How do I connect a knowledge base to my assistant?

Three source types

• Website / URL — a single page, a crawl that follows same-site links from the page you give us (up to Max pages — 1 to 200, default 50 — minus any Exclude paths), or a sitemap.xml whose every entry is fetched, nested sitemap indexes included. Re-indexed automatically on the cadence you pick (every 6 hours up to weekly; daily by default).
• File upload — PDF/MD/TXT/DOCX up to 20 MB each; ideal for handbooks and PDFs you don't host publicly.
• Pasted text — a doc or FAQ pasted directly as Markdown or plain text.

Sources vs. assistants

Sources are indexed once per workspace and shared. Each assistant only searches the sources attached to it — the Attach to assistants list when you add a source, or the assistant's Knowledge tab afterwards. A source attached to no assistant is indexed but never used, and the Knowledge page flags it.

How retrieval works

1. Each source is chunked heading-aware, roughly 1,000 characters per chunk with a small overlap.
2. Chunks are embedded once at ingest time and stored in your workspace; a re-sync replaces the previous chunks.
3. On every user question, we embed the query and pull the top-K nearest chunks.
4. Chunks are injected into the system prompt with source URLs so the model can cite.

When a source ends up in Error

A large crawl is indexed in slices, and each slice is checkpointed, so a sync that is interrupted resumes where it stopped rather than starting over. That applies from the very first step: listing the pages to crawl is itself checkpointed, so a big site that dies while still being mapped is retried rather than being reported as a plain timeout.

A source that has never finished a sync is retried on a fast ladder — roughly 10 minutes after the failed attempt, then 20, 40 and 80 minutes — instead of waiting for its normal cadence. A source that has indexed successfully before keeps the content it already has, so it waits for the next scheduled run rather than retrying immediately. After the fifth consecutive failure a never-indexed source falls back to the normal (slower) cadence, and Re-sync always retries it right away.

• "Sync kept crashing and was stopped early (N of M pages fetched)" — we retried and gave up. Whatever was indexed before that point is kept and stays searchable. Re-sync to continue.
• "Sync kept crashing before it could list the pages to index" — the crawl never got past mapping the site. Re-sync, or lower Max pages / point the source at a smaller section.
• "Indexing stopped early: N pages in a row timed out" — the site is most likely rate-limiting the crawler. Re-sync later.
• "Sitemap contained no page URLs" — the URL is not a sitemap, or it is an index with no entries. Point the source at the homepage instead.

Improving answer quality

• Prefer sitemaps over hand-picked URLs — broader coverage wins.
• Keep headings descriptive; chunk boundaries follow your H1/H2 structure.
• Re-sync after major content updates so embeddings stay fresh.
• Add a system-prompt rule: "Always cite sources. If unsure, say so."