Module · Knowledge

Create a Knowledge Base with AI

Lesson 5 of 6 · 7 min

Pasting works when the text is already good. Most content is not. A price list lives in a PDF with tables. A policy lives on a web page wrapped in navigation, a cookie banner and three calls to action. Paste either in and you get chunks full of fragments and furniture.

AI Import is the other door. You give it a source — a URL or a file — and it fetches the content, strips everything that is not content, restructures what is left, and puts that in your Knowledge Base.

What actually happens

It is worth knowing the pipeline, because you have already met every part of it in this course:

Step	What runs
	A web scrape for a URL, or the document parser for a file. The same two nodes from the previous path, with the same limits — a client-rendered page or a scanned PDF gives it very little to work with.
	A model is asked to turn that raw text into structured markdown and return a title alongside it.
	The cleaned markdown is added to the Knowledge Base and chunked exactly as pasted content is.

What runs: A web scrape for a URL, or the document parser for a file. The same two nodes from the previous path, with the same limits — a client-rendered page or a scanned PDF gives it very little to work with.

What runs: A model is asked to turn that raw text into structured markdown and return a title alongside it.

What runs: The cleaned markdown is added to the Knowledge Base and chunked exactly as pasted content is.

The cost lives in step two. Pasting content in costs nothing. Importing runs a model over the whole document, so a hundred-page PDF is a hundred pages of tokens. That is usually worth it once, per document — and worth remembering before you import a folder.

There is a third source the screen does not offer. The importer accepts a URL, a file, or a search — a topic or company name, which it looks up and then crawls. The modal has two tabs, so that third route is reachable by an agent asking to ingest something and not by a person clicking through. Worth knowing exists; not something you can use from this screen.

The trade you are making

The stored text is not your text. Plain import stores exactly what you pasted. AI Import stores what the model wrote after reading what you supplied. It is cleaner, better structured and easier to retrieve from — and it is a paraphrase. For a product description that is an improvement. For a returns policy, a warranty term, a legal notice or anything with numbers and obligations in it, a paraphrase is a liability, because your agent will quote it to a customer as your position.

So the rule is about what the content is, not how messy it is:

Import it when the content is descriptive and the source is messy — product pages, specifications, marketing copy, FAQs scraped from a site.
Paste it when the exact words matter — policies, terms, anything with a number somebody could hold you to. Clean it by hand if you must; it is one document and it is the one worth an hour.
Import, then read it, if you are unsure. The document viewer shows what was stored. Comparing that against your source takes a minute and is the only way to know what your agent will actually say.

What breaks

The screen still talks about a workflow. It is not one any more. AI Import used to be an ordinary Expedify workflow, and the modal still carries the furniture from then — a “Configure Workflow” link that opens the automation page, and a failure state telling you to create a workflow with a Click Trigger on a particular button id. Today the import runs as a built-in skill and none of that applies. If you followed either instruction you would build a workflow that never fires and change nothing about how importing behaves.

Garbage in is still garbage, just tidier. The model cleans structure, not facts. If the scrape returned a cookie banner and a footer because the page renders in the browser, the import produces beautifully formatted markdown of a cookie banner. Read the result whenever the source was a website — that is where this fails most often, and it fails looking like success.

Tables are the reason to use it and the thing to check first. A price table in a PDF is precisely what plain parsing destroys and what this is for. It is also where a model is most likely to drop a row or misalign a column. If the table matters, count the rows in the stored document against the source before you trust it.

Importing twice makes two documents. There is no matching on source. Re-importing an updated page adds a second copy, and your Knowledge Base now contains both the old prices and the new ones, with retrieval free to prefer either. Delete the old document as part of updating, every time.

Try it

Import a page from your own website and open the stored document. Compare it with the page — the difference between the two is what AI Import does.
Now paste the same page's text in as a plain document and compare the two chunk counts and the two bodies. That comparison is the whole lesson.
Import a PDF with a table in it, then count the rows in what got stored.
Import the same source a second time and watch a duplicate appear. Then delete one, and make that your habit.
Finally take a policy document — real words that matter — and read the imported version closely. If you would not send it to a customer as written, that is the one to paste by hand.

Next: the Knowledge Base node — putting the library you have built in front of a workflow, so an agent can actually search it.