Data sources

Source types

• Files — upload PDF, DOCX, TXT, or Markdown. Text is extracted, split into chunks, and embedded on upload.
• Website — give a starting URL and the crawler follows internal links, indexing each page it finds. The source card shows a live link count.
• Text — paste raw text directly, with a title, when there's no file or page to point at.
• Q&A — a question paired with an exact answer, for the specific phrasings you want answered a specific way rather than left to retrieval.
• Notion — connect a real Notion internal integration token and a page URL. The page (and its sub-pages) are pulled in as a source, re-fetchable any time your Notion content changes.
• Tickets — one click imports your own resolved and closed Helpdesk tickets as Q&A training data, so the agent learns from how your team has actually answered customers before.

Status and indexing

Each source shows a real character count, chunk count, and status (indexing, ready, or error with the actual error message). Only website sources need re-crawling to pick up changes — every other type is indexed as soon as you add it, since there's no external page for it to drift out of sync with.

Retrain re-crawls every website source on the agent. It's disabled until you have at least one website source, since there's nothing else for it to do.

Filtering, selecting, and deleting

Filter the list by type (All, File, Website, Text, Q&A, Notion, Tickets), select multiple sources with checkboxes, and delete them in bulk. Deleting a source removes its chunks immediately — the agent stops being able to retrieve from it on the very next message.

How retrieval actually works

A visitor's question is embedded and compared against every chunk on this agent by cosine similarity, across all source types at once — there's no separate ranking per type. The top 5 most relevant chunks are inserted into the model's context alongside your instructions, and the model is expected to answer only from what it was given, or say it doesn't know. Every citation shown to the visitor links back to the exact chunk and its similarity score, so you can audit exactly what the agent based an answer on.

Reusing data elsewhere

Backstage can list your data sources on request (see Backstage), and the source count shown on the Playground's Overview links straight back here — the two surfaces stay in sync automatically since they read the same underlying rows.