Chat over a Notion knowledge base kept fresh in Supabase

Chat over a Notion knowledge base kept fresh in Supabase

An n8n workflow that keeps a Supabase table in step with a Notion database. Every minute it looks for pages edited in the last
minute, deletes their old chunks, reads each page again, cuts it into chunks and embeds them with OpenAI. A chat window
answers questions from the stored chunks.

What is inside

• workflow.json: the workflow as n8n exports it, 22 nodes. Three are sticky notes and one, a Notion trigger, starts switched
  off. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The original used gpt-4o, which retires on 14 January 2027. The chat model is now gpt-6-luna. The chain uses no tools,
  so that model fits.
• The embedding step now names text-embedding-3-small itself. The older version of the node picked an older model when none
  was named.
• The step that loads the page text now reads the joined text. Before, it embedded the whole item written out as JSON.
• The chat trigger was open to anyone with the link. It is now set to not public, so the chat runs inside n8n.
• The Notion database was private. It is now the placeholder <YOUR_NOTION_DATABASE_ID>.
• Several nodes were renamed to say what they do, and the names no longer hold spaced hyphens. The wiring is unchanged.
• Three sticky notes were added in plain words.

Before you start

• A Notion integration token and a database that is shared with that integration.
• An OpenAI API key.
• A Supabase project, its host and its secret key, and a table named documents with content, metadata and a vector column
  of 1536 dimensions. Follow Supabase's guide to vector columns and n8n's notes on its Supabase vector store. Mindket did not
  paste SQL for it and did not check a table for this workflow.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add a Notion credential and select it in Get Edited Pages and Get Page Blocks. In Get Edited Pages, replace
   <YOUR_NOTION_DATABASE_ID> with the id of your database, the 32 characters in its link.
3. Make the table in Supabase. Add a Supabase credential and select it in Delete Old Chunks, Insert Page Chunks and
   Search Page Chunks.
4. Add an OpenAI credential and select it in OpenAI Embeddings and OpenAI Chat Model.
5. Publish the workflow, edit a page in the database and wait two minutes. Then open the chat window and ask about that page.

How a run goes

1. Check Every Minute starts the run, and Get Edited Pages lists the pages whose last edited time equals one minute ago.
2. Input Reference hands the pages to Loop Over Items, which takes one page at a time.
3. Delete Old Chunks removes the stored chunks whose metadata holds the page id. It always outputs an item, so a new page
   with no old chunks goes on. Keep One Delete Result makes sure only one item continues.
4. Get Page Blocks reads every block of the page, nested blocks included. Join Blocks Into One Text joins their text into
   the field concatenated_content.
5. Load Page and Add Metadata wraps the text as a document with the page id and name. Token Splitter cuts it into chunks
   of 500 tokens, OpenAI Embeddings turns them into vectors and Insert Page Chunks stores them. The loop then takes the
   next page.
6. In the chat, Chat Message Received passes a question to Question and Answer Chain. Search Page Chunks and
   Vector Store Retriever find the closest chunks, and OpenAI Chat Model answers from them.

Change the result

• The database: Get Edited Pages.
• How often pages are checked: Check Every Minute. The switched-off Notion Trigger does the same job and can replace it.
• The chunk size: Token Splitter.
• The metadata: Load Page and Add Metadata. Keep the id, because the delete step uses it.
• The model that answers: OpenAI Chat Model.
• The table: Delete Old Chunks, Insert Page Chunks and Search Page Chunks.

Good to know

• Pages that are deleted in Notion are not removed from the table.
• The edited page filter compares the last edited time with the exact time one minute ago. Mindket did not check that against
  Notion's timestamps. A page that the filter does not catch in its minute is not picked up later, so test with a few edits. If pages are missed, change
  the filter or use the switched-off Notion Trigger.
• The chat trigger is set to not public. Making it public with no authentication would let anyone spend your OpenAI credit.
• The vector column must match the model, and the metadata column must be able to hold the page id.
• The Supabase secret key skips row level security. Keep it in n8n and never in a page or a browser.
• The page text goes to OpenAI and to your Supabase project, and the questions go to OpenAI.
• Mindket did not run this workflow against Notion, OpenAI or Supabase.