📚 ICE V3 API Reference
This is the definitive technical reference for the ICE V3.0.0 SDK. All configurations are type-safe and validated via Pydantic.
1. Core Classes​
ice.Client​
The primary interface for ICE. Recommended for production.
Constructor:
tenant_id(str): Unique identifier for the tenant.client_budget(Optional[CognitiveBudget]): Instance-level execution budget configurations.pre_query_hooks(Optional[List[Callable]]): Async functions to run before a query.post_query_hooks(Optional[List[Callable]]): Async functions to run after a query.custom_headers(Optional[Dict[str, str]]): Headers for outbound proxy calls.
Methods:
await .ingest(source, config): Voxelizes data from a URI.await .query(prompt, model, config): Executes an LLM turn with context.
2. Configuration Models (Type-Safe)​
ice.CognitiveBudget / ice.QueryConfig / ice.IngestConfig​
These models share the same schema. Use QueryConfig or IngestConfig for per-request overrides.
| Field | Type | Default | Description |
|---|---|---|---|
post_limit | int | 8192 | Max tokens for the final LLM prompt. |
pre_limit | int | 5 | Max "needles" retrieved from memory. |
max_vram_mb | int | 8192 | Hard VRAM cap for the native kernel. |
fidelity | float | 0.5 | Ingestion resolution (0.0 to 1.0). |
physics_engine | str | "native" | native (RCC) or standard. |
3. Global Functions​
ice.set_defaults(**kwargs): Sets application-wide defaults for theCognitiveBudget.await ice.ingest(...): Top-level convenience ingest (uses default singleton).await ice.query(...): Top-level convenience query (uses default singleton).
4. Security & Environment​
| Variable | Description |
|---|---|
ICE_LICENSE_JWT | Required. Your hardware-bound JWT license. |
ICE_TELEMETRY_SINK | Destination for crash reports (internal). |
ICE_AUDIT_LOG_SINK | Destination for your operational logs (S3/File). |