Fork&Fetch 🍴

Fork&Fetch is a Retrieval-Augmented Generation (RAG) powered recipe assistant. Search for recipes by dish name, ingredient, or craving — the AI chef finds and presents the best match from a collection of 125,000 recipes.

📌 Features

• 🔍 Semantic recipe search using Chroma and HuggingFace sentence-transformers
• 🍳 Intelligent recipe suggestions powered by LLaMA 3.3 70B via OpenRouter
• 🧠 LangChain LCEL pipeline combining retrieval with LLM reasoning
• 🎨 Polished HTML/CSS frontend served by Flask
• 🐳 Fully containerized with Docker

📂 Project Structure

fork-and-fetch/
├── app.py                  # Flask API + static file server
├── Dockerfile
├── docker-compose.yml
├── pyproject.toml
├── uv.lock
├── frontend/
│   └── index.html          # HTML frontend
├── stages/
│   ├── preprocessor.py     # Dataset cleaning and preparation
│   ├── pipeline.py         # Runs preprocessing → embedding in sequence
│   └── embedder.py         # Chroma vector store generation
└── data/                   # Add dataset here (see Setup step 0)

📈 Results

Automated Metrics

Metric	Type	Base LLM	RAG	Improvement (%)
BLEU	1-gram	0.69	0.76	+8.9%
	2-gram	0.62	0.69	+11.2%
	3-gram	0.55	0.64	+18.1%
	4-gram	0.49	0.61	+26.7%
Rouge-1	F1 score	0.69	0.85	+22.95%
	Precision	0.69	0.90	+30.64%
	Recall	0.70	0.81	+15.95%
Rouge-2	F1 score	0.42	0.71	+69.20%
	Precision	0.41	0.76	+83.09%
	Recall	0.43	0.68	+56.19%
Rouge-L	F1 score	0.52	0.79	+51.23%
	Precision	0.52	0.84	+61.15%
	Recall	0.54	0.75	+41.03%

📚 Dataset

Sourced from Eight Portions — ~125,000 recipes scraped from various food websites, each including title, ingredients, instructions, and source URL.

🛠️ Technology Stack

Layer	Technology
Vector database	Chroma
Embeddings	HuggingFace all-MiniLM-L6-v2
LLM	LLaMA 3.3 70B via OpenRouter
RAG framework	LangChain LCEL
Backend	Flask
Frontend	HTML / CSS / JS
Container	Docker + uv
GPU acceleration	CUDA (when available)

🚀 Setup

0. Download the dataset

1. Download the dataset from Eight Portions
2. Unzip and place the CSV in data/:

data/
└── recipes_cleaned.csv

1. Configure environment

cp .env.example .env
# add your OPENROUTER_API_KEY

2. Build the vector database (once)

python stages/pipeline.py

3. Start the app

docker compose up --build   # first run
docker compose up           # subsequent runs

Then open http://localhost:5000.

📋 Prerequisites

• Python 3.11+
• Docker
• OpenRouter API key
• CUDA-compatible GPU (optional)