My answer looks wrong

If a Focus answer doesn't match what's in your documents, work through these in order. Most "wrong" answers fall to one of the first three.

1. Check the dataset

The fastest cause is that the dataset isn't what you thought it was. Open the chat's dataset picker:

• Are the documents you expected actually in it?
• For a label-scoped Focus: did you add the right label? Has the label's membership changed since you opened the chat?
• For a Smart Search-built selection: did the selection include something off-topic that's pulling the answer sideways?

Mid-chat, you can add or remove sources from the dataset and ask again.

2. Verify the citations

Click the citation tokens in the answer. Do they actually land on the page/section/cell the answer claims they say what the answer says?

• If citations exist but the cited content doesn't support the claim, the model is grounding incorrectly; try raising source adherence and asking again.
• If citations are missing entirely, the model didn't ground itself in your sources; also a source adherence issue.
• If citations point to an OCR'd page that's garbled, the underlying OCR is the problem; run Enhanced OCR.

3. Rephrase

Same question, different wording, often pulls in different passages of the sources and produces a different answer.

• Use vocabulary from the documents (see Phrasing tips).
• Ask for the specific shape: "Quote the exact clause" vs. "What does it say".
• Split a dense question into smaller ones.

4. Switch to a stronger model

If rephrasing and dataset checks don't help, try a stronger model. Reasoning models often handle ambiguous or dense documents better than fast baseline models.

• Open the model picker, pick a reasoning-capable or premium model.
• Re-ask the question.

If you're already on a strong model, it's likely a content or grounding issue rather than a model capability one.

5. Check the source adherence setting

If the answer feels like it's drawing from general knowledge instead of the documents, raise source adherence. At high adherence, the model is biased toward citation-anchored answers and is more likely to explicitly say "the documents don't say."

6. Try a workflow instead

If the question is shaped like "do X to each document" ("summarize each contract", "extract dates from these") a chat isn't the right tool. Run a workflow instead.

What it might not be

Wrong answers usually aren't:

• A bug in retrieval: the retrieval layer is the same one workflows use, and they work well across millions of documents.
• A missing document: easy to check by looking at the dataset.
• A capacity issue: those produce errors, not wrong answers.

What's next

• Phrasing tips: rephrasing as a first move.
• Tune source adherence: turn the bias up.
• Pick a model: model choice matters.
• Enhanced OCR: if scans are the underlying issue.