Post

Notion's AI can answer questions on any document you've created. But I have 1,000s of documents in my Notion—you can't put all these pages inside an AI prompt. How did they make it work? It's more interesting than you think ↓

Every time a document is created, you can use a tool like @cohere.com to create embeddings, which helps your system understand the document's meaning and context, and not just its keywords. Whole chunks of text are given an embedding, so that their context can be fully understood.

To save embeddings, you should use a vector database like @turbopuffer.bsky.social‬, which is specially optimised for handling numerical data of this type. When its queried, search text is converted into an embedding, passed to the DB, and semantically related data is scored & returned.

Now querying is set up, we need to show AI how to use it. This is possible through "tool calling", which is a way to let AI call your functions. First, define which "tools" the AI can use, and how to use them. For example, this is a `search` tool for our vector database.