config.small-team.example.yaml

Rama main · réplica de lectura del repositorio original.

# DeltaLLM small-team configuration
#
# One server (Docker Compose "single" profile) or a one-replica Helm release
# serving a team. Compared with config.example.yaml it adds:
# - durable spend and audit writes (outbox), which the Model Selector and
#   Realtime require
# - the Batch API with local artifact storage
#
# Every setting left out uses its built-in default. Which features to turn on,
# what each needs, and the other profiles: docs/deployment/feature-switches.md
#
# PostgreSQL connections: about 38 per DeltaLLM process (db_pool_size 20 +
# db_foreground_pool_size 8 + telemetry pools 5 + 5). Keep PostgreSQL
# max_connections above that for every process you run, plus one restart overlap.

model_list:
  - model_name: gpt-4o-mini
    deltallm_params:
      provider: openai
      model: openai/gpt-4o-mini
      api_key: os.environ/OPENAI_API_KEY
      api_base: https://api.openai.com/v1
      timeout: 60

router_settings:
  routing_strategy: simple-shuffle
  # One retry on the same deployment, then fail over to the next one.
  num_retries: 1
  retry_after: 1
  timeout: 600
  # A deployment rests for cooldown_time seconds after more than allowed_fails
  # rate-limit, server, or timeout errors within five minutes.
  cooldown_time: 60
  allowed_fails: 2
  response_validation:
    enabled: true
    on_exhausted: return_response
    json_healing: false

general_settings:
  instance_name: DeltaLLM

  master_key: os.environ/DELTALLM_MASTER_KEY
  salt_key: os.environ/DELTALLM_SALT_KEY
  database_url: os.environ/DATABASE_URL
  redis_url: os.environ/REDIS_URL

  platform_bootstrap_admin_email: os.environ/PLATFORM_BOOTSTRAP_ADMIN_EMAIL
  platform_bootstrap_admin_password: os.environ/PLATFORM_BOOTSTRAP_ADMIN_PASSWORD

  # Manage deployments in the Admin UI; the sample model seeds the first boot only.
  model_deployment_source: db_only
  model_deployment_bootstrap_from_config: true

  background_health_checks: true
  health_check_interval: 300

  # Durable spend and audit writes through a background worker in this process.
  # Restart after changing either mode.
  spend_ingestion_mode: outbox
  audit_ingestion_mode: outbox

  # Batch API (/v1/batches). local storage is safe with one server only; switch to
  # s3 (embeddings_batch_s3_bucket, region, prefix) before adding a second one.
  embeddings_batch_enabled: true
  embeddings_batch_storage_backend: local

  # Organization deletion requests. Startup fails while the deletion readiness check
  # finds missing migrations or unowned records, so enable only after it passes and
  # every API and worker replica reports the backend's current lifecycle
  # protocol version (v2 for this rollout).
  organization_deletion_requests_enabled: false

  # ---------------------------------------------------------------------------
  # Common next steps. Uncomment what you need.
  # ---------------------------------------------------------------------------
  #
  # Email for invitations, password reset, and notifications.
  # email_enabled: true
  # email_provider: smtp
  # email_from_address: no-reply@example.com
  # email_base_url: https://llm.example.com
  # smtp_host: smtp.example.com
  # smtp_port: 587
  # smtp_username: os.environ/SMTP_USERNAME
  # smtp_password: os.environ/SMTP_PASSWORD
  #
  # Single sign-on (needs Redis). Restrict sign-in to your company's domains.
  # enable_sso: true
  # sso_provider: oidc
  # sso_client_id: os.environ/SSO_CLIENT_ID
  # sso_client_secret: os.environ/SSO_CLIENT_SECRET
  # sso_authorize_url: https://idp.example.com/oauth2/authorize
  # sso_token_url: https://idp.example.com/oauth2/token
  # sso_userinfo_url: https://idp.example.com/oauth2/userinfo
  # sso_redirect_uri: https://llm.example.com/auth/callback
  # sso_sign_in:
  #   allowed_email_domains:
  #     - example.com
  #
  # Budget and key-expiry emails (need email).
  # governance_notifications_enabled: true
  # budget_notifications_enabled: true
  # key_lifecycle_notifications_enabled: true
  #
  # OpenAI Realtime sessions (need Redis and the spend outbox above). Restart.
  # realtime:
  #   enabled: true

deltallm_settings:
  success_callback: []
  failure_callback: []
  turn_off_message_logging: false