Reindexing and freshness

Content changes. These are the four controls that keep what the chatbot knows in step with what is true.

Reprocess one source

Reprocess puts a single source back through extract → chunk → embed. Use it when a source shows failed, when a URL's page changed and you do not want to wait for the schedule, or after you edit manual or FAQ text.

Reprocessing a disabled source re-enables it, so it has to pass the same plan-limit check as switching it back on. If you are at your source limit, the reprocess is refused until you free a slot.

Reindex the whole chatbot

Reindex queues every source on the chatbot for reprocessing. It is the right move after a large content change, or when answers look wrong across the board rather than on one topic.

It is also the slow option: a chatbot with many large sources takes a while, and each source is re-embedded. Reprocess the one source you changed when you can.

Enable and disable a source

The per-source toggle takes a source out of retrieval without deleting it. A disabled source:

• is never used to answer,
• does not count against your plan's knowledge-source limit,
• keeps its content, so re-enabling is instant.

That makes disable the right tool for seasonal content, and for fitting an existing knowledge base into a smaller plan. Re-enabling runs the limit check.

Scheduled re-crawl for URL sources

Each URL source re-crawls on its own cadence — daily, weekly, monthly, or off. The default is weekly. An hourly scan picks up only the sources that are actually due, so a monthly source is not touched 30 times.

Two details make re-crawls cheap:

• Content-hash diffing. If the page's content has not changed, the source is

  not re-embedded.

• Forced re-embed on a model change. If the embedding model changed, chunks

  are re-embedded even when the content is identical, so old-model chunks never fall out of search.

The Stale badge

A URL source shows Stale when it was last crawled more than a week ago — the same interval as the default cadence. The knowledge list also shows "crawled Xd ago" per source.

A source set to a monthly cadence will therefore wear a Stale badge for most of its cycle. That is the badge doing its job: it reports the age of the content, not a misconfiguration.

Inspecting chunks

Open a source and choose View chunks to see the passages the pipeline produced. This is the fastest way to diagnose a bad answer at the source:

• No chunks at all — extraction found no text. Common with scanned PDFs and

  JavaScript-rendered pages.

• Chunks full of navigation and boilerplate — the page's real content is

  thin; paste the useful part as manual text instead.

• The right text is there but never retrieved — that is a retrieval problem,

  not a knowledge one. See How retrieval works.

Where to go next

• URLs and crawling
• Evals and knowledge gaps