Three AI examples: email labels, PDF chat and booking agent

Three AI examples: email labels, PDF chat and booking agent

An n8n workflow that holds three separate AI examples on one canvas, plus a small webhook demo. One labels incoming Gmail
messages with a text classifier. One stores a PDF in Pinecone and answers questions about it in a chat. One is a Claude
agent that checks a Google Calendar for free time and books a 30-minute appointment. Each has its own trigger and its own
accounts, so you can set up one and ignore the rest.

What is inside

• workflow.json: the workflow as n8n exports it, 27 nodes and 6 sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The original copy carried seven sticky notes made for a live talk. They held slide and screenshot pictures hosted on
  other sites, a QR code, the speaker's name and links to a calendar, n8n's template library and its forum. Mindket's
  copy leaves them out and keeps the six notes that name each example.
• Several ids and addresses belonged to the author's own accounts. The Slack channel in Send message is
  <YOUR_SLACK_CHANNEL_ID>. The two Gmail label ids read <YOUR_AUTOMATION_LABEL_ID> and <YOUR_MUSIC_LABEL_ID>. The
  calendar address, which sat in the availability request, the booking address and the booking body, reads
  <YOUR_CALENDAR_EMAIL>.
• The booking agent's prompt, its chat title and its greeting named a person. They now say "the calendar's owner" or
  nothing at all. The Gmail trigger was named after an inbox owner and is now On new email.
• PDFs to download was a No-Op with nothing in it, so Download PDF had no address to fetch. It is now an Edit Fields
  node with one item: file_url, whitepaper_title, publish_year and author for the Bitcoin white paper. The
  address returned a PDF when Mindket tried it.
• The chat trigger carried the booking agent's title and greeting, but its only line led to the PDF chain, and the
  booking agent had no input at all. Mindket's copy gives the trigger to the booking agent as
  Chat with the booking agent and adds a plain Chat with the PDF for the PDF chain.
• Anthropic Chat Model was on version 1.2 with no model set. n8n's source gives that version
  claude-3-5-sonnet-20240620. Anthropic's deprecations page lists it as retired on 28 October 2025, with
  claude-sonnet-5-5 as the replacement. The node is now on version 1.6 and set to claude-sonnet-5-5. Both pages were
  read on 12 November 2026.

In the two calendar tools, the capital O in the name of the Google Calendar OAuth2 credential type is written as the JSON
escape \u004f. n8n reads the same name. The escape only stops a store check from taking that long name for a secret
key.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• An OpenAI API key, for the classifier, the PDF chat and the embeddings.
• For the labeller: a Gmail account with two labels, and for the webhook demo a Slack workspace.
• For the PDF chat: a Pinecone account and an index named whitepapers made for 1,536-dimension vectors, the size
  n8n's embeddings node lists as its default.
• For the booking agent: an Anthropic API key and a Google account with a calendar to book into.

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 Model1, Embeddings OpenAI, Embeddings OpenAI2 and
   OpenAI Chat Model.
3. For the webhook demo, add a Slack OAuth2 credential, select it in Send message and replace
   <YOUR_SLACK_CHANNEL_ID> with the channel's ID.
4. For the labeller, add a Gmail OAuth2 credential and select it in On new email, Add automation label and
   Add music label. Replace the two label ids with your own, then switch On new email on: it is disabled in the file.
5. For the PDF chat, add a Pinecone credential and select it in Insert into Pinecone vector store and
   Read Pinecone Vector Store. Create the whitepapers index first.
6. For the booking agent, add an Anthropic credential and select it in Anthropic Chat Model. Add a Google Calendar
   OAuth2 credential and select it in Get calendar availability and Book appointment. Replace <YOUR_CALENDAR_EMAIL>
   in the three places it appears.
7. Save. Run each example from its own trigger, as the next part describes.

How a run goes

Webhook demo:

1. Webhook receives a request. Whether email contains n8n checks whether its email query value contains @n8n.
2. When it does, Execute JavaScript adds a myNewField of 1 to the item, and Send message posts the address to Slack.

Email labeller:

1. On new email polls Gmail every minute for new messages.
2. Assign label with AI asks a gpt-4o model to pick one of two categories from the message text: automation for mail
   about workflows and automated processes, music for mail about music.
3. Add automation label or Add music label puts the matching label on the message. A message that fits neither is
   dropped.

PDF chat:

1. PDFs to download holds one item, and Download PDF fetches its file_url.
2. Insert into Pinecone vector store clears the whitepaper namespace and stores the file. Default Data Loader reads
   the PDF with its title, year and author as metadata, Recursive Character Text Splitter cuts it into chunks of 3,000
   characters with 200 of overlap, and Embeddings OpenAI makes the vectors.
3. Chat with the PDF takes a question. Question and Answer Chain adds the instruction to answer only from the stored
   passages, and Vector Store Retriever fetches them from Read Pinecone Vector Store. A gpt-4o model writes the
   answer.

Booking agent:

1. Chat with the booking agent opens a public chat titled Book an appointment.
2. Appointment booking agent runs on Anthropic Chat Model with Window Buffer Memory, which keeps the last 10
   messages. Its prompt makes appointments 30 minutes long and allows a booking only into a free slot.
3. Get calendar availability posts a free and busy request to Google Calendar. Book appointment creates the event, titled
   with the visitor's name, and adds the visitor and the calendar's owner as attendees.

Change the look

• The classifier's categories: the names and descriptions in Assign label with AI.
• The labels: the ids in the two Gmail nodes.
• What the PDF chat answers from: the item in PDFs to download, and the namespace and index in the Pinecone nodes.
• How the PDF chat answers: the text in Question and Answer Chain.
• The booking rules: the system message in Appointment booking agent, and timeZone in the bodies of both tools.
• The models: OpenAI Chat Model, OpenAI Chat Model1 and Anthropic Chat Model.

Good to know

• The model and agent facts below come from pages and source files read on 12 November 2026.
• The two OpenAI chat nodes set gpt-4o. OpenAI's deprecations page gives 14 January 2027 as the shutdown date of
  gpt-4o-2024-08-06 and gpt-4o-2024-11-20, with gpt-6-luna as the replacement. It does not list
  text-embedding-ada-002, the embeddings node's default for this version.
• Insert into Pinecone vector store clears its namespace before it stores a file, so each run replaces the last PDF.
• Assign label with AI has one category for each message and, with no fallback set, drops a message that fits neither.
• The booking chat is public. Once the workflow is active, anyone with the chat's address can book into your calendar and
  run up your Anthropic bill.
• Both calendar tools write Europe/Berlin into their requests, and the free and busy tool's description gives its example
  in 2024. Change both to your own time zone.
• Window Buffer Memory lives inside the n8n instance. n8n's source warns that it does not work in queue mode or with
  several main instances.
• n8n's source says the Anthropic node ignores its temperature on Claude Sonnet 5 and later, so the 0.4 in the file has
  no effect.
• The webhook demo posts whatever its email value holds to a Slack channel. Its path is a fixed string in the file, so
  change it before you switch the workflow on.
• Billing: OpenAI bills the classifier, the embeddings and the PDF chat. Anthropic bills each booking chat turn.
• Email text goes to OpenAI. The PDF's chunks go to OpenAI and Pinecone, the booking chat to Anthropic, and calendar
  requests to Google.