Telegram RAG bot answering from uploaded knowledge base files

How it works

1. 1.A Telegram trigger listens for all updates and a Get a chat node resolves the chat from its ID.
2. 2.The message text is sent to an AI Agent whose system message limits answers to the knowledge base, with max 10 iterations.
3. 3.The agent calls the Retrieve documents tool to fetch the top 10 chunks from the shared vector_store_key store.
4. 4.The agent runs on the Groq chat model and keeps context via window memory keyed by the Telegram chat ID.
5. 5.The agent output is sent back to the same Telegram chat with attribution disabled.
6. 6.Files uploaded via the web form are loaded, split, embedded, and inserted into the in-memory vector store.

Use cases

• Internal Q&A assistant over uploaded policy PDFs or CSV reports on Telegram.
• Customer support bot restricted to product documentation without outside knowledge.
• Quick RAG prototype before moving to a persistent vector database.

Requirements

• Telegram credentials with a valid bot token for the trigger, get-chat, and send nodes.
• Groq API credential to run the llama-3.1-8b-instant model.
• HuggingFace Inference access token for both embedding nodes (insert and retrieve).
• The form trigger URL must be shared so users can upload PDF/CSV files.
• Note that the in-memory store is not persistent and is lost when executions restart.

Setup and testing in n8n

1. Activate the workflow, open the form URL, and upload a small PDF or CSV file.
2. Ask a question in Telegram about the file content and confirm the reply arrives in the same chat.
3. Ask about a topic not in the documents and verify the bot declines to answer beyond the knowledge base.
4. Send follow-up questions to confirm chat memory retains conversation context.
5. Inspect the execution log to confirm the retrieval tool is called and iterations stay within 10.

Service Value

Ready-made workflow template for automation delivery and service execution.