PDF report questions answered from Qdrant through a webhook

PDF report questions answered from Qdrant through a webhook

An n8n workflow in two parts. The first reads a PDF from Google Drive, cuts it into chunks and stores them in a Qdrant
collection. The second waits at a webhook: you post a question, it finds the five closest chunks, hands them to an
OpenAI chat model and sends the answer back in the reply.

What is inside

• workflow.json: the workflow as n8n exports it, 13 nodes beside 2 sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

Mindket fixed a few things the copy would have tripped over:

• The Google Drive node held its author's own file id. It now reads YOUR_FILE_ID, and you put in your own.
• The webhook's path was a long random string. It now reads pdf-question.
• The copy had a chat trigger wired into the answering chain. The trigger sends the message in a field called input,
  and the chain reads body.input, so a question from the Chat window would have failed. Mindket removed that trigger.
• Two sticky notes spoke of Supabase while the nodes use Qdrant. Mindket reworded them, and removed a third note that
  only pointed to a video.
• The workflow arrives switched off, and no credential is inside.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• An OpenAI API key.
• A Google account, for a Google Drive OAuth2 credential, and a PDF in your Drive.
• A Qdrant instance that n8n can reach, with its address and, if it needs one, an API key.

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 and select it in Embeddings OpenAI, Embeddings OpenAI1 and OpenAI Chat Model.
3. Add a Google Drive credential and select it in Google Drive. Replace YOUR_FILE_ID with the id of your PDF, the
   long part of its address in Drive.
4. Add a Qdrant API credential with your Qdrant address and key, and select it in Qdrant Vector Store and
   Qdrant Vector Store1.
5. Open OpenAI Chat Model and pick a current model. The file names none, so n8n fills in gpt-5-mini. OpenAI's
   deprecations page, read on 12 November 2026, gives 11 December 2026 as its shutdown date and names gpt-5.6-terra
   as its replacement.
6. Press Execute Workflow once. The PDF is stored in a collection called crowd, which is created if it is missing.
7. Listen for a test event on Webhook1, then post a question:

curl -X POST https://YOUR-N8N-ADDRESS/webhook-test/pdf-question \
  -H "Content-Type: application/json" \
  -d '{"input": "What are the main risks this report names?", "company": "crowd"}'

Once the workflow is published, the same call goes to /webhook/pdf-question.

How a run goes

The first part runs when you press Execute Workflow:

1. When clicking "Execute Workflow" starts it.
2. Google Drive downloads the file whose id you set.
3. Binary to Document reads the PDF's binary field data and turns its pages into documents.
4. Recursive Character Text Splitter cuts them into chunks of 3000 characters that overlap by 200.
5. Embeddings OpenAI turns each chunk into a vector, and Qdrant Vector Store inserts the vectors in the collection
   crowd.

The second part runs when a question arrives:

1. Webhook1 takes a POST on pdf-question and holds the reply until a node answers.
2. Retrieval QA Chain reads the question from body.input.
3. Qdrant Vector Store1 searches the collection named in body.company, with Embeddings OpenAI1 turning the
   question into a vector. Vector Store Retriever keeps the five closest chunks.
4. OpenAI Chat Model writes the answer from those chunks.
5. Respond to Webhook sends the answer back as plain text.

Change the look

• The collection: the name in Qdrant Vector Store for storing, and the company field you post for asking. Keep them
  the same.
• Chunk size and overlap: Recursive Character Text Splitter.
• How many chunks the chain sees: topK in Vector Store Retriever, 5 in the file.
• The model: OpenAI Chat Model.
• A different file: the id in Google Drive. Use a new collection name for each report you want to keep apart.

Good to know

• The company field decides which collection is searched, so a name that matches no collection cannot return an answer.
• Each run of the first part adds the file's chunks again and does not clear the collection, so run it once per file.
• Both embedding nodes must use the same model, because the collection's vectors have one fixed size. They arrive on the
  node's default.
• Each question costs at least one embedding call and one model call.
• The report's text and each question go to OpenAI. Qdrant holds the chunks.
• The webhook has no sign-in. Once the workflow is published, anyone with its address can ask questions and spend your
  OpenAI key.
• The answer rests on the five chunks it was given, so check anything that matters against the PDF.
• The JSON holds no credential.