Telegram bot answering from an uploaded PDF or CSV

Telegram bot answering from an uploaded PDF or CSV

An n8n workflow with two ways in. A web form takes a PDF or CSV, and a Telegram bot answers questions about it. The file is cut into chunks, embedded with Hugging Face and kept in n8n's own memory.

What is inside

• workflow.json: the workflow as n8n exports it, fifteen nodes. Thirteen do the work and two are empty sticky notes that frame the two halves. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The chat model was Groq's llama-3.1-8b-instant. Groq's deprecations page, read on 12 November 2026, lists it as shut down on 16 August 2026 for free and developer plans. Mindket's copy uses openai/gpt-oss-20b, the model that page names as its replacement.
• The data loader read each item as JSON. A form upload carries the file as binary data, so the loader would have stored only the file's name, type and size. It now reads the binary field Upload_your_file, which holds the file itself.
• The form field took several files and was called Upload your file(s). It now takes one file and is called Upload your file, so the binary field always has the same name.
• The system message had typos. Mindket's copy says the same thing in clean words.
• Several nodes were linked to credentials in the author's n8n. The links are gone, so select your own.
• Nothing else. The chunk size, the memory and the two embedding nodes are the file's own.

Before you start

• An n8n instance with a public HTTPS address, on n8n Cloud or your own server. Telegram sends every message to that address.
• A Telegram bot made with BotFather, and its token.
• A Groq API key.
• A Hugging Face access token.
• A PDF or CSV file to upload.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add a Telegram credential with your bot token. Select it in Telegram Message Trigger1, Get a chat1 and Reply in Telegram1.
3. Add a Groq credential and select it in Groq Chat Model.
4. Add a Hugging Face credential and select it in Embeddings HuggingFace Inference and Embeddings HuggingFace Inference1. Both nodes use the node's default model, so leave them the same.
5. Activate the workflow.
6. Open the production address shown in Upload your file here. Upload a PDF or CSV and wait for the page to finish.
7. Write to your bot in Telegram and ask something the file can answer.

How a run goes

1. A file sent through the form starts Upload your file here. Default Data Loader1 reads it, and Recursive Character Text Splitter cuts it into chunks of 3000 characters with an overlap of 200.
2. Embeddings HuggingFace Inference turns each chunk into a vector. Simple Vector Store saves the vectors in n8n's memory under the key vector_store_key.
3. A message to the bot starts Telegram Message Trigger1. Get a chat1 asks Telegram for the chat's details and passes the run on to AI Agent1.
4. AI Agent1 reads the question from the trigger. It has Retrieve documents as a tool, which searches the store for the ten closest chunks. Embeddings HuggingFace Inference1 embeds the question for that search.
5. Groq Chat Model writes the answer from those chunks, and Simple Memory keeps the recent messages of the chat.
6. Reply in Telegram1 sends the answer to the chat.

Change the result

• Rules and tone: the system message in AI Agent1.
• Model: model in Groq Chat Model.
• Chunk size and overlap: chunkSize and chunkOverlap in Recursive Character Text Splitter.
• Chunks read for each question: topK in Retrieve documents.
• File types: acceptFileTypes in Upload your file here.

Good to know

• Mindket did not run this workflow. Telegram, Groq and Hugging Face were not called. Every model, field and limit above comes from their docs, their model data or n8n's source, read on 12 November 2026.
• The store is a map inside the n8n process. A restart empties it, so upload the file again afterwards. By default nothing limits its size or age. The environment variables N8N_VECTOR_STORE_MAX_MEMORY and N8N_VECTOR_STORE_TTL_HOURS set limits.
• Version 1.3 of the store node uses the key vector_store_key as it is. Another workflow with the same key reads and writes the same store.
• Each upload adds to the store. Nothing is cleared between uploads, because clearStore is off.
• Both embedding nodes use the node's default model, sentence-transformers/distilbert-base-nli-mean-tokens. Hugging Face's model data listed it as live on its inference provider on 12 November 2026. Hugging Face bills inference through your account.
• The embedding nodes must use the same model, or the search compares vectors that do not match.
• Telegram's Markdown mode rejects text with an unpaired _ or *, and a message holds at most 4096 characters (Telegram's Bot API page).
• Telegram needs an HTTPS address for its webhook, so a local n8n without a public address cannot receive messages.
• The chunks of your file and your questions go to Hugging Face. The questions and the chunks found go to Groq.
• The file holds no credential and no key.