Configuration

server:
  port: "8080"
  read_timeout: 35m       # Long for reasoning models (o3, GPT-5.2)
  write_timeout: 35m
  idle_timeout: 120s
  coalescence_window: 50ms  # Batch streaming tokens (0 = disabled)

redis:
  address: "localhost:6379"
  username: ""
  password: ""
  database: 0
  pool_size: 10
  max_retries: 3
  dial_timeout: 5s

database:
  path: "./data/ensemble.db"
  max_open_conns: 25
  max_idle_conns: 5

cache:
  enable_session_affinity: true
  max_cache_entries: 100000
  crc_algorithm: crc32
  cache_wait_threshold: 0.25   # >$0.25 estimated value: strong affinity
  load_balance_threshold: 0.05 # <$0.05: prefer least-utilized endpoint
  ttls:
    anthropic: 8m
    openai: 24h
    gemini: 6m

rate_limit:
  window_size: 1m
  ttl_seconds: 65
  sync_interval: 1s        # Background Redis sync interval
  default_rpm: 1000
  default_tpm: 1000000
  redis_eval_timeout: 50ms
  redis_rollback_timeout: 25ms

# Provider configs are in separate files under config/providers/
# anthropic.yaml, openai.yaml, gemini.yaml, xai.yaml, openrouter.yaml,
# bedrock.yaml, vertex.yaml, fireworks.yaml, self-hosted-*.yaml

# Per-model streaming timeouts (stall vs overall)
streaming_timeouts:
  "o1":
    stall_timeout: 20m
    overall_timeout: 25m
  "o3":
    stall_timeout: 20m
    overall_timeout: 30m
  "gpt-5":
    stall_timeout: 20m
    overall_timeout: 30m

# Provider HTTP client timeouts
provider_timeouts:
  default: 60s
  bedrock: 90s
  api_call_default: 15m
  api_call_extended: 50m  # GPT-5.2 Pro reasoning

# YAML-driven parameter validation per model pattern
parameter_validation:
  enable: true
  model_drop_rules:
    "gpt-5": ["temperature"]
    "o1": ["temperature", "top_p"]
  conditional_rules:
    "claude*opus*":
      - if_parameter: "temperature"
        drop_parameters: ["top_p"]