The six kinds of entry we write
HighLevel doesn't prescribe topics. In our builds we sort knowledge base content into six kinds of entry, because it makes a missing answer easy to spot:
- About: who the business is, what it does and who it serves.
- Pricing and Programs: what is offered and what it costs, as the business publishes it.
- Client Experience: what working with the business is like, from first contact on.
- Business and Operations: hours, locations, and how scheduling and service work.
- AI Agent Instructions: how the AI should behave, written for the AI.
- FAQ: the questions customers really ask, each answered in a sentence or two.
Match each entry to a source type
That grouping is our convention. HighLevel's own source tabs are a different list: Web Crawler, FAQ, Web Search, Tables, Rich Text and File Upload. Any of our six can live in whichever tab suits its shape.
- Tables suit structured data, such as a pricing grid or a list of services.
- FAQ suits single questions with short answers.
- Rich Text suits content you'll keep editing directly in HighLevel. It is a separate source from any uploaded document.
- File Upload takes documents. HighLevel's document article lists DOC, DOCX and PDF; another of its articles names more types, so check the upload screen before relying on anything else.
- Web Crawler pulls in pages from the website, which is where we start.
Start from the website, then clean up the crawl
The business's own website is our first source. It is already written for customers, and it says what the business has chosen to say in public. Call transcripts are never a direct source in our builds. When a call settles a fact, the fact gets written up as a clean entry and approved before it goes in.
Read what the crawler brought back before you save it. Look for outdated pricing, navigation text repeated from page to page, and anything that has nothing to do with what a customer would ask.
You can edit crawled content inside HighLevel without changing the live website. A later re-crawl can overwrite those edits, though, so recheck your corrections every time the site is crawled again.
A re-crawl can undo your edits. Check them after every crawl.
Rather have it set up for you?
We write every knowledge base entry out in a plan you approve, then load and test it, starting with a 30-minute 1:1 with Rob.
What to keep out
Everything in a knowledge base is something the AI might say. These don't belong there:
- Transcript fragments: speaker labels, timestamps, half-sentences from a meeting.
- Internal notes, or anything else written for the team instead of the customer.
- Placeholders and template text that was never filled in.
- Sensitive material. HighLevel's document guidance is to remove it before you upload.
- Stale pricing, whether it came from the website or an old file.
- Old copies of a revised document. Uploading a revision adds another source, it doesn't replace the first, and conflicting or duplicate sources can lead to the wrong one being retrieved.
- Anything that exists only as an image. Document ingestion extracts the text and skips embedded images.
Write documents the AI can search
Give uploaded documents clear headings. HighLevel notes that headings help structure the content it searches, and they also make your own review faster.
Then attach the knowledge base, because creating one doesn't connect it to any agent. Attach it through Conversation AI > Agent List > Bot Training, wait for indexing, then ask the Retrieval Tester the questions customers really ask and read which sources come back. A wrong source points to the content: a duplicate to delete, a heading to add, or a stale copy to remove.