Configuration

Settings resolve in this order, later winning:

1. defaults
2. a YAML or JSON file, from piramid serve --config or CONFIG_FILE
3. PIRAMID__ environment variables, spelled from the path: runtime.wal.max_log_size is PIRAMID__RUNTIME__WAL__MAX_LOG_SIZE
4. --port and --data-dir

PIRAMID_API_KEY and OPENAI_API_KEY are environment-only. An unknown, misspelled or unimplemented key fails startup with a message naming it.

The file has three blocks. startup: applies once at boot. runtime: is re-read by POST /api/config/reload, except that inference is fixed at startup, and quantization, memory, wal.enabled and wal.sync_on_write cannot change while a collection is open. search.metric is copied into a collection when it is created. console: configures the console.

On a GPU, set startup.hardware.profile: gpu and runtime.execution: gpu; the model loads onto cuda:N at startup.hardware.gpu.device_ordinal. On the CPU, leave both out and set a KV cache budget:

runtime:
  inference:
    enabled: true
    model_path: ./models/Qwen2.5-0.5B-Instruct
    kv_cache:
      max_bytes: 2147483648