Use case 08 · AI context & RAG
Feed your AI clean context, not HTML soup.
Language models answer as well as the context you give them. Paste in a scraped web page and half the tokens go to menus, cookie banners and share buttons. Give them Markdown instead and the answers get sharper, cheaper and easier to cite.
Free account, 10 markdown + 10 AI credits included. No credit card needed.
Today
- Copy-paste from the browser drags in navigation, footers and ads.
- PDF text arrives as one long paragraph with broken line breaks.
- Every prompt starts with “ignore the menu and the cookie notice”.
With output.md
- Only the article, with headings, lists and tables intact.
- Structured text a model can chunk on headings instead of arbitrary character counts.
- Fewer tokens per source, more sources per context window.
Step by step
How it works, in 4 steps.
-
01
Bring in the sources
You do
Paste a URL, drop in a list of URLs for bulk mode, add a YouTube link for the transcript, or upload the PDF or Word file.
You get
One clean Markdown file per source, inside a project you name.
TipUse a project per topic or per assistant so exports stay tidy.
-
02
Trim what the model should not see
You do
Open the file in the editor. Cut the author bio, the “related reading” list, the legal boilerplate.
You get
A source that says exactly what you want the model to know — nothing else.
-
03
Group them into a Collection
You do
Click Collect on each article, or upload Markdown you already have. Give the collection a name.
You get
A knowledge base you can generate from or ask questions of.
-
04
Ask, or export
You do
Use Knowledge chat to ask questions grounded in the collection, or export the project as a ZIP and load the files into your own vector store or custom GPT.
You get
Answers that quote your documents, or a folder of .md files ready for any pipeline.
TipMarkdown headings make excellent chunk boundaries — most RAG frameworks split on them out of the box.
Why it works
What you get out of it.
Better answers per token
Removing chrome typically shrinks a page by 80–95%. That is room for more sources, or a cheaper call.
Structure the model can follow
Headings stay headings and tables stay tables, so the model sees the same hierarchy a reader would.
Portable by design
Plain .md files work with LangChain, LlamaIndex, custom GPT knowledge files, Notion AI — anything that reads text.
What it costs
1 markdown credit per source, 1 AI credit per question.
A free account includes 10 markdown and 10 AI credits — enough to build a first collection of ten sources and ask ten questions before you spend anything.
Packs are one-time top-ups — no subscription. See all packs.
Questions
Before you start.
Why is Markdown better than HTML for a language model?
HTML carries markup the model has to read past: tags, classes, scripts, navigation. Markdown keeps only the semantic structure — headings, emphasis, lists, tables, links — so a higher share of every token is meaning.
Can I use the files in my own RAG pipeline?
Yes. Export a project as a ZIP and you get plain .md files with no lock-in. They drop straight into LangChain, LlamaIndex, a vector database or a custom GPT knowledge base.
Does Knowledge chat make things up?
It answers from the documents in your collection and tells you when the documents do not cover a question. It is designed for grounded answers, not general chat.
Try it on your own material.
Paste a link or upload a file, and see the Markdown before you decide anything else.