Check the Sections and Summaries
The Contents outline starts from the source document’s own headings. The reasoning model adds sub-sections by meaning, so the outline can differ from the document’s table of contents. Each section carries a summary. Research searches use these summaries to decide which sections to read. A summary that misses the point makes relevant content harder to find. What to look for: sections that merged two unrelated topics, or a summary that doesn’t capture what the section is about. How to fix it: make the source content’s headings clearer and its topic sentences more explicit, so the model has better signals for where sections start and what each one is about.Check the Chunks
Within each section, the text is split into chunks—the units retrieval searches and returns. Each chunk should hold one complete topic. A table with many rows, like a price list, gets one chunk per row by design. Chunks are capped at 1,200 characters. Longer content, such as a long code block, is cut at the limit. What to look for: chunks that mix unrelated topics, or a procedure, definition, list, or code block cut in half. How to fix it: restructure the source content so each heading covers one topic, each paragraph holds one idea, and anything that must be retrieved together sits under one heading within the maximum chunk length.Check the Images
With text-only reasoning and embedding models, only an image’s caption or alt text is kept, so retrieval can find the image only by that text. What to look for: images your users will ask about that carry no caption or alt text in the source. How to fix it: add a caption or alt text in the source content, or switch to a vision-capable embedding model and click Re-index on the document, so images become searchable by what they show. See Multimodal Retrieval for details.Apply a Fix
After editing the source content, bring the new version in:- For a document from a connected source, the next sync picks it up. To bring it in right away, click Sync now on the Source page.
-
For an uploaded file, delete the old document and upload the new one.
Uploading a file with the same name adds a second document instead of replacing the first.