Chat with the PDF and text files in a Supabase bucket

Chat with the PDF and text files in a Supabase bucket

An n8n workflow in two halves. The first reads the PDF and text files in a Supabase storage bucket, cuts their text into
pieces and stores each piece as an OpenAI embedding in a Supabase table. The second is a chat agent that searches those
pieces and answers from them.

What is inside

• workflow.json: the workflow as n8n exports it, 31 nodes, nine of them sticky notes. It holds no credential. Six notes
  remind you to replace credentials, two name the halves, Scenario 1 for the files and Scenario 2 for the chat, and one
  gives the workflow's title.
• README.md: these notes.

What we changed from the original copy

• Every node was linked to a credential in its author's n8n. The links are gone, so select your own.
• Get All files and Download held the address of their author's Supabase project. It now reads
  <YOUR_SUPABASE_PROJECT_REF>.supabase.co.
• Supabase Vector Store filtered on one file's id from its author's database. It now reads <YOUR_FILE_ID>.
• Three sticky notes carried their author's promotion, a video link and setup steps. One now holds a short title, and the
  other two are gone.
• The original copy had no settings block. Mindket's sets the execution order to v1.
• Nothing else. The file has no Code node, and its expressions needed no repair.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• A Supabase project with a storage bucket named private, holding PDFs or plain text files.
• A table files with the text columns name and storage_id.
• A table documents and a match_documents function, set up the way n8n's Supabase vector store expects. The table has
  content, metadata and embedding columns, and the embedding is 1536 numbers long, the size text-embedding-3-small
  returns.
• An OpenAI API key.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Create the bucket, the files table and the documents table with its function in Supabase.
3. Add a Supabase API credential with your project's URL and a secret key. Select it in Get All Files,
   Create File record2, Insert into Supabase Vectorstore and Supabase Vector Store. In Get All files and
   Download, which are HTTP requests, pick the Supabase credential type and the same credential.
4. In Get All files and Download, replace <YOUR_SUPABASE_PROJECT_REF> in the address with your project reference,
   the part before .supabase.co. If your bucket has another name, change private in both addresses.
5. Add an OpenAI credential and select it in Embeddings OpenAI, Embeddings OpenAI2, OpenAI Chat Model1 and
   OpenAI Chat Model2.
6. Put one PDF and one text file in the bucket. Run the workflow from When clicking ‘Test workflow’, then check that
   files and documents gained rows.
7. In Supabase Vector Store, replace <YOUR_FILE_ID> with the id of a row in files to answer from that one file, or
   remove the metadata filter to search every file.
8. Save, press Chat on the canvas and ask a question about a file.

How a run goes

Scenario 1, the files:

1. Get All Files reads the files table and Aggregate folds its rows into one list.
2. Get All files asks Supabase Storage for the bucket's file list, 100 files at most, sorted by name.
3. Loop Over Items takes one file at a time. If lets it through when no row in files has its storage_id and its
   name is not .emptyFolderPlaceholder. A file that fails goes back to the loop.
4. Download fetches the file. Switch sends a download with no file extension in its binary data, a text file, down
   the txt output, and a PDF down the pdf output through Extract Document PDF. Merge joins the two.
5. Create File record2 adds the file to files. Insert into Supabase Vectorstore then stores its text. `Default Data
   Loader reads data for a text file or text for a PDF and stamps each piece with the row's id as file_id`.
   Recursive Character Text Splitter cuts the text into pieces of 500 characters that overlap by 200, and
   Embeddings OpenAI turns each into a vector. The loop then moves on.

Scenario 2, the chat:

1. When chat message received takes your message. AI Agent answers it, with OpenAI Chat Model1 as its model.
2. Its one tool, Vector Store Tool1, is named knowledge_base. It searches Supabase Vector Store through
   Embeddings OpenAI2 and returns the 8 closest pieces, and OpenAI Chat Model2 reads them for the tool.

Change the look

• Which files are read: the two rules in Switch.
• Piece size: chunkSize and chunkOverlap in Recursive Character Text Splitter.
• Pieces per question: topK in Vector Store Tool1.
• The embedding model: text-embedding-3-small in both embedding nodes. Keep them the same, and keep the table's vector
  size to match.
• The chat: the models in OpenAI Chat Model1 and OpenAI Chat Model2, and the tool's name and description.

Good to know

• Models: the two chat model nodes set no model, so they take the node's own default, gpt-5-mini in n8n's repository on
  12 November 2026. OpenAI's deprecations page, read that day, gives 14 January 2027 as the shutdown of gpt-4o and
  gpt-4o-mini, with gpt-6-luna as the replacement. If your n8n offers an older default, pick a model in each node.
• File types: only PDFs and plain text files are meant to pass. Another type, such as a Word file, matches no rule of
  Switch, so nothing returns to the loop and the run stops with later files unread.
• Limits: Get All files sees the first 100 files in the bucket. Get All Files has Return All off, so n8n reads only 50
  rows of files. Past 50 rows, older files look new and are added again. Turn on Return All in that node.
• Each file goes in once. To read a file again, delete its row in files and its pieces in documents first.
• The metadata filter keeps the chat to one file. Without it the agent searches every file in documents.
• Billing: OpenAI bills the embeddings for every piece of every new file, and each question takes an agent call plus a call
  for the knowledge tool.
• Your data: the text of each file goes to OpenAI and into your Supabase project. Your questions go to OpenAI.
• The Supabase credential holds a secret key, so use a project meant for this.