Personal shopper chatbot for WooCommerce with store RAG

Personal shopper chatbot for WooCommerce with store RAG

An n8n workflow for the chat of a WooCommerce shop. A shopper either asks for a product or asks about the shop itself.
Product questions go to a search of the shop's catalogue. Questions about opening hours or the address are answered from
the shop's own documents, which sit in a Qdrant collection.

What is inside

• workflow.json: the workflow as n8n exports it, twenty-five nodes: twenty-one that do the work and four sticky notes.
  It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The credential links are gone: OpenAI on five nodes, Google Drive on two, Qdrant on two, WooCommerce on the product
  search and a Header Auth credential on the delete call. Select your own.
• The original named one shop's Google Drive folder and one Qdrant collection, and its delete call went to
  https://QDRANTURL/collections/NAME/points/delete. They are YOUR_FOLDER_ID, YOUR_COLLECTION and
  https://YOUR_QDRANT_HOST/collections/YOUR_COLLECTION/points/delete now.
• The RAG tool had an Italian name and description. It is now called RAG, the name the agent's instructions use, with
  an English description.
• The personal_shopper tool read its filters from Information Extractor fields named SKU, keyword, price_max
  and price_min. The extractor's schema has no fields with those names, so those four options came out empty. They now
  use $fromAI, which lets the agent fill them from what the shopper said.
• The extractor's schema was an example object, not a JSON Schema. It is now a full JSON Schema with the same two fields.
• The two notes that describe step 1 are reworded. The other notes stay as they were.
• Nothing else.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• A WooCommerce shop and REST API keys for it, for the WooCommerce credential.
• A Qdrant instance and an API key, with a collection already created. The vector size must match the embedding model.
  OpenAI's embeddings guide gives 1536 for text-embedding-3-small.
• A Google Drive folder with the shop's documents, such as opening hours, delivery terms and returns. Google Docs and plain
  text files work. PDFs do not, because the download step asks for plain text.
• 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. Add the credentials and select them: OpenAI in OpenAI Chat Model, OpenAI Chat Model1, OpenAI Chat Model2,
   Embeddings OpenAI and Embeddings OpenAI3, Google Drive in Google Drive2 and Google Drive1, Qdrant in both
   Qdrant nodes, WooCommerce in personal_shopper, and a Header Auth credential in HTTP Request.
3. For the Header Auth credential, set the name to api-key and the value to your Qdrant API key.
4. Put your collection's name where YOUR_COLLECTION stands, in Qdrant Vector Store, Qdrant Vector Store1 and the
   HTTP Request address, and your Qdrant address where YOUR_QDRANT_HOST stands.
5. In Google Drive2, replace YOUR_FOLDER_ID with the ID of your documents folder.
6. Click Test workflow to load the documents into Qdrant. Do this again whenever the documents change.
7. Open the chat from When chat message received. Ask for a product, then ask when the shop opens.

How a run goes

Loading the documents, started by hand:

1. HTTP Request sends Qdrant a delete request with an empty filter, meant to clear the collection first.
2. Google Drive2 lists the files in the folder and Google Drive1 downloads each one as plain text.
3. Default Data Loader2 turns each file into a document, and Token Splitter1 cuts it into pieces of 300 tokens with 30
   tokens of overlap. Embeddings OpenAI3 embeds them and Qdrant Vector Store1 stores them.

Answering a shopper:

1. When chat message received takes the message, and Edit Fields keeps its session ID and text.
2. Information Extractor asks a chat model whether the shopper wants a product. If so, it pulls out the keyword, any
   prices, a product code and a category.
3. AI Agent reads the message and the extractor's JSON. For a product it calls personal_shopper, which searches the
   shop's in stock products. For a shop question it calls RAG, which reads the Qdrant collection through
   OpenAI Chat Model1. Calculator is there for sums, and Window Buffer Memory keeps the chat's earlier messages.

Change the look

• Which questions go where: the system messages of Information Extractor and AI Agent.
• The shop's documents: the folder in Google Drive2.
• How the documents are cut: the size and overlap in Token Splitter1.
• The models: the three chat model nodes and the two embeddings nodes. Both embeddings nodes must use the same model.
• What a product search returns: the options of personal_shopper.

Good to know

• The prompts were written for a shop that sells shoes and bags. They say so in the extractor and call the shop a clothing
  store in the agent. Change both to describe yours.
• personal_shopper returns the full records of up to 50 in stock products, which is a lot of text for the model. Lower its
  Limit if answers get slow or costly.
• The extractor reads a category name, but personal_shopper has no category option, so a category does not narrow the
  search.
• Qdrant's reference page, read on 12 November 2026, does not say what an empty filter matches in a delete request. Try the
  delete call on a throwaway collection first. It is meant to clear the collection, and it may do so.
• None of the three chat model nodes names a model. n8n's current source, read on 12 November 2026, gives
  gpt-5-mini as the default. OpenAI's deprecations page gives 11 December 2026 for it and names gpt-5.6-terra.
• The two embeddings nodes name no model either. The default for their version in n8n's current source is
  text-embedding-3-small. Load and search must use the same model.
• Shopper messages and the shop's documents go to OpenAI.
• The JSON holds no credential.