Skip to content

Knowledge base overview

The four knowledge source types, the extract-chunk-embed-store pipeline that makes them searchable, and what each source status means.

The knowledge base is what your chatbot answers from. Nothing else is: the system prompt shapes how the agent talks, but every fact in a reply comes from a knowledge source you added.

The four source types

FAQ pairs are boosted in retrieval, which makes them the sharpest tool for fixing a single bad answer. See FAQ and manual text.

Each type has its own page: Files, URLs and crawling.

How a source becomes answerable

Adding a source does not make it searchable straight away. Every source goes through the same four-stage pipeline on a background worker:

  1. 1
    Extract — pull plain text out of the PDF, DOCX, HTML page, or CSV.
  2. 2
    Chunk — split that text into overlapping passages of roughly 1,500

    characters with a 200-character overlap, so a passage keeps enough context to stand alone.

  3. 3
    Embed — turn each chunk into a vector.
  4. 4
    Store — save the chunks and their vectors, scoped to your workspace.
Note:

This runs on a queue, not in the request. A source you added a second ago is not answerable yet, and "processing" is a real state rather than a spinner. Failures are retried automatically. A big PDF or a 50-page crawl takes noticeably longer than a pasted FAQ.

Source statuses

See Reindexing and freshness for reprocess, reindex, and the per-source toggle.

Isolation

Retrieval is scoped to the chatbot and its workspace. One workspace's content can never be used to answer another workspace's question, and one chatbot does not read another chatbot's sources — even inside the same workspace. If two bots need the same document, add it to both.

Limits

Your plan caps how many knowledge sources a workspace may hold and how large an uploaded file may be. Disabled sources do not count. If you hit the cap, disable or delete something, or upgrade — see Plans.

Where to go next

Last updated