ShopBot AI Server

ShopBot AI is an e-commerce support chatbot I built to explore how far you can push LLMs when they’re connected to real systems like databases and retrieval pipelines.

Instead of relying only on prompt-based answers, the bot can actually:

• check real orders stored in MySQL
• search a knowledge base using embeddings (FAISS)
• decide when it should use tools vs when it should just answer normally

The idea was simple:
a chatbot that doesn’t hallucinate when it matters (like order status).

🔗 Live demo: https://shopbot-ai-os4z.onrender.com

---

Why I built this

Most chatbot demos feel impressive at first, but break down quickly when you ask:

• “Where is my order?”
• “What exactly is your refund policy?”
• “Is this product in stock?”

They usually respond confidently… even when they shouldn’t.

I wanted to build something closer to how a real support system should behave:

• if it’s factual → query a database
• if it’s semantic → search knowledge base
• otherwise → let the model respond normally

This pushed me into combining RAG, tool calling, and routing logic in one system.

---

How it works

User message ↓ Flask API ↓ Gemini decides intent (route query) ↓ MCP tool layer ├── MySQL tool → order / customer data └── FAISS tool → FAQ / policy retrieval ↓ ↓ Final response to user

---

Key design choices (and why I made them)

MCP instead of direct function calls

I used MCP because I didn’t want tool logic mixed directly into the LLM layer. It keeps things modular, so adding a new tool later doesn’t require rewriting the whole routing logic.

---

FAISS instead of a hosted vector DB

I went with FAISS mainly because I wanted everything to run locally without extra infrastructure costs. It keeps the project self-contained and easier to demonstrate.

---

Gemini for routing

Gemini is only used to decide what kind of question this is, not to store any “truth.” That separation was intentional to reduce hallucination risk.

---

Flask instead of FastAPI

This started as a prototype, so Flask was simply faster to iterate with. I prioritized building the system over framework perfection.

---

What I learned from building this

• Tool use matters more than prompt engineering for real-world reliability
• RAG alone isn’t enough unless retrieval is well scoped
• A simple routing layer improves accuracy more than expected
• Most “AI chatbot problems” are actually system design problems