Telegram bot that answers questions from a PDF in Pinecone

Telegram bot that answers questions from a PDF in Pinecone

An n8n workflow with two jobs on one Telegram bot. Send the bot a PDF, and it cuts the text into pieces and stores them in a
Pinecone index. Send it a question, and it finds the pieces that fit, hands them to a model on Groq and sends back the answer.

What is inside

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

What we changed from the original copy

• Groq Chat Model asked for llama-3.1-70b-versatile. Groq's models page, read on 12 November 2026, lists
  openai/gpt-oss-120b as a production model and does not list that name, so this copy uses openai/gpt-oss-120b.
• Both embeddings nodes left the model unset on node version 1, which means text-embedding-ada-002. They are now on
  version 1.2 and name text-embedding-3-small. Both vectors have 1536 numbers, but the two models give different
  vectors, so fill a new index and do not reuse one that the original copy filled.
• Both Pinecone nodes pointed at an index named telegram, which was the author's. They now name YOUR_PINECONE_INDEX, and the node that
  searches has its mode set to retrieve on purpose.
• Telegram Trigger is now on node version 1.2, which can limit who may use the bot. See Good to know.
• The comments in Change to application/pdf were in Portuguese. They are in English now, and the code does the same work.
• The two sticky notes are rewritten in plain words. The time zone and the caller policy of the author's n8n are removed from
  the settings.
• Every credential reference is removed, so you pick your own.

Before you start

• An n8n instance with a public HTTPS address. Telegram delivers updates by an HTTPS POST to the webhook address.
• A Telegram bot and its token. You make one by chatting with BotFather inside Telegram.
• A Pinecone account and an index with 1536 dimensions. Pinecone's guide says to set the dimension and the metric to match your
  embedding model, and OpenAI's guide gives 1536 for text-embedding-3-small. Pick the cosine metric.
• An OpenAI API key for the embeddings, and a Groq API key for the answers.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. In Pinecone Vector Store and in Pinecone Vector Store1, replace YOUR_PINECONE_INDEX with the name of your index and
   add a Pinecone credential with your API key.
3. In Embeddings OpenAI and in Embeddings, add an OpenAI credential with your API key. Keep the same model in both.
4. In Groq Chat Model, add a Groq credential with your API key.
5. In Telegram Trigger, Telegram get File, Telegram Response and Telegram Response about Database, add a Telegram
   credential with your bot token.
6. Press Test workflow and send a small PDF to your bot. The bot answers with the number of pages that it saved.
7. Press Test workflow again and send a question about that PDF. The answer arrives in the same chat.
8. Turn the workflow on, so that the bot keeps answering when you are away.

How a run goes

1. Telegram Trigger starts the run when someone sends a message to your bot.
2. Check If is a document looks for a file in the message. A file goes the upper way, and anything else goes to the question path.
3. File path: Telegram get File downloads the file, and Change to application/pdf marks it as a PDF.
4. Pinecone Vector Store inserts it. Default Data Loader reads the PDF, Recursive Character Text Splitter cuts it into
   pieces of 3000 characters with 200 characters of overlap, and Embeddings OpenAI turns each piece into numbers.
5. Limit to 1 keeps one item, and Telegram Response about Database tells the sender how many pages were saved.
6. Question path: Question and Answer Chain takes the text of the message. Vector Store Retriever asks
   Pinecone Vector Store1 for the closest pieces, with Embeddings turning the question into numbers.
7. Groq Chat Model writes the answer from those pieces, and Telegram Response sends it to the chat.
8. A failed send goes to a Stop and Error node, so the run shows as failed in n8n.

Change the look

• The model: the model name in Groq Chat Model.
• The size of the pieces: Chunk Size and Chunk Overlap in Recursive Character Text Splitter.
• How many pieces the model sees: the limit in Vector Store Retriever, which keeps its default of four.
• The wording of the question: the text in Question and Answer Chain.
• Who may use the bot: see the first item under Good to know.

Good to know

• Every PDF goes into one pool. The index has no namespace, so anyone who can write to the bot can ask about any file that
  anyone sent. To limit it, open Telegram Trigger, add the field Restrict to User IDs under Additional Fields and list the
  Telegram user numbers that may use it. It is enforced on node version 1.2 and later.
• Sending the same PDF twice stores its pieces twice. The option Clear Namespace is off, and it only works when you name a
  namespace.
• Telegram lets a bot download files up to 20 MB, so a larger PDF fails at Telegram get File.
• The code marks every file as a PDF, so a Word file or a picture fails. A PDF made of scanned pictures has no text to store.
• A message that is neither a file nor text, such as a photo, stops at the chain with an empty text error.
• openai/gpt-oss-120b is a reasoning model, so an answer can take a little longer than with a plain chat model.
• Replies go out in Telegram's Markdown mode, which is the n8n default. A lone _ or * in an answer can make Telegram
  refuse the message.
• Your data: the text of every PDF goes to OpenAI for the embeddings and is stored in your Pinecone index. Each question and
  the pieces that fit it go to Groq.
• Not run: we did not call any of the services to test this file. We ran it in a test runner with stand-in answers. The
  question path ran from the message to the reply. The file path ran with a text file in place of a PDF, because the
  runner cannot read PDFs, so the page count in the last message was not tried.
• Credentials: the JSON holds none.