Decision Models
A small, fast model that makes typed judgments for Fabric — yes/no, pick one, score — in a fraction of a second. It powers Guarded mode, adaptive thinking, the decide tool and more.
A decision model answers questions with probabilities instead of prose: Is this command risky? How much reasoning does this message need? Which of these six folders does each email belong in? It answers in roughly 0.05–0.5 seconds for a fraction of a cent, which makes it cheap enough to ask before every tool call or every message.
Fabric uses one for a set of features you switch on individually. It is off by default and you bring your own key (or run one locally for free).
Which decision model?
Fabric speaks one protocol, so any of these work:
| Model | Runs | Good for |
|---|---|---|
| Jev (TypeSafe AI) | In the cloud — directly, or through OpenRouter or Vercel AI Gateway | The reference model. Reuse an OpenRouter or Vercel AI Gateway key you already have in Fabric. |
| Clef and Clef Flash (Cloudflare) | Cloudflare Workers AI, optionally through your AI Gateway | Very low latency (Clef Flash: median ~40 ms), 64K context. Ideal for Guarded mode. |
| Laya | On your machine | Free and open source. Nothing leaves your computer. |
| Custom | Any compatible server | Self-hosted deployments. |
See Set up a decision model for step-by-step instructions for each.
What it does in Fabric
Each feature has its own switch under Settings → AI → Decision model (Jev) → Advanced settings:
Guarded mode
Let the agent work without prompts, and ask only before risky actions.
Adaptive thinking
Think less on simple messages: faster and cheaper turns, never below what a request needs.
Smart features
Large results, suggestions, mid-turn messages, titles, risk badges, automations and tasks.
The decide tool
Let the agent classify, route and score hundreds of items at once.
Quick start
- Open Settings → AI and find the Decision model (Jev) card.
- Switch on Enable decision model.
- Choose a Provider and add its key (or pick Laya and start the local server).
- Press Test. You should see
Connected · <model> · <latency> ms. - Open Advanced settings and switch on the features you want. A good first set: Guarded mode, Adaptive thinking and Risk badges.
Safe by design
- A decision never grants permission. It can add a prompt, a hint or a skip; it can never turn a prompt into an approval.
- When the model is unsure, slow or unreachable, Fabric behaves exactly as without it — except Guarded mode, which asks you when it can't check a call.
- Each feature reports what it did. Under Advanced settings every feature shows its checks, changes and failures over the last 7 days, and says so when it has run 30 checks without changing anything — a hint that you can switch it off.
Privacy
What a feature sends goes to the provider you chose (with Laya, nowhere). Every call is recorded in ~/.fabric-agent/logs/decisions.jsonl with the questions, the answers and a hash of the input — never the input text itself.
Databricks Mosaic AI
Connect Databricks Mosaic AI Model Serving endpoints to Fabric Agents using a workspace host and bearer token. Serving endpoints are discovered automatically, with optional AI Gateway route overrides.
Set Up a Decision Model
Connect Jev (TypeSafe AI, OpenRouter, Vercel AI Gateway), Cloudflare Clef (Workers AI or AI Gateway), a local Laya server or your own endpoint.