← All projects

(Case study · 09)

Support AI BotTelegram Support with RAG

By Md. Habibur Rahman Shohel · 2026

  • Node.js
  • Express.js
  • Next.js
  • MongoDB
  • OpenAI
  • Telegram
Support AI Bot preview

A Telegram support bot that answers from a company's own documents, hands hard cases to a person, and comes with an admin panel.

A support bot for Telegram. Customers ask questions in chat; the bot finds the answer in the company's own documents — PDFs, text, Markdown, JSON and Telegram channel exports — and replies in context, remembering the conversation. When a question needs a person, it hands over and alerts an admin. Admins manage the knowledge base, review answers and correct them in a Next.js panel. Under the hood: retrieval-augmented generation with OpenAI embeddings and MongoDB Atlas Vector Search, on a Node.js and Express API.

Year
2026
Stack
Node.js · Express.js · Next.js
Private client project

(How it works)

Telegram · Support AI BotOnline
↑

Animated illustration of the real flow, with sample data — not a live system.

(What I built)

  • 01Telegram chat that remembers the conversation
  • 02Answers from the company's documents (RAG with OpenAI and vector search)
  • 03Upload PDF, TXT, Markdown, JSON or Telegram exports
  • 04Hands a chat to a person and alerts an admin
  • 05Admin panel to review, correct and re-teach answers
  • 06Links to matching YouTube videos in replies

Process

(Problem → what I did)
  1. 01

    Problem

    Finding the right context in a knowledge base that keeps growing.

    What I did

    • Semantic search with OpenAI embeddings and MongoDB Atlas Vector Search.
  2. 02

    Problem

    Remembering the conversation across Telegram sessions.

    What I did

    • Stored each user's conversation context in MongoDB, so multi-turn chats keep their memory.
  3. 03

    Problem

    Making admin feedback actually improve the answers.

    What I did

    • Built an admin panel to review and correct answers, and re-embed the corrected content.
Show all 5 steps
  1. 04

    Problem

    Documents arrive as PDF, TXT, Markdown, JSON and Telegram exports.

    What I did

    • Wrote one set of parsers so every format is parsed and chunked the same way.
  2. 05

    Problem

    Handing a chat to a person only when it's really needed.

    What I did

    • Keyword and intent detection decides when to escalate, and notifies an admin on Telegram.

The product

8 screens from the real product. Click any screen to see it full size.

(Shipped)

The bot answers around the clock from one searchable knowledge base instead of scattered documents, hands anything it shouldn't answer to a person, and gets better as admins correct its answers.

(Guide from this project)How an AI chatbot answers from your own documents5 min read

Next project

TrueTerra Oils AI Chatbot

Need something like Support AI Bot? I build web, mobile and backend — end to end.

Hiring? See my experience →