app.yaml

# Databricks App spec for the Unity Gateway Voice Studio. One container runs the
# stdlib web/token tier (app/web_server.py) and the LiveKit agent worker
# (app/agent.py) side by side, launched by start_app.py (pure Python; the Apps
# image has no Node). The web tier binds the port immediately; the worker boots
# behind it in an isolated /tmp/agent-venv.
#
# Credentials come from app SECRET resources backed by the `ug-voice-studio`
# secret scope (no secret values live in this file). Everything else is
# non-sensitive config. The app authenticates to Databricks with the injected
# PAT (DATABRICKS_TOKEN); start_app.py drops the runtime's OAuth SP creds so the
# SDK has one unambiguous auth method (matches local behavior).
command: ["python", "start_app.py"]

env:
  # --- runtime / agent identity (non-secret) ---
  - name: PORT
    value: "8000"
  - name: AGENT_NAME
    value: "ug-agent"

  # --- secrets (valueFrom -> app secret resource -> ug-voice-studio scope) ---
  - name: LIVEKIT_URL
    valueFrom: livekit-url
  - name: LIVEKIT_API_KEY
    valueFrom: livekit-api-key
  - name: LIVEKIT_API_SECRET
    valueFrom: livekit-api-secret
  - name: DEEPGRAM_API_KEY
    valueFrom: deepgram-api-key
  - name: DATABRICKS_TOKEN
    valueFrom: database-token
  # Plan 5 Halloween voice (ElevenLabs). `elevenlabs-api-key` is an app secret
  # resource bound to a key in the ug-voice-studio scope (suggested key name:
  # `elevenlabs_api_key`). The secret VALUE and that resource binding are created
  # OUT-OF-BAND by the owner (see skills/building-voice-agents-on-databricks/
  # references/deployment.md) -- no key value ever lives in this repo. Attach the
  # resource BEFORE deploying this file: a `valueFrom` only resolves once the app
  # resource exists. If the key is missing or unusable at runtime the Halloween
  # voice drops to the Deepgram fallback voice and the call carries on.
  - name: ELEVEN_API_KEY
    valueFrom: elevenlabs-api-key

  # --- Databricks workspace (non-secret) ---
  - name: DATABRICKS_HOST
    value: "https://your-workspace.cloud.databricks.com"

  # --- Lakebase (non-secret config; token minted at runtime via the SDK) ---
  - name: LAKEBASE_ENDPOINT
    value: "projects/your-project/branches/production/endpoints/primary"
  - name: LAKEBASE_DATABASE
    value: "databricks_postgres"
  - name: UG_SCHEMA
    value: "ug"

  # --- Unity Gateway tier->model routing map (non-secret) ---
  - name: UG_MODEL_STANDARD
    value: "system.ai.gpt-5-nano"
  - name: UG_MODEL_PREMIUM
    value: "system.ai.gpt-5-5"
  - name: UG_MODEL_VIP
    value: "system.ai.gpt-6-sol"
  - name: UG_MODEL_FALLBACK
    value: "system.ai.gpt-5-5"
  - name: UG_EMBED_MODEL
    value: "system.ai.gte-large-en"
  - name: UG_GEN_MODEL
    value: "system.ai.gpt-5-4"

  # --- Plan 5: AI Decide, the per-turn voice-mode decision (non-secret) ---
  # AI Decide decides the voice mode each turn by calling the Databricks
  # `ai_decide` function. It needs `ai_decide` enabled on the workspace (admin ->
  # Previews); without it, explicit requests still work through the rule-based
  # check. UG_AI_DECIDE=0 is the kill switch.
  - name: UG_AI_DECIDE
    value: "1"                        # kill switch: "0" = never classify, stay in the standard voice
  - name: UG_DECIDE_TIMEOUT_S
    value: "3.0"                      # limit on the background decision call
  - name: UG_DECIDE_CUE_WAIT_S
    value: "0.8"                      # longest a turn waits for the decision before replying in the current voice

  # --- Plan 5: Halloween voice (non-secret; the key is ELEVEN_API_KEY above) ---
  - name: UG_HALLOWEEN_TTS
    value: "elevenlabs"               # elevenlabs | openai | deepgram
  - name: UG_HALLOWEEN_TTS_MODEL
    value: "eleven_v3_conversational" # only eleven_v3* models perform inline [tags]; any other model gets none
  - name: UG_HALLOWEEN_STABILITY
    value: "0.5"                      # ElevenLabs voice setting, 0 to 1
  - name: UG_HALLOWEEN_FALLBACK_VOICE
    value: "aura-2-zeus-en"           # Deepgram voice used when ElevenLabs is unavailable
  # UG_HALLOWEEN_VOICE_ID is REQUIRED for the ElevenLabs voice and is the owner's
  # pick (chosen by ear). It is deliberately left unset here: with no voice id the
  # Halloween profile is the Deepgram fallback voice above, never ElevenLabs' stock
  # voice. To enable the ElevenLabs voice, add this entry:
  #   - name: UG_HALLOWEEN_VOICE_ID
  #     value: "<elevenlabs-voice-id>"

  # --- optional tracing (fail-soft if unavailable) ---
  - name: OTEL_SERVICE_NAME
    value: "ug-voice-agent"
  - name: DATABRICKS_TRACE_CATALOG
    value: "your_catalog"
  - name: DATABRICKS_TRACE_SCHEMA
    value: "your_schema"
  - name: DATABRICKS_TRACE_TABLE_PREFIX
    value: "ugvoice"
  - name: MLFLOW_EXPERIMENT_NAME
    value: "/Users/you@example.com/ugvoice"
  - name: MLFLOW_TRACING_SQL_WAREHOUSE_ID
    value: "your-sql-warehouse-id"