render.yaml

services:
  - type: web
    name: bimo-backend
    runtime: python
    plan: free
    region: oregon
    rootDir: backend
    buildCommand: pip install -r requirements.txt
    startCommand: gunicorn -w 2 -k gthread --threads 8 --timeout 600 --bind 0.0.0.0:$PORT app.main:app
    autoDeploy: true
    envVars:
      - key: PYTHON_VERSION
        value: "3.11.9"
      - key: SUPABASE_URL
        sync: false
      - key: SUPABASE_SERVICE_ROLE_KEY
        sync: false
      - key: SUPABASE_JWT_SECRET
        sync: false
      - key: SUPABASE_STORAGE_BUCKET
        value: bimo-attachments
      - key: NVIDIA_API_KEY
        sync: false
      - key: NVIDIA_BASE_URL
        value: https://integrate.api.nvidia.com/v1
      - key: NVIDIA_MODEL
        value: meta/llama-3.3-70b-instruct
      # Friendly-mode model IDs — change these on Render to swap Aeon / Stanza /
      # Nexos / vision / Iris without touching code. Redeploy after editing.
      - key: NVIDIA_AEON_MODEL
        value: mistralai/mistral-small-4-119b-2603
      - key: NVIDIA_STANZA_MODEL
        value: qwen/qwen3-next-80b-a3b-instruct
      - key: NVIDIA_NEXOS_MODEL
        sync: false
      - key: NVIDIA_VISION_MODEL
        value: google/diffusiongemma-26b-a4b-it
      # Per-reply output cap. Higher = longer answers / full code builds
      # without mid-stream truncation. At ~40 tok/s, 8192 ≈ 3.5 min.
      - key: NVIDIA_MAX_TOKENS
        value: "8192"
      # Client read timeout (seconds). Must comfortably exceed the time to
      # generate NVIDIA_MAX_TOKENS at the model's speed, and stay under the
      # gunicorn --timeout above so a slow build isn't cut as an "API error".
      - key: NVIDIA_TIMEOUT
        value: "540"
      # Retry budget for transient NVIDIA failures. The free tier caps each
      # model at 40 requests/minute and returns 429 + Retry-After when hit;
      # the SDK backs off and honours that header within this budget.
      - key: NVIDIA_MAX_RETRIES
        value: "3"
      # Image generation (Iris). Served via NVIDIA's GenAI endpoint.
      # Klein 4B: distilled model, ~4 steps, fast text-to-image.
      - key: NVIDIA_IMAGE_BASE_URL
        value: https://ai.api.nvidia.com/v1/genai
      - key: NVIDIA_IMAGE_MODEL
        value: black-forest-labs/flux.2-klein-4b
      - key: CORS_ORIGINS
        value: https://bimo.qzz.io,http://localhost:5500,http://127.0.0.1:5500
      # Rate limiting is ON for the live deployment (per-user keying with IP
      # fallback). Point RATELIMIT_STORAGE_URI at a Redis URL so limits are
      # accurate across workers (in-memory only limits per worker).
      - key: RATELIMIT_ENABLED
        value: "1"
      - key: RATELIMIT_STORAGE_URI
        value: memory://
      # WhatsApp Cloud API Integration
      - key: WHATSAPP_PHONE_ID
        sync: false
      - key: WHATSAPP_BUSINESS_ACCOUNT_ID
        sync: false
      - key: WHATSAPP_TOKEN
        sync: false
      - key: WHATSAPP_VERIFY_TOKEN
        value: bimo_whatsapp_verify_token_2026
      - key: WHATSAPP_APP_SECRET
        sync: false
      # Groq AI for WhatsApp Aeon model
      - key: GROQ_API_KEY
        sync: false
      - key: GROQ_AEON_MODEL
        value: qwen/qwen3.8-27b
      # Web Search & Scraping (TinyFish)
      - key: TINYFISH_API_KEY
        sync: false
      # Autonomous search routing (Groq first, Mistral fallback, rules if neither)
      - key: SEARCH_CLASSIFIER_MODEL
        value: qwen/qwen3.8-27b
      - key: SEARCH_CLASSIFIER_TIMEOUT_S
        value: "1.5"
      # Mistral AI (Stanza 2.5)
      - key: MISTRAL_API_KEY
        sync: false
      - key: MISTRAL_MODEL
        value: ministral-8b-2512
      # Deepgram Flux Streaming TTS
      - key: DEEPGRAM_API_KEY
        sync: false
      - key: DEEPGRAM_TTS_MODEL
        value: flux-hannah-en