Notion pages stored in Pinecone with Gemini embeddings

Notion pages stored in Pinecone with Gemini embeddings

An n8n workflow that watches a Notion database for new pages. It reads each page, joins the text, cuts it into chunks and
turns the chunks into vectors with a Gemini embedding model. Pinecone stores the vectors, with the page id, the title and
the creation time as metadata.

What is inside

• workflow.json: the workflow as n8n exports it, nine nodes. One is a sticky note. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The original used the Gemini model models/text-embedding-004, which Google shut down on 14 January 2026. The step now uses
  models/gemini-embedding-001, the default of n8n's Gemini embeddings node.
• That model makes vectors of 3072 numbers, not the 768 in the original name. The Pinecone index must have that size.
• The Notion database and the Pinecone index were private. They are now the placeholders <YOUR_NOTION_DATABASE_ID> and
  <YOUR_PINECONE_INDEX>.
• The page title was read from a property called Name. It is now read from whichever property holds the title, so it works
  when your database calls it something else.
• Several nodes were renamed to say what they do, and the names no longer hold spaced hyphens. The wiring is unchanged.
• A sticky note was added in plain words.

Before you start

• A Notion integration token and a database that is shared with that integration.
• A Google Gemini API key.
• A Pinecone account, an API key and an index with 3072 dimensions.

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 Page Added in Notion and Get Page Blocks. In Page Added in Notion, replace
   <YOUR_NOTION_DATABASE_ID> with the id of your database, the 32 characters in its link.
3. Make a Pinecone index with 3072 dimensions. In Store in Pinecone, add a Pinecone credential and replace
   <YOUR_PINECONE_INDEX> with the index name.
4. Add a Google Gemini credential and select it in Gemini Embeddings.
5. Publish the workflow, then add a test page to the database and wait a minute.

How a run goes

1. Page Added in Notion finds a new page. Simplify is off, so the item holds the raw page with its properties.
2. Get Page Blocks reads the blocks that sit directly on the page.
3. Filter Non-Text Content drops blocks of the type image or video.
4. Join Blocks Into One Text joins the text of the remaining blocks, one block to a line, into the field
   concatenated_content.
5. Load Page and Add Metadata wraps that text as a document with the page id, the creation time and the page title.
   Token Splitter cuts it into chunks of 256 tokens with an overlap of 30.
6. Gemini Embeddings turns each chunk into a vector, and Store in Pinecone inserts the vectors into your index.

Change the result

• The database: Page Added in Notion.
• Which blocks are skipped: the conditions in Filter Non-Text Content.
• The chunk size and the overlap: Token Splitter.
• The metadata: Load Page and Add Metadata.
• The embedding model: Gemini Embeddings. Change the size of the index with it.
• The index or the namespace: Store in Pinecone.

Good to know

• This workflow only fills the index. Searching it is a separate workflow.
• n8n's Gemini embeddings step does not ask for a smaller vector size. Check the index size against the length of the first
  vector. Mindket did not check which distance measure Google recommends for the model.
• Only new pages are handled. A page that you edit or delete later stays as it was in the index, and a page that is run twice
  is stored twice.
• Only the blocks directly on the page are read. A toggle or a nested list does not add its inner text.
• The page text goes to Google's embedding service and to Pinecone.
• Mindket did not run this workflow against Notion, Gemini or Pinecone. The model name, the vector size and the input limit
  follow Google's published notes, as read by Mindket on 12 November 2026.