What type of PR is this?

This is a bug fix because it reduces noisy visual image retrieval results for normal text-based chat questions.

What does this do?

This PR gates visual image embedding retrieval behind a lightweight visual-intent check.

Previously, multimodal retrieval always searched both text chunks and visual image embeddings whenever image embeddings were configured. That could cause unrelated image-
only matches to appear for ordinary text queries, for example when the user is asking about a date, invoice, payment, contract, or record name.

With this change:

• text and OCR chunk retrieval still run for every query
• image records can still be found through extracted OCR/text content
• visual image embedding retrieval only runs when the query appears to be asking about images, screenshots, scans, photos, or visual content
• normal document questions are less likely to receive unrelated “Visual image match” context

Which issue(s) does this PR fix/relate to?

N/A

Have you included tests for your changes?

No, I have not included new tests for this change. The update is intentionally small and was manually tested through the prompt retrieval/debug flow.

Did you document any new/modified features?

No, this is an internal retrieval behavior fix and does not add a user-facing feature.

Notes

This is a pragmatic guard to reduce noisy multimodal retrieval. It does not remove OCR/text retrieval for image records; it only avoids running visual embedding search for
clearly non-visual questions.