AIDataNordic — Food Recipe MCP

A production-grade semantic search server for food recipes — built for AI agents using the Model Context Protocol (MCP). Search across 50,000+ recipes with hybrid dense + sparse retrieval and cross-encoder reranking.

---

What This Is

A self-hosted MCP server exposing a recipe dataset through semantic search. Agents can query by natural language, filter by diet, difficulty, time, and servings — and get back ranked, structured recipe data including ingredients, instructions, and nutrition.

Designed for autonomous machine-to-machine consumption via FastMCP 3.2 over HTTP.

---

Architecture

Query (natural language)
        ↓
  Dense embedding          Sparse embedding
  (e5-large-v2)            (BM25 / fastembed)
        ↓                        ↓
       Qdrant — Hybrid Fusion (RRF)
                    ↓
          Cross-encoder reranking
          (mmarco-mMiniLMv2-L12-H384-v1)
                    ↓
          Structured JSON response
                    ↓
           MCP tool / AI agent

---

Data Coverage

Field	Details
Total recipes	50,000+
Source	food.com and others
Fields	title, description, ingredients, instructions, nutrition, rating, difficulty, diet, total_time, servings
Diet tags	vegetarian, vegan, gluten-free, dairy-free
Difficulty	easy, medium, hard

---

Technical Stack

Search

• Dense embeddings: intfloat/e5-large-v2 (1024 dim)
• Sparse embeddings: Qdrant/bm25 via fastembed
• Fusion: Reciprocal Rank Fusion (RRF) in Qdrant
• Reranker: cross-encoder/mmarco-mMiniLMv2-L12-H384-v1

Serving

• FastMCP 3.2 over HTTP (/mcp endpoint)
• Compatible with Claude, LangChain, and any MCP-capable agent

Infrastructure

• Ubuntu Server 24 LTS, self-hosted
• Qdrant vector database (self-hosted)

---

MCP Tool

search_recipes(
    query="quick chicken pasta",       # required — natural language
    diet="vegetarian",                 # optional: vegetarian, vegan, gluten-free, dairy-free
    difficulty="easy",                 # optional: easy, medium, hard
    max_minutes=30,                    # optional: maximum total time in minutes
    servings=4,                        # optional: number of servings
    limit=5                            # optional: number of results (default 5)
)
# Returns semantically ranked recipes with ingredients, instructions, nutrition, and ratings

Example response

[
  {
    "rerank_score": 7.96,
    "title": "quick and easy chicken pasta salad",
    "description": "great use for left-over chicken.",
    "total_time": 25,
    "difficulty": "medium",
    "diet": [],
    "main_ingredient": "chicken",
    "ingredients": ["cooked chicken", "pasta shells", "tomatoes", "italian dressing"],
    "instructions": ["combine ingredients", "pour dressing", "chill 1 hour"],
    "nutrition": {"calories": 424, "protein_g": 26, "carbs_g": 38.5, "fat_g": 19.5},
    "rating": 4.8,
    "rating_count": 5
  }
]

---

Quickstart

1. Install dependencies

pip install -r requirements.txt

2. Start the server

python recipe_mcp_server.py

Server starts at http://localhost:8004/mcp

3. Connect with FastMCP client

import fastmcp, asyncio

async def main():
    async with fastmcp.Client("http://localhost:8004/mcp") as client:
        result = await client.call_tool("search_recipes", {
            "query": "quick chicken pasta",
            "max_minutes": 30,
            "limit": 3
        })
        for recipe in result.structured_content["result"]:
            print(recipe["title"], "-", recipe["total_time"], "min")

asyncio.run(main())

4. Connect with Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "food-recipes": {
      "url": "https://recipes.aidatanorge.no/mcp"
    }
  }
}

---

Live Demo

Try the search interface at recipes.aidatanorge.no

---

Files

File	Description
recipe_mcp_server.py	FastMCP server with hybrid search
mcp_client.py	Example Python client
requirements.txt	Python dependencies

---

Built and operated as part of AIDataNordic — self-hosted AI data infrastructure for autonomous agents.