# ---------------------------------------------------------------------------
# Base de données externe (Supabase)
# Project Settings → Database → Connection string
# ---------------------------------------------------------------------------
DB_HOST=aws-0-eu-west-1.pooler.supabase.com
DB_PORT=5432
DB_NAME=postgres
DB_USER=postgres.ogjzkmbvorbcdbfltiib
DB_PASSWORD=o2FeWHSz04kgVRL4
DB_SSL_MODE=require

# ---------------------------------------------------------------------------
# Secrets (générer chacun avec : openssl rand -hex 32)
# ---------------------------------------------------------------------------

# Token pour les endpoints /admin/* de admin-api (header X-Admin-API-Key)
ADMIN_API_TOKEN=3209a9a3af263757e2803625ff7c5931c7f2d61c37aaf9629cd888fd172b553d68
ADMIN_TOKEN=5622b7e0de1a38e613ba8442231120fb3b94e49ec427abaacf629ad8ac440a81
# Secret partagé entre admin-api et api-gateway pour /internal/validate
# DOIT être identique dans les deux services
INTERNAL_API_SECRET=a499796b63273356f599a5d592a1c4d0af8009ecd1f6d1729235331197639ffa

# Token partagé entre meeting-api et runtime-api
RUNTIME_API_TOKEN=408fdf4e7e80f015f6ceb2da182a70d9439e1504c613dd9044ec8eba2cfd4bc2

# Token read-only pour les endpoints analytics (optionnel)
# ANALYTICS_API_TOKEN=

# ---------------------------------------------------------------------------
# Bot
# ---------------------------------------------------------------------------
# En prod (image publique) :
BOT_IMAGE_NAME=vexaai/vexa-bot:latest

# URL callback baked dans la config des bots spawned
# En local (réseau interne Docker) :
MEETING_API_URL=http://meeting-api:8080
# En prod :
# MEETING_API_URL=https://cr-meet.ashia.sn

# ---------------------------------------------------------------------------
# Services optionnels
# La gateway démarre avec ces URLs fictives si les services ne sont pas déployés.
# Les endpoints /transcripts et /mcp retourneront des erreurs à l'usage.
# ---------------------------------------------------------------------------
 TRANSCRIPTION_COLLECTOR_URL=http://meeting-api:8080
 TRANSCRIPTION_SERVICE_URL=http://transcription:8083/v1/audio/transcriptions

# ---------------------------------------------------------------------------
# Divers
# ---------------------------------------------------------------------------
LOG_LEVEL=INFO
CORS_ORIGINS=*
BOT_STOP_DELAY_SECONDS=90
VEXA_ENV=development   # désactive /docs et /openapi.json sur admin-api

# ---------------------------------------------------------------------------
# Divers
# ---------------------------------------------------------------------------
# Model configuration
MODEL_SIZE=medium

# Available models (all multilingual):
# - tiny, base, small, medium, large-v2, large-v3, large-v3-turbo
#
# Recommended: large-v3-turbo + INT8
# - GPU VRAM: ~2.1 GB (validated)
# - Quality: Excellent (95-98% accuracy)
# - Speed: Very fast (>10x real-time)
# - Multilingual: 99+ languages

# Device configuration
# DEVICE=cuda  # For GPU (default)
 DEVICE=cpu   # For CPU-only

# Compute type (optimization)
 COMPUTE_TYPE=int8     # Default: 50-60% VRAM reduction, 2-4x CPU speedup, minimal accuracy loss
# COMPUTE_TYPE=float16  # GPU only: Maximum speed, higher VRAM usage (~6-8 GB)

# CPU optimization (only used when DEVICE=cpu)
# CPU_THREADS=4  # Set to number of physical CPU cores (0 = auto-detect)

# Load management / backpressure
# These control how the service behaves under load.
# Recommended for WhisperLive streaming: FAIL_FAST_WHEN_BUSY=true
MAX_CONCURRENT_TRANSCRIPTIONS=2   # Max concurrent model calls per worker
MAX_QUEUE_SIZE=10                # Max requests waiting (ignored when FAIL_FAST_WHEN_BUSY=true)
FAIL_FAST_WHEN_BUSY=true         # Return 503 immediately when busy (lets WhisperLive keep buffering/coalescing)
BUSY_RETRY_AFTER_S=1             # Retry-After header value (seconds) for busy/overload responses

# Quality parameters (derived from WhisperLive best practices)
# These parameters improve transcription quality and accuracy
BEAM_SIZE=5                    # Beam size: 1 = greedy (fast), 5 = beam search (better quality, slower)
BEST_OF=5                      # Number of candidates when sampling with non-zero temperature
COMPRESSION_RATIO_THRESHOLD=1.8  # If gzip compression ratio > this, treat as failed (hallucination detection) - lowered to catch repetitions
LOG_PROB_THRESHOLD=-1.0        # If avg log probability < this, treat as failed
NO_SPEECH_THRESHOLD=0.6         # If no_speech_prob > this AND log_prob < threshold, consider silent
CONDITION_ON_PREVIOUS_TEXT=false # Use previous output as prompt for next window - DISABLED to prevent repetition loops
PROMPT_RESET_ON_TEMPERATURE=0.3 # Reset prompt if temperature > this (prevents stuck loops) - lowered for more aggressive reset
REPETITION_PENALTY=1.1          # Penalize repeated tokens (>1.0 = penalize). Prevents "they are saying they are saying..."
NO_REPEAT_NGRAM_SIZE=3          # Hard-block any 3-word phrase from repeating

# VAD (Voice Activity Detection) parameters
VAD_FILTER=true                # Enable VAD to filter out non-speech audio
VAD_FILTER_THRESHOLD=0.5       # VAD onset threshold (0-1): higher = less sensitive to noise
VAD_MIN_SILENCE_DURATION_MS=160  # Minimum silence duration in milliseconds

# Temperature fallback chain (for quality improvement)
# Uses multiple temperatures and falls back if compression_ratio or log_prob thresholds are exceeded
USE_TEMPERATURE_FALLBACK=true  # Enable temperature fallback chain [0.0, 0.2, 0.4, 0.6, 0.8, 1.0]

# API Token for securing the service
# This token must match TRANSCRIPTION_SERVICE_API_TOKEN in the gateway
# If not set, service will accept all requests (not recommended for production)

API_TOKEN=271f1e185abcebdf7f610b95d388d7ce3fe7a3a7c28c565cc88ad51a165198c6