Voice RAG chatbot for an ElevenLabs agent

Voice RAG chatbot for an ElevenLabs agent

An n8n workflow that answers a voice agent's questions from your own documents. One part loads the files of a Google
Drive folder into a Qdrant collection. The other is a webhook. An ElevenLabs agent posts the caller's question to it, an
OpenAI agent looks the answer up in Qdrant, and ElevenLabs speaks what comes back.

What is inside

• workflow.json: the workflow as n8n exports it, 18 nodes and five sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• Create collection sent a POST to /collections/<name> with a body of {"filter": {}}, which is a points filter.
  Qdrant creates a collection with a PUT and a vectors setting. The node now sends a PUT with a vector size of 1536
  and the distance Cosine, and its On Error setting is Continue, so a second run does not stop when the collection
  already exists.
• Window Buffer Memory had no session setting. Its default reads a sessionId field that a webhook call does not
  carry, so the agent would stop with "No session ID found". It now uses a fixed key, voice-rag-chat.
• The address placeholders are now YOUR_QDRANT_URL, YOUR_COLLECTION and YOUR_FOLDER_ID. The original copy wrote
  them as QDRANTURL, COLLECTION and a sample folder name after an equals sign.
• The five sticky notes are rewritten. The widget snippet in the last one uses the script address ElevenLabs shows now.
• The original copy was exported with no credential, as these notes expect, and with the workflow switched off.

Before you start

• An n8n instance with a public address, on n8n Cloud or your own server, so ElevenLabs can reach the webhook.
• An OpenAI API key. It covers the two chat models and the embeddings.
• A Qdrant instance, its address and its API key.
• A Google account, a Drive folder with your documents in it and a Google Drive OAuth2 credential.
• An ElevenLabs account with an agent that can call a webhook tool.

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, OpenAI Chat Model, Embeddings OpenAI and Embeddings OpenAI1.
3. Add a Qdrant credential with your address and key. Select it in Qdrant Vector Store and Qdrant Vector Store1.
   In both, replace YOUR_COLLECTION with your collection's name.
4. Add an HTTP Header Auth credential with the header name api-key and your Qdrant key as its value. Select it in
   Create collection and Refresh collection. In both, replace YOUR_QDRANT_URL with your Qdrant host, with its port
   if it uses one, and YOUR_COLLECTION with the same collection name.
5. Add a Google Drive OAuth2 credential. Select it in Get folder and Download Files. In Get folder, replace
   YOUR_FOLDER_ID with the ID of your folder, the last part of its web address.
6. Press Test workflow once. The collection is made, emptied and filled from the folder. Check in Qdrant that it holds
   points.
7. Save and press Publish. Copy the production address of Listen. It ends in test_voice_message_elevenlabs.
8. In ElevenLabs, create an agent and give it a first message and a system prompt that tells it to call your tool for
   every question. Add a Webhook tool with the method POST and the production address. Add one body parameter named
   question, a string whose value the model fills in from the conversation.
9. Test the agent in ElevenLabs. Then add the widget to your site, as the last sticky note shows.

How a run goes

Loading the documents runs from When clicking ‘Test workflow’:

1. Create collection makes the collection in Qdrant. Refresh collection then posts a delete of every point, with an
   empty filter, so each load starts from nothing.
2. Get folder lists the files directly inside your folder, up to 50. Download Files downloads each one, and turns a
   Google Doc into plain text.
3. Qdrant Vector Store1 stores them. Default Data Loader reads each file's binary data and picks a loader by file
   type. Token Splitter cuts the text into chunks of 300 tokens that overlap by 30. Embeddings OpenAI1 turns each
   chunk into a vector.

Answering runs from the webhook:

1. Listen takes a POST on the path test_voice_message_elevenlabs and waits for Respond to ElevenLabs to answer.
2. AI Agent takes body.question from the request as its message. OpenAI is its chat model, and `Window Buffer
   Memory` keeps the last five exchanges.
3. The agent can call Vector Store Tool, named company. Qdrant Vector Store finds the four closest chunks, using
   Embeddings OpenAI for the question, and OpenAI Chat Model writes an answer from them.
4. Respond to ElevenLabs sends the first item back as JSON, {"output": "..."}, and the ElevenLabs agent speaks it.

Change the look

• What the agent is allowed to say: add a system message to AI Agent, and tell the ElevenLabs agent how to behave in
  its own prompt.
• When the agent searches: the name and description of Vector Store Tool. The description is written in Italian, so
  put yours there, in the language of your documents.
• The chunks: 300 and 30 in Token Splitter.
• How many chunks an answer uses: the Limit of Vector Store Tool, 4 by default.
• The chat models: OpenAI and OpenAI Chat Model. The file sets no model, so n8n fills in its own default.
• The embedding model: Embeddings OpenAI and Embeddings OpenAI1. Keep them the same, and change the vector size in
  Create collection to match.

Good to know

• Neither OpenAI chat node names a model in the file, so n8n fills in gpt-5-mini. OpenAI's deprecations page, read on
  12 November 2026, lists it for shutdown on 11 December 2026, with gpt-5.6-terra as the replacement. Choose another
  model in both nodes before then. The embedding default, text-embedding-3-small, is not on that page.
• Qdrant's API page does not say what an empty filter in a delete matches. Refresh collection expects every point to
  go, so look at the point count after a load.
• The vector size of 1536 matches text-embedding-3-small. A different embedding model needs its own size and a new
  collection.
• Every press of Test workflow empties the collection and embeds every file again, and OpenAI bills the embeddings each
  time.
• Get folder reads only the files directly in the folder, not its subfolders.
• Every caller shares one conversation, because the memory key is fixed. The history lives in the n8n process, and n8n
  drops it after an hour without use.
• The webhook has no authentication. Anyone with its address can ask questions on your OpenAI key. Turn on the webhook's
  authentication and send the same header from the ElevenLabs tool.
• The test address works only while n8n is listening, so the ElevenLabs tool needs the production address.
• ElevenLabs only shows a widget for an agent that is public, with authentication off.
• Your documents go to OpenAI to be embedded and are stored in Qdrant. Each question goes to OpenAI.