Troubleshooting: knowledge
Under Knowledge you set what your assistant knows about your business. When a source doesn't work as expected, the cause is usually the document's settings, the format of the source, or the way the question was phrased. Below are the most common problems and their fixes; check every change with Test search, at the top of Knowledge.
Symptoms, causes and fixes
| Symptom | Likely cause | Fix |
|---|---|---|
| PDF is ignored | Document disabled, not ready yet, in another collection or not chosen on the Tasks tab, under Use knowledge | Enable the document and choose it on the assistant |
| PDF yields no knowledge | The file is a scan with no text layer | Upload a text PDF or a text file |
| Website scan comes back empty | Site blocks automated reading, loads content via JavaScript or requires login | Upload a PDF or add Q&A entries |
| Only some pages read | Limit of 50 pages (25 without a subscription) or a lower Page limit you set, same domain only | Raise the Page limit or add key pages as Q&A or documents |
| No proposals appear | No sources and no finished conversations yet, or another collection selected | Add sources, select the right collection, check back after a few minutes |
| Exact values changed | The AI model paraphrases | Add the rule "Repeat exact values literally" and put values in Q&A form |
| Q&A shows up in the wrong assistant | Q&A entries apply to every assistant that uses the same collection | Give the assistant its own collection |
| Answer not found although it is in a document | Question phrased very differently from the document | Add a Q&A phrased the way callers ask |
| Assistant uses an old version of a document | The old document is still enabled, or the old value is also in a Q&A entry | Disable or delete the old document, and update the Q&A entries |
A PDF is ignored
A document you've uploaded is only used once it has been processed, is enabled, is in the assistant's collection and may be searched by that assistant.
- Go to Knowledge → Sources, select the assistant's collection at the top, and check the document has the status Ready and is not set to Disabled. If it is, click Enable.
- Open the assistant, go to the Tasks tab, open Use knowledge and check Collection. If Documents is set to Only the documents I choose, tick the document. With All documents that isn't needed.
- Test with Try it out, using a question whose answer only appears in that document.
See Knowledge for how documents work.
A PDF yields no knowledge
A scanned PDF consists of page images with no text we can read. The document then gets the status Failed with the message "This PDF contains no text we can read. Is it a scan?". If only some pages are scanned, the document does reach Ready, but those pages are missing. Export the source file again as a text PDF (from Word or your quoting software, for example), or copy the text into a text file (.txt or .md) and upload that.
The website scan comes back empty
Some websites can't be read: they block automated reading (through robots.txt, for example), only show content after running JavaScript, or sit behind a login page. If nothing readable is left as a result, the website gets the status Failed under Knowledge → Sources, with the message "No readable pages were found. The site may block automated reading. Put the content in a PDF or in Q&A instead." Click View and check Pages skipped for the reason. If some text does come through, such as only the login page, the website shows Ready but the real content is missing; you can see this under View, at Pages read. Upload the most important pages as a PDF, or turn the content into Q&A entries, which also gives the most reliable answers.
Only some pages are read
The scan reads at most 50 pages (25 without a subscription) within 5 minutes and only follows links within the same domain; with or without www counts as the same domain. It reads the sitemap when there is one, and it prefers contact, services and pricing pages over the blog. Parts of large sites and pages on a subdomain are therefore left out. On the website's row in Knowledge → Sources, click View and check Pages skipped for the reason; pages on a subdomain aren't listed there. Add a subdomain as a website of its own. The same panel has the Page limit: if it is set lower, the scan stops there. Raise it, click Save rules, close the panel and click Refresh scan on the website's row.
To read only certain pages, choose Custom paths in the same panel, set Only these paths (for example /pricing) and click Save rules. Close the panel and click Refresh scan on the website's row. Every scan counts towards your limit: 3 a day and 6 in total without a subscription, 10 a day with one. You can still add prices, the service area and warranty as a Q&A entry.
No proposals appear
Proposals under Knowledge → Questions & answers → To review are generated from the sources you've added and from questions callers actually asked that the assistant couldn't answer. They belong to a collection, so first check you've selected the right one at the top. Without sources or finished conversations there is nothing to suggest.
Add one or more sources first and check back after a few minutes: processing isn't instant, a thorough scan takes about 3 to 10 minutes, and while it runs the filter shows To review (creating…). If the list stays empty after that, click Process again or Refresh scan on the source's row. If you ticked Add the proposals to Q&A right away for the website, the proposals go straight to Active and not to To review.
A test with Try it out usually doesn't create a proposal. Only Real call always counts as a real conversation here. Any other test can only count with Perform actions for real on, and when the assistant itself ends the conversation.
Exact values are changed
The assistant is told to repeat prices, opening hours, addresses and numbers exactly, but the AI model can still paraphrase: €89.50 becomes "about ninety euros" and "8am to 5.30pm" becomes "during the day". On the Instructions tab, under Important rules, add "Repeat exact values literally, including prices and times" (in Advanced mode it goes in Full prompt), and put key values in Q&A form with the amount or time in Short answer. See How to write a great prompt.
A Q&A shows up in the wrong assistant
Q&A entries apply to every assistant that uses the same collection. If one assistant for emergencies and one for sales share a collection, each also knows the other's answers. To separate them, select Manage collections at the top of Knowledge, click New collection, and choose that collection on the right assistant, on the Tasks tab, under Use knowledge, at Collection. A new collection starts empty, with its own business profile.
If you only want to separate documents, you can do that within one collection: on the assistant, choose Only the documents I choose and tick the right documents.
The answer is not found although it is in a document
First check the assistant may search the document (see A PDF is ignored). The assistant searches by meaning, so different words can still match, but it only gets back the best-matching pieces of text. If the question differs a lot from the document's wording ("call-out fee" versus "attendance charge", for example), the right piece can be missing. Add a Q&A entry with the question phrased the way callers ask it: the assistant always knows its Q&A entries. Use Test search with the Same meaning option, which is how the assistant searches, to confirm the answer is now found.
The assistant uses an old version of a document
If you've uploaded a new price list but the assistant still quotes old prices, the old price list is probably still enabled alongside the new one. On the old document's row in Knowledge → Sources, click Disable or Delete. If the assistant is set to Only the documents I choose, tick the new price list as well.
Then check Knowledge → Questions & answers and Knowledge → Business profile. Deleting a document doesn't remove Q&A entries or profile text taken from it, and the assistant keeps using those. Update them, so there is only one source of truth.
Still stuck?
Can't find your problem here, or did the fix not help? Let us know via Report a problem and mention which source is affected and which question you tested with Test search.