Building
Sources
A source is one item of material inside a knowledge base. There are three kinds, and choosing the right kind for a given piece of content has more effect on answer quality than anything else you can control.
The three kinds
| Kind | Best for | Limits |
|---|---|---|
| Question and answer | Policies, prices, opening hours, anything with one correct wording | Question up to 2000 characters, answer up to 8000. Up to 100 pairs per batch |
| Document | Handbooks, manuals, contracts, existing written material | PDF or DOCX, up to 25 MB per file |
| Text | Short notes and anything you can paste, with no file to hand | 20 to 200,000 characters of body text; optional title of 1 to 500 characters |
Question and answer pairs are matched most precisely, because you have written both sides. When a fact must come back in exact wording, such as a refund window or a warranty period, prefer a pair over a paragraph buried in a document.
Working with each kind
Question and answer
Enter pairs one at a time, or paste a batch of up to 100 when you are importing an existing FAQ. Write the question the way a visitor would ask it, rather than the way your team refers to it internally. Write "How do I return something?" instead of "RMA policy".
Several phrasings of the same question are useful rather than wasteful. Visitors ask the same thing in different words, and each pair gives the assistant another route to the same answer.
Documents
Upload a PDF or a DOCX file up to 25 MB. Text is extracted and indexed for you. Older .doc files are not accepted, so save them as DOCX first.
A scanned document is an image of text, not text. If a PDF was produced by a scanner or a camera and you cannot select words in it, nothing can be extracted and the assistant will not be able to use it. Run it through optical character recognition first, or add the key facts as question and answer pairs.
Text
Paste content directly, with an optional title. Useful for material that lives in a wiki or an email thread, where exporting a file would be more work than copying the paragraph that matters.
Use the text-source form for pasted text. A .txt file is not an accepted document upload. Give longer passages a descriptive title so you can find and maintain them later.
What happens after you add a source
Indexing runs in the background, so the page stays usable while it works. A source carries a status while that happens.
| Status | Meaning |
|---|---|
| Processing | The content is being read and indexed. The assistant is not using it yet |
| Ready | Indexed and in use. The assistant can answer from it immediately |
| Failed | The content could not be read. The usual cause is a scanned PDF with no text layer, or a corrupted file |
| Archived | Kept, but excluded from answers |
Large documents take longer than short ones. You do not need to keep the page open; the status updates on its own.
Writing material the assistant can use well
The assistant answers from what it finds. These habits have the largest effect on whether it finds the right thing.
- State each fact once. Two documents that contradict each other produce inconsistent answers, because both are legitimate matches.
- Keep answers self contained. A paragraph that begins "as mentioned above" loses its meaning when retrieved on its own.
- Prefer specifics over promotional language. "Returns accepted within 30 days of delivery" is usable; "easy returns" is not.
- Remove material you no longer stand behind. Deleting a source removes it from answers straight away.
- Add the questions you actually get. Support inboxes and chat logs are better sources of questions than an internal brainstorm.
When the material is in place, confirm the result on the Playground before pointing any visitors at it.
Resolve conflicting material
If two active sources disagree, the assistant may retrieve either. Review the source status, dates, and scope; then update or archive the obsolete material. Retest the same question in the Playground after indexing completes.
The public chat response does not include source citations or a provenance field. Use the dashboard's source and Knowledge Gaps workflows to investigate answers rather than expecting a citation in the API payload.
After adding or removing material, retest in the Playground using the exact wording a customer used. Indexing changes what gets found, and a question that worked last week can start matching something else.
Indexing is automatic and runs in the background, so a source is not usable the instant you add it. Answer quality is a direct product of source quality. Contradictory or outdated material produces answers that look inconsistent, and no setting compensates for that.