MODEL=openai/whisper-1
# NOTE: vertex_ai/chirp does NOT work via LiteLLM audio proxy.
# For Vertex AI speech, use TRANSCRIBE_PROVIDER=google (Gemini inline audio).
#
# LiteLLM TTS
# TTS_PROVIDER=litellm (auto-detected when LITELLM_API_BASE set)
# LITELLM_TTS_MODEL=tts-1 # Use model name from your LiteLLM model_list
# If your config uses full paths: LITELLM_TTS_MODEL=vertex_ai/google-tts
# LITELLM_TTS_VOICE=en-US-Journey-F # Google Cloud voice name (or alloy/nova for OpenAI)
# ============================================================
# TRANSCRIPTION (speech-to-text)
# ============================================================
# Option A: OpenAI Whisper (default if no AWS configured)
OPENAI_API_KEY=sk-your-openai-key
# Option B: Amazon Transcribe (HIPAA eligible, no S3 needed)
# Uses same AWS credentials as Bedrock above.
# Set TRANSCRIBE_PROVIDER=aws to force AWS even if OPENAI_API_KEY is set.
# Leave unset to auto-detect (uses AWS when AWS_BEDROCK_REGION is configured).
# TRANSCRIBE_PROVIDER=aws
# Option C: Local Whisper (privacy-first, no cloud API needed)
# Requires whisper.cpp or faster-whisper installed on the server.
# TRANSCRIBE_PROVIDER=local
# WHISPER_MODEL_SIZE=small # tiny, base, small, medium, large
# WHISPER_BINARY=whisper-cpp # or: whisper, faster-whisper
# WHISPER_MODEL_PATH= # custom path to .bin model file
# WHISPER_LANGUAGE=en
# WHISPER_THREADS=4 # defaults to CPU count - 1
# Amazon Transcribe Medical — better accuracy for clinical dictation
# Knows drug names, diagnoses, procedures, SOAP terminology
# HIPAA eligible (ensure your AWS account has a BAA)
# AWS_TRANSCRIBE_MEDICAL=true
# AWS_TRANSCRIBE_SPECIALTY=PRIMARYCARE
# Other options: CARDIOLOGY, NEUROLOGY, ONCOLOGY, RADIOLOGY, UROLOGY
# Optional
ELEVENLABS_API_KEY=
# Push Notifications (ntfy — self-hosted, optional)
# NTFY_URL=https://ntfy.yourdomain.com
# NTFY_TOKEN=tk_your_token_here
# App
PORT=3000
APP_URL=https://your-domain.com
# Cloudflare Turnstile (anti-bot on registration, optional)
# TURNSTILE_SITE_KEY=your-site-key
# TURNSTILE_SECRET_KEY=your-secret-key
JWT_SECRET=generate-a-random-64-char-string-here
# Application-layer encryption key for PHI at rest (Nextcloud tokens, audio backups)
# Generate with: openssl rand -hex 32
# REQUIRED in production. Rotating invalidates existing encrypted data.
DATA_ENCRYPTION_KEY=generate-with-openssl-rand-hex-32
# Email (for verification & password reset)
SMTP_HOST=smtp.gmail.com
SMTP_PORT=587
SMTP_USER=your-email@gmail.com
SMTP_PASS=your-app-password
SMTP_FROM=noreply@yourdomain.com
# Nextcloud (optional)
NEXTCLOUD_URL=https://cloud.yourdomain.com
# S3 Document Storage (optional — works with AWS S3, Backblaze B2, MinIO)
# S3_BUCKET=your-bucket-name
# S3_REGION=us-east-1
# S3_PREFIX=documents/
#
# For AWS S3: uses same AWS credentials as Bedrock above, or set S3-specific keys:
# S3_ACCESS_KEY_ID=...
# S3_SECRET_ACCESS_KEY=...
#
# For Backblaze B2:
# S3_ENDPOINT=https://s3.us-west-004.backblazeb2.com
# S3_REGION=us-west-004
# S3_ACCESS_KEY_ID=your-b2-application-key-id
# S3_SECRET_ACCESS_KEY=your-b2-application-key
#
# For MinIO (self-hosted):
# S3_ENDPOINT=http://minio:9000
# S3_REGION=us-east-1
# S3_ACCESS_KEY_ID=minio-access-key
# S3_SECRET_ACCESS_KEY=minio-secret-key
# S3_FORCE_PATH_STYLE=true
# ============================================================
# EMBEDDINGS (for Learning Hub semantic search)
# ============================================================
# Enables vector-based semantic search in Learning Hub
# Requires pgvector extension: apt-get install postgresql-16-pgvector
# Default model (Vertex AI text-embedding-005, 768 dims, English + code optimized)
EMBEDDING_MODEL=vertex_ai/text-embedding-005
EMBEDDING_DIMENSIONS=768
# Other Vertex AI embedding models:
# - vertex_ai/text-embedding-005 → 768 dims, English + code (recommended)
# - vertex_ai/gemini-embedding-001 → up to 3072 dims, multilingual + code
# - vertex_ai/text-multilingual-embedding-002 → 768 dims, multilingual focus
#
# LiteLLM usage (if using LiteLLM proxy):
# EMBEDDING_MODEL=text-embedding-005 # LiteLLM will route to configured provider
#
# OpenAI fallback (NOT HIPAA-eligible):
# Uses text-embedding-3-small if OPENAI_API_KEY is set and no Vertex/LiteLLM configured
# ============================================================
# DATABASE
# ============================================================
DATABASE_URL=postgresql://pedscribe:<password>@postgres:5432/pedscribe
DB_PASSWORD=pedscribe_secret_change_me
# ── Audio backups (optional) ────────────────────────────────────────────────
# Every recording is kept for 24 hours, whether its transcription succeeded or
# not. Unset, the audio lives in the audio_backups table; set, it goes to a
# bucket instead and only metadata stays in Postgres. The generated-images key
# is scoped to that bucket and cannot be reused here — create a bucket and a
# user of its own, and give the bucket a 24h expiry rule as a backstop.
# Audio is gzipped and AES-256-GCM encrypted before it is stored, either way.
# AUDIO_BACKUPS_S3_ENDPOINT=http://assets:9000
# AUDIO_BACKUPS_S3_BUCKET=audio-backups
# AUDIO_BACKUPS_S3_REGION=us-east-1
# Prefer the _FILE variants: credentials then stay out of the process environment.
# AUDIO_BACKUPS_S3_ACCESS_KEY_FILE=/run/secrets/audio-backups-access-key
# AUDIO_BACKUPS_S3_SECRET_KEY_FILE=/run/secrets/audio-backups-secret-key
# AUDIO_BACKUPS_S3_ACCESS_KEY=
# AUDIO_BACKUPS_S3_SECRET_KEY=
# ============================================================
# EVERY REMAINING VARIABLE THE APP READS
# Completed 2026-09-11: this file listed 18 of the 67 variables src/ actually
# reads. Each entry below shows the default the code falls back to, so an
# unset line means "this default", not "broken".
# ============================================================
# ── Mail ────────────────────────────────────────────────────────────────────
# SMTP_SECURE=false # true for implicit TLS (port 465)
# ── Admin lockdown ──────────────────────────────────────────────────────────
# With several admins, everything in the admin panel is editable by all of
# them. Set this and configuration becomes read-only in the panel and refused
# by the server: prompts, model policy, retrieval limits and budgets, TTS/STT,
# embeddings, SMTP and email templates. Day-to-day operation stays available —
# announcements, registration and invitations, feature flags, site details.
# A setting added later is locked until it is deliberately added to the
# editable list, rather than defaulting to open.
#
# It is deliberately an environment variable: a setting could be switched off
# by the very admin it restrains, so lifting it needs host access and a
# restart. Read access is unaffected — locked settings stay visible.
# ADMIN_LOCKDOWN=false
# ── Identity and limits ─────────────────────────────────────────────────────
# SITE_NAME=Pediatric AI Scribe
# API_RATE_LIMIT_MAX=200 # requests per window across /api
# LOGIN_RATE_LIMIT_MAX=10 # login attempts per 15 minutes
# Codes emailed for sign-in, per IP per hour. Lower than the login limit
# because each request sends mail to somebody else's address (default 5).
#LOGIN_CODE_RATE_LIMIT_MAX=5
# NODE_ENV=production # with APP_URL, puts the app in production mode:
# refuses to start without JWT_SECRET or a CORS origin
# CORS_ORIGINS= # extra allowed origins, comma-separated, beyond APP_URL
# ── Services ────────────────────────────────────────────────────────────────
# REDIS_URL=redis://ped-ai-redis:6379
# LOKI_URL=http://monitoring-loki:3100 # log shipping; unset disables it
# MCP_SERVER_URL= # legacy single MCP URL
# ── LiteLLM gateway ─────────────────────────────────────────────────────────
# LITELLM_MASTER_KEY= # falls back to OPENAI_API_KEY, then 'sk-litellm'
# LITELLM_DEFAULT_MODEL= # falls back to OPENAI_MODEL
# LITELLM_FALLBACK_MODEL= # tried when the default model fails
# OPENAI_MODEL= # legacy name for the default chat model
# LITELLM_TTS_VOICES= # comma-separated voices offered in the pickers
# ── Clinical Assistant: retrieval ───────────────────────────────────────────
# CLINICAL_ASSISTANT_MCP_URL=http://mcp:8000/mcp
# The indexer (a separate container) reconciles the library only when the
# admin panel asks. Address and token can also be set in the admin panel.
# CLINICAL_ASSISTANT_INDEXER_URL=http://mcp-indexer:8001
# CLINICAL_ASSISTANT_INDEXER_TOKEN=
# CLINICAL_ASSISTANT_MCP_URLS= # comma-separated, tried in order
# CLINICAL_ASSISTANT_SEARCH_TOOL=clinical_semantic_search # the only accepted value
# CLINICAL_ASSISTANT_MCP_INITIALIZE_TIMEOUT_MS=30000
# CLINICAL_ASSISTANT_MCP_REQUEST_TIMEOUT_MS=90000
# CLINICAL_ASSISTANT_MCP_SESSION_TTL_MS=600000
# CLINICAL_ASSISTANT_MCP_CONCURRENCY=3 # library searches in flight at once; they used to run one at a time
# CLINICAL_ASSISTANT_MCP_WARMUP= # open a session at boot
# CLINICAL_ASSISTANT_MCP_WARMUP_DELAY_MS=
# ── Clinical Assistant: models and prompts ──────────────────────────────────
# CLINICAL_ASSISTANT_IMAGE_MODEL=openai-gpt-image-1
# CLINICAL_ASSISTANT_PROMPT_MODEL= # model that writes the starter prompts
# CLINICAL_ASSISTANT_PROMPT_POOL_TARGET=1000
# CLINICAL_ASSISTANT_PROMPT_POOL_KEY=clinical-assistant:prompt-pool:v2
# CLINICAL_ASSISTANT_PROMPT_POOL_REFRESH_MS=
# CLINICAL_ASSISTANT_PROMPT_POOL_WARMUP_DELAY_MS=
# CLINICAL_ASSISTANT_EXAMPLE_CACHE_MS=600000
# PATIENT_TAKEHOME_BEHAVIOR= # overrides the admin-editable take-home guidance
# ============================================================
# OPENBAO (optional)
# The entrypoint fetches secrets from OpenBao only when OPENBAO_ADDR is set;
# unset, it uses this file and nothing else. A value set here or in the compose
# environment always wins over the OpenBao copy, so an override needs no
# OpenBao change.
#
# To use OpenBao: set OPENBAO_ADDR, OPENBAO_ROLE_ID, OPENBAO_SECRET_ID
# (and OPENBAO_KV_PATH if not kv/ped-ai/prod)
# To stop using it: unset OPENBAO_ADDR, and put the keys it held in this
# file. List them without printing values:
# bao kv get -format=json kv/ped-ai/prod | jq 'keys'
# ============================================================
# OPENBAO_ADDR=
# OPENBAO_ROLE_ID=
# OPENBAO_SECRET_ID=
# OPENBAO_KV_PATH=kv/ped-ai/prod
# Optional. Lets something outside this host scrape /metrics with
# `Authorization: Bearer <token>`. Unset by default, and an unset token can
# never match — Prometheus scrapes pediatric-ai-scribe:3000 directly over the
# Docker network, which needs no token. Requests arriving through the reverse
# proxy (they carry X-Forwarded-For) get a 404 either way.
METRICS_TOKEN=
# Gotenberg (LibreOffice behind an HTTP API), used to turn a generated deck or
# document into PDF. Defaults to http://gotenberg:3000 on the danvics_convert
# network. PDF is the one export allowed to fail: if this is unreachable the
# PowerPoint and Word downloads still work.
GOTENBERG_URL=