🍽️ California Restaurant Concierge

A chatbot that recommends California restaurants by cuisine, city, vibe or budget. An LLM agent answers questions by calling tools on a Model Context Protocol (MCP) server, so every recommendation comes from real data rather than the model's imagination. It runs entirely on your own machine with a local model through Ollama: no API keys, no cost.

How it works

flowchart LR
    U[User] -->|question| UI[Gradio chat<br/>app.py]
    UI --> AG[Tool-calling agent<br/>agent.py]
    AG <-->|chat + tool calls| LLM[Local LLM<br/>Ollama · qwen2.5]
    AG <-->|MCP over stdio| SV[MCP server<br/>server.py]
    SV --> DB[(restaurants.json)]

1. The agent starts the MCP server and discovers its tools at runtime.

2. The LLM decides which tools to call and with which arguments.

3. The agent runs those calls through MCP and feeds the results back, looping until the LLM can answer.

4. The chat shows which tools were used for each answer.

MCP server

Dataset

data/restaurants.json holds fictional restaurants generated by scripts/generate_dataset.py. The script prompts a local model from seeds (city, neighborhood, cuisine, vibe), constrains its output to a JSON schema, and validates every record with Pydantic before saving. Regenerate or extend it by editing the seed list and running:

python scripts/generate_dataset.py

Related MCP server: Dine-Discover-AI MCP Server

Getting started

You need: Python 3.10+ and Ollama. A GPU with ~6 GB of memory makes responses fast, but a CPU also works.

git clone https://github.com/abellucido/restaurant-concierge-mcp.git
cd restaurant-concierge-mcp

ollama pull qwen2.5:7b

python -m venv .venv && source .venv/bin/activate   # or use conda
pip install -r requirements.txt

python src/concierge/app.py

Open http://127.0.0.1:7860 and try one of the example questions.

Optional settings (model, Ollama host, public share link) are listed in .env.example; copy it to .env to change them. With less GPU memory, try OLLAMA_MODEL=qwen2.5:3b.

Tests

The MCP tools are tested in-memory against a small fixture dataset, with no LLM needed. GitHub Actions runs them on every push.

pytest -v tests

Project structure

├── data/restaurants.json        Generated restaurant dataset
├── scripts/generate_dataset.py  Dataset generator (Ollama + JSON schema + Pydantic)
├── src/concierge/
│   ├── server.py                MCP server: tools and resource (FastMCP)
│   ├── agent.py                 MCP client + tool-calling loop (LangChain + Ollama)
│   └── app.py                   Chat UI (Gradio)
└── tests/                       Server tests and fixture data

Tech stack

Python · Model Context Protocol (FastMCP) · LangChain · Ollama (Qwen 2.5) · Gradio · Pydantic · pytest · GitHub Actions

Background

Inspired by what I learned in the IBM RAG and Agentic AI Professional Certificate on Coursera. This repository is an independent implementation with its own code and dataset; it contains no course materials.

License

MIT