PDF chat that cites its sources from Pinecone

PDF chat that cites its sources from Pinecone

An n8n workflow that answers questions about a PDF and names where each answer came from. One run loads a PDF from Google Drive
into Pinecone. In the chat, each reply ends with a citation for every piece the model used.

What is inside

• workflow.json: the workflow as n8n exports it, 22 nodes, four of them sticky notes. It holds no credential. Of the notes,
  two are coloured frames that group the nodes, one holds try-it steps and one is a short hint.
• README.md: these notes.

What we changed from the original copy

• The Google Drive, OpenAI and Pinecone nodes were linked to credentials in their author's n8n. The links are gone, so select your
  own.
• The Drive link in Set file URL in Google Drive pointed at one public file and had a space in front of it. It is now a
  placeholder, <GOOGLE_DRIVE_FILE_ID>, that you replace with your own link. The try-it note and the small note beside the node
  were reworded to match.
• Compose citations read [$json.citations], which asks for the entry at a position named by the whole list of numbers. With one
  cited chunk it worked, and with two or more it found nothing and the run failed. It now reads the entry for each number in turn,
  and it skips a number that has no chunk.
• The Structured Output Parser's schema now lists answer and citations as required, so a reply without citations fails at the
  parser, not later.
• OpenAI Chat Model had no model set, so it used n8n's default, gpt-5-mini. It is now set to gpt-5.6-terra.
• Both embeddings nodes had no model set, so they used text-embedding-ada-002. They are now set to text-embedding-3-small, which
  makes vectors of the same size, 1536.
• The Chat Trigger's webhook id is gone, and n8n makes a new one when you import the file.
• Nothing else.

Before you start

• An n8n instance, on n8n Cloud or on your own server.
• An OpenAI API key.
• A Google account that can open your PDF, linked to n8n through its Google sign-in.
• A Pinecone account, an API key, and an index with 1536 dimensions.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add an OpenAI credential. Select it in OpenAI Chat Model, Embeddings OpenAI and Embeddings OpenAI2.
3. Add a Google Drive OAuth2 credential and select it in Download file.
4. In Set file URL in Google Drive, replace the whole value with the share link of your PDF. In Drive, open the PDF, choose Share
   and copy the link.
5. Make an index with 1536 dimensions in Pinecone. Add a Pinecone credential in n8n, then select the credential and the index in
   Add to Pinecone vector store and in Get top chunks matching query. Both nodes must use the same index.
6. Press Execute workflow to load the PDF. Check in Pinecone that the index now holds records.
7. Open the chat from the button at the bottom of the canvas and ask a question about the PDF.

How a run goes

Loading, started by hand:

1. Set file URL in Google Drive holds the link in a field named file_url, and Download file fetches that file.
2. Add in metadata copies the file name, its extension and the link onto the item.
3. Add to Pinecone vector store takes the file through Default Data Loader as binary data. The loader keeps the file name and
   the link as metadata. Recursive Character Text Splitter cuts the text into pieces of 3,000 characters that overlap by 200.
   Embeddings OpenAI turns each piece into a vector.

Asking, started from the chat:

1. Set max chunks to send to model adds a field chunks with the value 4.
2. Get top chunks matching query embeds your question with Embeddings OpenAI2 and asks Pinecone for that many pieces.
3. Prepare chunks numbers the pieces from 0 and joins them into one block of text, headed --- CHUNK 0 --- and so on.
4. Answer the query based on chunks gives the block and your question to OpenAI Chat Model. The Structured Output Parser makes the
   model reply with an answer and a list of the chunk numbers it used.
5. Compose citations turns each number into the file name and the line range of that piece, such as [report.pdf, lines 12-30].
6. Generate response joins the answer and the citations into the text that comes back in the chat.

Change the look

• Which PDF it loads: the link in Set file URL in Google Drive.
• How many pieces the model sees: the chunks value in Set max chunks to send to model.
• How big a piece is: the chunk size and the overlap in Recursive Character Text Splitter. Load again after a change.
• The wording of the instructions: the prompt text in Answer the query based on chunks.
• The look of a citation: the expression in Compose citations.
• The model: the model field of OpenAI Chat Model.

Good to know

• Copies: the insert node does not clear anything. Each time you load, the pieces are added again, and a question may then return
  the same text twice. Empty the index, or use a namespace, before you load again.
• Models: OpenAI's deprecations page, read on 12 November 2026, lists gpt-5-mini for shutdown on 11 December 2026 with
  gpt-5.6-terra as the replacement. The chain has no tools, so Mindket's copy uses gpt-5.6-terra. The page does not list
  text-embedding-ada-002. The new embeddings model gives vectors that do not mix with the old one's, so load again if your index
  already holds vectors from the old model.
• Citations come from the model: it reports the numbers of the chunks it says it used, and nothing checks that. Read the cited
  lines before you rely on a quote.
• Line numbers: they are the ones the text splitter records for each piece, so they count inside the text it was given.
• Your data: the text of the PDF goes to OpenAI and to Pinecone, and each question goes to OpenAI.
• Not run: nothing here was run against Google, OpenAI or Pinecone. These notes come from n8n's source and the vendors' pages, read
  on 12 November 2026.