Setup & performance

Cookbook: Hardware-Aware Model Tuning

The Cookbook is GCTRL's built-in model tuning assistant. It profiles the machine GCTRL is running on and recommends the right model for each pipeline stage - instead of you guessing at model names and finding out the hard way that a stage is too slow, or that the hardware can't carry it at all.

Status: this page is a skeleton for the Cookbook feature shipping in this release. It will expand with screenshots and a full walkthrough as the page sees real usage.

What it does

  • Detects your hardware - CPU, RAM, and GPU/VRAM where applicable - and classifies the machine's tier.
  • Recommends a model per pipeline stage, matched to that tier:
    • KEX embedding
    • KEX relation extraction
    • FUSE distill
    • Pi agent (the built-in GCTRL agent)
    • Talk-to-Graph
  • Shows an impact estimate for each recommendation - the expected quality/speed tradeoff of switching, so the choice isn't blind.
  • Guardrails the choice - if a recommended model repeatedly fails at runtime (crashes, timeouts), the Cookbook automatically reverts that stage to its safe default rather than leaving the pipeline stuck on a model the hardware can't run.

Where to find it

Settings → Cookbook, alongside Settings → AI Models (the manual picker). The Cookbook is the guided, hardware-aware path to the same underlying settings - use either, they stay in sync.

See also

LLM Providers · Infrastructure & Ollama · Performance Guide