RAGatouille

// Overview

RAGatouille is a Python library that makes state-of-the-art retrieval models easy to use in a RAG pipeline. Right now it focuses on ColBERT, a late-interaction model that matches queries and documents at the token level instead of comparing single dense embeddings.

It is aimed at developers who already run dense embedding retrieval and want stronger results without reading years of information-retrieval research. The README notes that ColBERT models generalise well to new domains and are data-efficient, and that the pretrained ColBERTv2 model often works well zero-shot, so you can prototype before training anything of your own.

Within the rerankers and hybrid-search space, RAGatouille offers strong but adjustable defaults: a few lines of code get you started, while components such as the training data processor and negative miners can be reused or swapped on their own.

// What it does

• Wraps ColBERT late-interaction retrieval behind a small Python API (RAGPretrainedModel and RAGTrainer)
• Works zero-shot with the pretrained colbert-ir/colbertv2.0 model, so no training is required to start
• Built-in RAGTrainer.prepare_training_data() converts pairs, labelled pairs, and triplets into training triplets
• Automatically mines hard negatives and removes duplicates during data preparation
• Supports both training a new ColBERT from a transformer and fine-tuning an existing ColBERT model
• Modular components (data processor, negative miners) are usable stand-alone or replaceable with your own

// Getting started

Install the package with pip, then either use a pretrained ColBERT model or prepare data to train one. Note: Windows is not supported (the README reports it only works under WSL2), and in a script you must run code inside an if __name__ == "__main__" guard.

pip install ragatouille

from ragatouille import RAGTrainer

my_data = [
    ("What is the meaning of life ?", "The meaning of life is 42"),
    ("What is Neural Search?", "Neural Search is a terms referring to a family of ..."),
]  # Unlabelled pairs here
trainer = RAGTrainer()
trainer.prepare_training_data(raw_data=my_data)

from ragatouille import RAGTrainer

trainer = RAGTrainer(
    model_name="MyFineTunedColBERT",
    pretrained_model_name="colbert-ir/colbertv2.0",
)

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Add a stronger retriever to an existing RAG pipeline where dense embeddings underperform on your domain
• Run zero-shot retrieval with pretrained ColBERTv2 while prototyping, before committing to any training
• Fine-tune ColBERT on your own query/passage pairs to improve retrieval for a specific corpus
• Reuse the data processor or hard-negative miner as stand-alone components in a custom training pipeline

// How RAGatouille compares

RAGatouille alongside other open-source rerank, search & hybrid tools AI/TLDR tracks, ranked by GitHub stars.