Knowledge sources: beyond crawling

Whatever you add, a crawled page, an uploaded file, or a manual entry, ends up as one or more chunks in the same searchable knowledge base. The assistant doesn’t treat one source as more trustworthy than another; it just retrieves whichever chunk actually answers the question.

The Documents page accepts PDF, DOCX, TXT, Markdown, and CSV files up to 10 MB each, uploaded one at a time. This is separate from crawling and is the right place for content that doesn’t live on a public page: internal policy documents, spec sheets, pricing tables you don’t want indexed by search engines. You can also filter your uploads by type and status, and re-index a file if processing fails.

If you already maintain a list of frequently asked questions, the Documents page has an FAQ import that turns each question and answer into its own chunk. This tends to produce sharper answers than a crawled FAQ page, since the question-answer pairing is explicit rather than inferred from surrounding text.

For a single fact you want the assistant to know right now, without uploading a file, the Knowledge Base page has an “Add knowledge” option: an optional heading plus a block of content. It’s indexed and answerable immediately, with no crawl or file involved.

Every chunk, regardless of source, is listed on the Knowledge Base page and can be searched by heading or content, or filtered by source type. You can edit a chunk’s text directly or delete it. One thing worth knowing: if you edit a chunk that came from a crawled page, your edit will be overwritten the next time that page is recrawled, since recrawling rebuilds page chunks from the live page. For a correction you want to stick, add it as a manual entry or fix it on the source page itself.