Knowledge base

On this page: Ways to teach · Documents · CSV files · FAQ CSV · Web training · Credits · Troubleshooting

Ways to teach your agent

Your knowledge base is what your chat agent answers from. In the dashboard it is the Knowledge Base section of the sidebar:

Upload a document

1. Open Documents → Upload.
2. Give it a title, pick the document type and choose the file: PDF, Word (.docx), plain text (.txt) or CSV. Keep each file under 10 MB, the limit the upload page states.
3. Press upload. The file is read in the background; its status changes to processed when your agent can use it.

A document is split into passages of about 400 characters (each overlapping the next by 80). Each passage is one knowledge record and uses one knowledge credit.

Upload a CSV file

A file whose name ends in .csv is always read as a table. The first row is the header. Column names are trimmed and not case-sensitive (Question and question are the same column). Empty rows are ignored. Save the file as UTF-8 (“CSV UTF-8” in Excel); Latin-1 is also read.

What happens to the rows depends on the header:

• With question and answer columns, the file is an FAQ CSV: one FAQ per row.
• Any other header: each row becomes one knowledge record, written as column: value | column: value from its non-empty cells. Rows with almost no text are skipped. This suits price lists, opening hours, branch lists and product tables.

FAQ CSV

question,answer,title,url,description,price
"What are your opening hours?","Saturday to Thursday, 10 AM to 8 PM.","Opening hours",,,
"How much is the Oslo sofa?","The Oslo 3-seater sofa is 48,500 taka.","Oslo sofa",https://yourshop.com/oslo-sofa,"Solid wood frame",48500

A row without a question or an answer is skipped. Each FAQ uses one knowledge credit. The FAQs appear in Training Data, where you can edit or delete them one by one.

Add one FAQ by hand

In Training Data, add an FAQ with a question and an answer. It is the same as one row of an FAQ CSV.

Train on your website

Available on these plans: Starter, Growth, Scale, Enterprise.

1. Open Web Training and enter your website address (for example https://yourshop.com; https:// is added when you leave it out).
2. Discovery. Central AI reads /sitemap.xml at the root of your site. For a sitemap index it follows the first 5 sitemaps. It lists up to 100 pages, your homepage included. When there is no sitemap, it lists your homepage and up to 50 links found on it that stay on exactly the same host (so www.yourshop.com and yourshop.com count as different sites). Links to login, sign-up, cart, checkout, account and admin pages, and to files, are left out.
3. Choose pages. Untick pages you do not want, or add a page by hand.
4. Scrape. The chosen pages are read in the background. Pages that need JavaScript are opened in a browser. Links to files (PDF, images, Office documents, archives) are skipped: upload those as documents instead.
5. Embed. Press embed to add the pages to your agent’s knowledge. Each page becomes one knowledge record and uses one knowledge credit.

• Your homepage is checked against your robots.txt before anything is read. Other listed pages are not checked one by one, so untick any page you do not want read.
• The agent is given about the first 4000 characters of each page. Put the facts that matter near the top of a long page, or upload them as a document.
• You can edit the text of a scraped page, or delete it, from the web training job.
• Pages are not re-read on their own. When your site changes, run web training again, and delete the old job so the agent does not hold two versions of a page.

Knowledge credits

Each knowledge record uses one knowledge credit: one FAQ, one CSV row, one document passage, one web page, or one record created through the Knowledge Sync API. Your plan includes:

• Uploads: when a file has more records than you have credits left, only the first ones are processed and the page tells you how many were skipped. With no credits left, the upload is refused.
• Web training charges one credit per page when the pages are embedded.
• Deleting a document, page or record removes it from your agent. It does not give credits back.
• Offerings (what you sell) never use knowledge credits.

Troubleshooting

The agent does not know something I uploaded.

Check the document’s status in Documents. A failed status carries its reason, for example no subscription or no knowledge credits left.

My FAQ CSV became plain rows instead of FAQs.

The header must contain both question and answer. Check for typos, and that the first line of the file is the header.

Web Training is not in my plan.

It needs one of: Starter, Growth, Scale, Enterprise. Until then, upload your pages’ content as documents or an FAQ CSV.

Discovery found only my homepage.

Your site may have no /sitemap.xml and few links on the homepage. Add pages by hand.