Tune how strictly answers stick to sources

A slider between source-bound and free-to-generalize. Set the default in account settings, override per chat.

Every Focus answer sits somewhere on a spectrum between "only what's in the sources" and "also general knowledge." The source adherence slider is how you set that bias.

Two poles

High adherence Low adherence
Answers come almost entirely from the dataset Model draws on general knowledge to fill gaps
Heavy on citations Lighter on citations
Won't speculate past what's in sources Will speculate when sources are thin
Good for: fact-checking, citation-heavy work, regulated content, audits Good for: ideation, broad analysis, sparse datasets, exploratory questions

The slider sets the chat's grounding behavior consistently across every model in the catalog; switching models doesn't change the adherence setting.

When to push toward high adherence

  • The answer must be defensible against the source documents (legal, compliance, finance).
  • You don't want the model filling in what's missing. You'd rather see an explicit "the documents don't say."
  • You're building something where wrong-but-confident is worse than slow-but-grounded.

When to push toward low adherence

  • You're brainstorming or exploring; the sources are inputs, not the only allowed material.
  • Your dataset is small and you want the model to use general background.
  • You're asking how-to or methodology questions that aren't in the documents anyway.

Default and per-chat override

Set your default in account settings → Adherence. Most Focus chats inherit it.

When a specific chat needs different behavior, open the chat's settings and override the slider for that conversation only. Useful when the rest of your work is exploratory but this one chat needs to hold tight to the sources (or vice versa).

What it isn't

  • Not a hallucination filter. High adherence biases the model toward citation-anchored answers, but doesn't make hallucination impossible. Citations and a stronger model help further.
  • Not a model setting. Switching models doesn't change your adherence; the slider applies to every model the same way.
  • Not a hard rule. Like custom instructions, it's strong guidance, not enforcement. For high-stakes claims, ask the chat to quote the source verbatim.

What's next

Was this helpful?