Telegram assistant with Pinecone contacts and tool workflows

Telegram assistant with Pinecone contacts and tool workflows

An n8n workflow for a Telegram bot that works through other workflows. Message it in plain words. An agent on gpt-4o
picks one of six tool workflows for email, calendar and a project sheet, finds a contact's email in Pinecone when you
name someone, and answers in the chat.

What is inside

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

The six tool workflows are not in the file. Import or build them first (below).

What we changed from the original copy

• The six tool nodes pointed at workflows in their author's own n8n. They now point at none until you pick yours.
• Pinecone Vector Store named the author's index, and its note listed two contacts, one with a personal email
  address. The index is now YOUR_PINECONE_INDEX, and the note shows a sample contact.

Before you start

• An n8n instance that Telegram can reach over HTTPS, on n8n Cloud or your own server.
• A Telegram bot token from BotFather.
• An OpenAI API key.
• A Pinecone account, its API key, and an index with 1,536 dimensions, the size of text-embedding-3-small. Your
  contacts go in a namespace named Contacts.
• The six tool workflows, saved in the same n8n.

The six tool workflows

Each tool node runs a workflow of yours and hands it the agent's words in query. Each workflow must end on a node
whose output has a response field: that field is what goes back to the agent. These listings on Mindket take query
and return response, one for each node:

| Node | Tool name | What it is asked for | A listing that fits |
|---|---|---|---|
| ⚙️Get Emails | get_emails | Email from a given day | Summarize unread Gmail by date |
| ⚙️Send Email | send_email | An email to send | Write and send Gmail from a plain request |
| ⚙️Get Calendar | get_calendar | A day's events | Google Calendar day summary tool |
| ⚙️Set Calendar Event | set_event | A new event | Plain-text Google Calendar event tool |
| ⚙️Update Database | update_database | A project's notes and status | Log project updates to Google Sheets |
| ⚙️Summarize Database | Summarize_Database | A summary of every project | Google Sheets project summary tool |

Five of them get query as text. ⚙️Send Email has a schema, so its query is an object:

{
  "sendTo": "sam@example.com",
  "subject": "Review moved",
  "message": "Hi Sam, the review moves to Thursday at 10.",
  "ccList": [],
  "bccList": [],
  "emailType": "Text",
  "senderName": "Alex"
}

sendTo, subject and message are required, the rest optional. A send workflow that reads query as JSON copes
with either shape.

Set it up

1. Import the six tool workflows, add their own credentials and save each one.
2. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
3. In each tool node, from ⚙️Get Emails to ⚙️Summarize Database, pick the matching workflow from the list.
4. Add a Telegram credential with the bot token and select it in Telegram Trigger and Telegram.
5. Add an OpenAI credential and select it in OpenAI Chat Model, OpenAI Chat Model1 and Embeddings OpenAI.
6. Add a Pinecone credential, select it in Pinecone Vector Store and replace YOUR_PINECONE_INDEX with your index's
   name.
7. Load your contacts into the Contacts namespace, one record for each person with a name and an email, embedded
   with text-embedding-3-small.
8. In Telegram Trigger, under Additional Fields, add Restrict to User IDs with your own Telegram user ID.
9. Save, press Publish and message the bot.

To find your user ID, message the bot once while you test the workflow and read message.from.id in the trigger's
output.

How a message is answered

1. Telegram Trigger passes the message to AI Personal Assistant. The agent reads message.text, with today's date
   in its system message.
2. When the request names a person, the agent asks Vector Store Tool, Contacts_Data to the model. It searches
   Pinecone and has OpenAI Chat Model1 answer from the matches.
3. The agent calls the tool workflows it needs and reads each response.
4. Telegram sends the agent's answer to the sender's user ID.

Change the result

• What the agent knows and may do: the system message of AI Personal Assistant.
• When each tool is used: the description in each tool node, which the model reads.
• Memory: there is none, so each message stands alone. Add a Simple Memory node to AI Personal Assistant, keyed by
  {{ $('Telegram Trigger').item.json.message.chat.id }}, to let it follow a conversation.
• Models: gpt-4o in OpenAI Chat Model and OpenAI Chat Model1.

Good to know

• Until you restrict it, the trigger answers every chat that finds the bot, and the agent can send mail as you.
• Email goes out the moment the agent decides to send it, with no question first.
• The agent reads text only. A photo, a file or a voice note reaches it with no words.
• Nothing in this file writes to Pinecone. Records embedded with another model will not match the search.
• OpenAI bills the agent, the contact search's model and each embedding. Pinecone bills by its plan, and each tool
  workflow adds its own calls.
• OpenAI's deprecations page, read on 2 November 2026, lists GPT-4o (gpt-4o-2024-08-06 and gpt-4o-2024-11-20) for
  removal from the API on 14 January 2027, with gpt-6-luna as its replacement. Change both chat models before then.
• Your messages go to OpenAI, and your contacts sit in your Pinecone index.