Caner Olcay
All projects

Cocktail RAG — MCP server + retrieval

A FastMCP server with 8 tools over a LangChain and FAISS retrieval layer, built on TheCocktailDB.

Role
Python / AI Developer
Timeline
2025
Status
Personal project
Area
Retrieval / MCP
Source on GitHub (opens in a new tab)

Overview

A small applied-NLP project for learning how an MCP server and a retrieval layer fit together. A FastMCP server exposes 8 tools for querying cocktails, and LangChain with a FAISS index handles semantic search over recipes and ingredients from TheCocktailDB.

Problem

Recipe data is structured, but people ask for drinks in plain language: by ingredient, taste, or what they have at home.

Approach

A retrieval layer over the recipe data, exposed as MCP tools so an assistant can query it directly.

How it works

  1. 01

    Prepare

    TheCocktailDB recipes

  2. 02

    Index

    Embeddings in FAISS

  3. 03

    Retrieve

    LangChain semantic search

  4. 04

    Serve

    8 tools on a FastMCP server

A preparation script loads TheCocktailDB into a FAISS index through LangChain; a FastMCP server exposes the query tools over the Model Context Protocol.

Outcomes

  • Exposed 8 cocktail query tools through an MCP server built with FastMCP.
  • Indexed 134 cocktails with full ingredient data for semantic search.
  • Practised the MCP and retrieval patterns later used on the ToPWR assistant.

Challenges

  • Choosing which queries deserve a tool and which belong to free-text retrieval.
  • Keeping ingredient matching useful when recipes name the same thing differently.

Stack

Python · FastMCP · LangChain · FAISS · OpenAI