Improvements to how the AI Assistant reads your documents
Hi everyone đź‘‹
We're rolling out a set of improvements to how the AI Assistant reads your documents. Some are available now, and more are coming between now and mid-December. Here's what's changing and when.
You don't need to do anything. Your content picks up the changes automatically when it's reprocessed, and it stays searchable the whole time.
Â
âś… Available now
Available in Japan, Canada, the EU (Frankfurt), Australia, the UK, US-East, US-West and GovCloud.
Better handling of document structure
The assistant splits your documents into short passages before it searches them. Those passages follow the structure of the document:
- Table headers stay with their rows, including in tables with merged cells or cells that span several lines.
- Headings, lists and nested sections stay intact, including in Word documents.
- Very short fragments, such as a stray line or a heading on its own, are combined with the content around them instead of becoming separate passages.
What you may notice: more accurate answers to questions about tables, and fewer answers that stop partway through a section.
Links inside PDFs are retained
Links in PDFs keep their URLs, so the assistant can cite and link to the resources your documents point to.
⏠Early December: wider context for answers
The assistant matches queries to the chunks that best answer a question. With this update, it also reads the text just before and after each match, and adjusts how much based on the question. Search still finds the most precise passage; the assistant just sees more of the document when it writes the answer.
What you may notice: more complete answers, especially for step-by-step procedures and policies where the key detail is a paragraph or two away from the best match.
This is already available in GovCloud. It will be enabled progressively for other regions in the first week of December.
🔄 Now through mid-December: PDF and SharePoint improvements
Better PDF reading
PDFs will now be read much more like web pages:
- Sections, headings and reading order are preserved.
- Tables with ruled lines keep their rows and headers.
- Words split across two lines with a hyphen are joined back together (for example, "reimburse- / ment" becomes "reimbursement").
- When the same link text appears more than once, each link points to the right URL.
Note: scanned, image-only PDFs aren't covered by this change.
FAQ questions stay with their answers (PDF)
In FAQ-style PDFs, numbered questions (for example, "3. How do I request a new laptop?") stay together with their answers, so the assistant can match a question to the right answer.
Spacing fix for SharePoint pages
Spaces next to links or bold text on SharePoint pages are kept, so answers no longer show run-together words like "benefitson offer".
We expect these to be available in all regions by mid-December.
Â
Â
Â