Sources and citations

What it is

A source is a piece of your own knowledge that supported an answer — a chunk of an uploaded file, a paragraph of pasted text, or a crawled web page. Sources are the evidence trail behind every grounded reply, and they are what makes an Agentency chatbot auditable rather than merely plausible.

There are two separate audiences for them, and confusing the two causes most of the questions on this page:

• You, the owner, always see them. Every answer in the transcript and in Test Chatbot has an inspectable list of what was used, regardless of any setting.
• Visitors see links only when you turn them on, and only when the source has an address to link to.

A source is knowledge. A channel is where the person was sitting. Different things entirely.

When you would use it

• Verifying that the chatbot cited the right policy, not a similar-looking one.
• Investigating an answer that looks invented.
• Deciding whether a wrong answer is a retrieval problem or a knowledge problem.

Where to find it

Open Dashboard → Chatbots → your chatbot → Conversations → a thread, and open the sources control on a reply. The same control appears inside Test Chatbot.

Steps

1. Open a conversation with View details, or open Test Chatbot from the chatbot header.
2. Find an answer you want to check and open its sources control. It shows the number of sources used.
3. Read the panel. For each source you get its title, the matched text, its scores, and — where one exists — a link to the original page.
4. Compare the original query with the search query used. If they differ, that rephrasing is what retrieval actually looked for, and a bad rephrasing is a diagnosable cause of a bad answer.
5. Check the scores. A source flagged as a weak match means the chatbot was scraping the bottom of what it had.
6. Fix what the panel reveals: the wrong document means your titles or content are ambiguous; no sources at all means the knowledge is missing.

What the sources panel shows

• Source number and title — the knowledge item. Meaningful titles pay for themselves here; "Document 1" tells you nothing.
• The matched text, with a Read more control for longer chunks.
• Semantic score — how closely the meaning matched.
• Reranker score — a second, more careful pass over the shortlist.
• Weak match label when a source scraped in below the comfortable bar.
• Collection — which knowledge collection it came from. See Collections and ownership.
• Open source page — only for knowledge that has an address, meaning crawled pages.
• "No longer available" — the knowledge was updated or removed after this conversation. The transcript is historical; the knowledge base moved on.

Showing sources to visitors

The switch is Show reference links on Settings → AI configuration. See Chatbot AI settings.

On — answers can include links back to the source page.

Off — the chatbot produces a links-free, prose-only reply. This does more than hide a chip: the answer text itself is generated without URLs or a resources block.

The constraint that surprises everyone

Only crawled web pages can be linked. A crawled page has an address. An uploaded PDF and a block of pasted text do not — there is nothing to point a link at.

So a chatbot whose knowledge is entirely files and pasted text will show visitors no links no matter how the toggle is set. That is correct behaviour. If you want clickable citations, crawl the website that publishes the same material, and let the chatbot use the crawled version.

Either way, you keep the full inspection panel in the dashboard. The toggle only governs the visitor's view.

Using sources to diagnose a bad answer

The panel turns "the chatbot is wrong" into one of four specific, fixable problems:

What the panel shows	Diagnosis	Fix
No sources at all	Nothing matched — the knowledge is missing or untrained	Add the document; confirm it is trained
The right document, wrong section	Retrieval landed in the wrong part of a long file	Split the file, or add clear headings
The wrong document, high scores	Two documents look too similar	Retitle them distinctly; delete the outdated one
Right sources, wrong answer	Retrieval worked, generation did not	Add an instruction; consider Long Response

The third row is the most common in mature knowledge bases: an old and a new version of the same policy both sitting there, and no way for the chatbot to know which one you meant. Deleting the old one fixes it immediately. See Remove knowledge.

What you will see

A sources control on answers that used knowledge, opening into a panel of numbered sources with their scores and text. Turns with nothing to cite — greetings, small talk, fallbacks — have no control at all, which is not a failure.

Limits and plan notes

Citations require trained knowledge. Turning the toggle on does not invent sources, and an untrained chatbot has nothing to cite. See Untrained mode.

The chatbot never browses the open web. Every source is something you put into its knowledge base.

Reading sources needs permission to view conversations and spends no credits.

Common problems

Visitors do not see links even though I enabled them.

In order of likelihood: your knowledge has no crawled pages (uploads and pasted text have no URL), the answer used no knowledge at all, or you have not saved the AI configuration tab.

A source title is useless, like "Document 1".

Rename the knowledge item. Titles appear in citations, and clear ones make both your debugging and the visitor's experience better.

The chatbot cited a document that does not answer the question.

Look at the scores. Weak matches mean it had nothing better. Add the document that does answer it, then re-ask in Test Chatbot after training completes.

The search query used is different from what the visitor typed.

That is deliberate — the question is rephrased for retrieval, especially in a multi-turn conversation where "how much is it?" only makes sense with the previous message. If the rephrasing looks wrong, that is a genuine finding and usually points at a confusing earlier turn.

Can I hide sources from visitors but keep them for myself?

That is exactly what the setup does. Turn Show reference links off; the transcript and Test Chatbot panels are unaffected.