The problem

Customer questions were coming in around the clock, but support only had eyes on it during business hours. Anything after hours sat unanswered until someone got to it, and repeat questions kept eating time that should've gone to harder problems.

Why a simple chatbot wasn't enough

A scripted bot with pre-set replies breaks the moment someone phrases a question differently than expected. It needed to actually understand the question and pull the right answer from real information, not a rigid decision tree.

What I built

A RAG-based support assistant deployed on Telegram. It retrieves answers from an actual knowledge base instead of guessing or hallucinating a response, and it runs 24/7 without anyone monitoring it.

Architecture

Incoming message → query embedded and matched against the knowledge base → relevant context retrieved → language model generates a grounded answer → response sent back on Telegram.

Tools used

n8n, RAG / vector search, Telegram API

Challenges

Getting retrieval to pull the right chunk of information consistently took the most tuning. Too broad and the answers got vague, too narrow and it missed context it needed.

What I'd improve next

Add a fallback path that flags a human when the model's confidence is low, instead of answering anyway.