Knowledge is stuck processing

What it is

Knowledge does not become usable the moment you add it. Each source moves through a short pipeline, and the badge on its row tells you where it is:

Badge	What is happening	Do you need to act?
Not Trained	Added, not yet queued	Yes — start training
Pending Training	Queued, waiting for a worker	No — wait
Reading Document	A scanned or image-based file is being transcribed into text first	No — wait
Generating Embeddings	Being made searchable	No — wait
Embeddings Ready	Done	No
Embedding Failed	Something went wrong	Yes — retry or fix the source
Re-syncing	Content changed and is being refreshed	No — wait

Most "stuck" reports are a source in a waiting state that simply has not finished yet. Genuine sticking exists, and there is an automatic recovery for it, described below.

When you would use it

When a badge has not moved for longer than feels reasonable, when a crawl has clearly stalled, or when the chatbot is answering as though content you added does not exist.

Where to find it

Dashboard → Chatbots → your chatbot → Knowledge. Every source is listed with its badge, and failed rows can be opened for the reason.

Steps

1. Open Chatbots → your chatbot → Knowledge.
2. Read the badge on the row that is bothering you and find it in the table above.
3. If it is waiting, give it time. A short text note is quick; a large PDF, a scanned document, or a whole-site crawl legitimately takes minutes. Refresh rather than re-adding.
4. If it says Embedding Failed, open the row and read the reason. Match it in the section below.
5. Retrain what failed. Select the failed rows and use Retrain, or use the retry-all option if there are many. A banner appears when something needs it: "A knowledge update didn't finish, so some answers may be missing the latest content. Retry to restore it."
6. If it says the item is currently being processed, you cannot retry yet: "This item is currently being processed. Wait for it to finish (or stop it) before retrying." Wait for it to leave the processing state, or stop it first.
7. Check the size meter on the same page. A source rejected on size looks like a failure, not like a spinner. See Chatbot and knowledge limits.
8. Once it reads ready, test it. Ask a question that only that content can answer.

What each failure reason means

• The content produced nothing usable. An empty document, a PDF that is entirely images with no readable text, or a page that was blank when it was fetched. Fix: open the source and confirm it actually contains text. For a scanned file, see OCR and scanned documents.
• The processing service returned an error. Usually transient. Fix: retry. If it fails three times in a row, report it.
• It stalled and was recovered automatically. The system noticed and re-queued it. Fix: nothing — but check the badge later to confirm it finished the second time.
• You are over your knowledge limit. Fix: remove content you are not using, or buy more room. See Extra chatbots and extra knowledge.
• Removal failed or syncing failed. Fix: retry. These usually clear on a second attempt.
• An internal error. Fix: retry once, then report it with the source name and the time.

Failure reasons are deliberately plain-language rather than raw technical output, so what you see on the row is the whole story.

Automatic recovery

You are not the only safety net. Agentency runs a watchdog on a short cycle that looks for sources genuinely stuck rather than merely slow.

• A source queued but never picked up is re-dispatched.
• A source processing with no progress is marked failed and re-queued.
• A source stranded partway through document reading is restarted.
• A removal stuck partway through is retried.

The thresholds are in the region of an hour, so the honest guidance is: if it is still spinning after about an hour, the system will pick it up itself — and you can hit Retrain yourself at any point once the row is no longer processing.

What this means practically: waiting is a legitimate action, and re-adding the same file five times is not. Duplicates make your knowledge base worse and consume your size allowance.

While it processes

• Leave the chatbot Active. Pausing it does not speed training up and does break your hosted page and widget for real visitors.
• The chatbot keeps answering from whatever it already learned. Adding new knowledge never takes the old knowledge offline.
• Do not start overlapping crawls of the same site. They queue behind each other and make everything slower.
• Removal is asynchronous too. Several rows showing as removing is normal; wait for them to disappear.

Limits and plan notes

One source at a time per chatbot. Work on a chatbot's knowledge is serialised on purpose so that ingesting, retraining, and removing cannot collide. A queue behind a large crawl is expected.

Large crawls take proportionally longer. A hundred-page site is not a hundred times slower than one page, but it is not instant either. See Crawl status and history.

Failed training never deletes your source. The row stays so you can fix and retry it.

Being over the knowledge cap blocks new sources but leaves existing ones working.

A chatbot that has never completed a training run cannot answer at all. Visitors get a "not ready yet" message — see Untrained mode.

If nothing here works

Write to support with:

• the chatbot name and the account owner's email,
• the source name and its type — pasted text, uploaded file, crawl, or cloud import,
• the badge it is showing and the failure reason if there is one,
• when you added it, with your timezone,
• whether you have already retried and how many times.

See Contact Support. Do not attach the file itself if it contains anything confidential — the name and type are enough to start.

Common problems

Everything is pending and nothing moves.

Wait through one refresh cycle. If it is still pending an hour later, the watchdog should have re-queued it; if the badge has not changed after that, report it.

The badge says ready but the chatbot ignores the content.

Ask using the words that appear in the source. If it still misses, the passage may be too thin to match anything — add a short, plainly worded text note that states the fact directly.

Retrain is greyed out.

The row is currently processing. Wait for it to leave that state, or stop it first.

I uploaded a scanned PDF and it produced nothing.

Image-only documents need transcribing before they contain any text. See OCR and scanned documents.

Should I delete everything and start again?

No. It rarely helps, it loses content you will have to re-add, and it puts you at the back of the same queue. Retry the specific failed rows instead.