⚙️ Customizing ICE V3

ICE is designed for surgical control. You can configure your cognitive engine at four distinct levels of granularity.

1. Global Defaults

Set app-wide defaults before initialization.

import ice

ice.set_defaults(
    post_limit=16384,
    max_vram_mb=12288
)

2. The Configured Client

For isolated environments (e.g., multi-tenant SaaS), instantiate a dedicated Client.

# Creates an isolated instance for a specific tenant
acme_client = ice.Client(
    tenant_id="acme-corp",
    client_budget=ice.CognitiveBudget(post_limit=32000)
)

3. Surgical Request Overrides

For individual queries that require extreme precision or massive context windows, use QueryConfig.

response = await acme_client.query(
    "Deep analysis of financial contracts.",
    model="claude-3.5-sonnet",
    config=ice.QueryConfig(post_limit=128000)
)

Available Customization Features

The following settings can be configured via ice.CognitiveBudget:

Setting	Type	Description	Default
post_limit	int	Total token budget for the final LLM prompt.	8192
pre_limit	int	Max number of memory "needles" to inject.	5
max_vram_mb	int	VRAM limit for native kernel operations.	8192
fidelity	float	Semantic resolution of ingestion (0.0-1.0).	0.5
physics_engine	str	Execution mode: native or standard.	"native"

---

Next: Enterprise Escape Hatches