Skip to main content
KBs are not static. Products change, policies update, support articles get rewritten. This page covers the operational side of keeping a KB current.

​
Re-ingesting a source

Source content is frozen at ingestion, and there is no in-place “refresh.” To pull the latest version of a page or file, delete the source and add it again:
  1. Open the KB → click the source.
  2. Delete it.
  3. Add it again (same URL or file). It goes back through converting_to_markdown → chunking → embedding → ready.
For URL content you can also re-run a URL-pattern attach to re-pull a batch of pages.

​
Delete a source

From the source list, click the menu → Delete. Embeddings are removed immediately; subsequent queries will not include this source’s chunks.

​
Bulk operations

For KBs with many sources, the UI gets tedious fast. Use the API (all paths under /knowledge-base):
OperationEndpoint
List sources in a KBGET /knowledge-base/datasets/{id}/sources
Attach source(s)POST /knowledge-base/datasets/{id}/sources
Attach by URL patternPOST /knowledge-base/datasets/{id}/sources/pattern
Detach a sourceDELETE /knowledge-base/datasets/{id}/sources/{dataset_source_id}
Detaching uses the dataset_source_id (the attachment id from the sources list), not the underlying source_id. Detaching removes the source’s embeddings from that KB without deleting the source itself.
A common pattern is a nightly workflow that:
  1. Pulls a list of canonical URLs from your CMS
  2. Diffs against the KB’s current sources
  3. Adds new ones, re-ingests changed ones (delete + re-add), deletes orphaned ones

​
Switching embedding models

Changing the embedding model on an existing KB triggers re-embedding of every source. During re-embedding:
  • The KB stays queryable using the old embeddings
  • New embeddings are computed in the background
  • Once complete, queries switch over
Plan re-embedding for off-hours on large KBs — it incurs the corresponding embedding cost.

​
Versioning content

KBs themselves are not versioned. There’s no “publish” step on a KB — changes are live as soon as a source’s status reaches ready. If you need versioning (e.g. to roll back content updates), the pattern is:
  1. Create a new KB (Product FAQ v2)
  2. Update agent attachments to point at it
  3. Test
  4. Delete the old KB once confident

​
Monitoring

The KB list page shows daily source counts and total source counts per KB. For deeper monitoring:
  • Use the API to fetch source counts and statuses
  • Set up a workflow on a daily cron that alerts if any source status is failed

​
Limits

  • Plan-based source counts per KB
  • Embedding rate limits inherited from your OpenAI/Azure quota when using your own keys