Home

Awesome

Welcome to RAGatouille

Easily use and train state of the art retrieval methods in any RAG pipeline. Designed for modularity and ease-of-use, backed by research.

GitHub stars Python Versions Downloads Documentation Twitter Follow

<p align="center"><img width=500 alt="The RAGatouille logo, it's a cheerful rat on his laptop (branded with a slightly eaten piece of cheese) and a pile of books he's looking for information in." src="RAGatouille.png"/></p>

The main motivation of RAGatouille is simple: bridging the gap between state-of-the-art research and alchemical RAG pipeline practices. RAG is complex, and there are many moving parts. To get the best performance, you need to optimise for many components: among them, a very important one is the models you use for retrieval.

Dense retrieval, i.e. using embeddings such as OpenAI's text-ada-002, is a good baseline, but there's a lot of research showing dense embeddings might not be the best fit for your usecase.

The Information Retrieval research field has recently been booming, and models like ColBERT have been shown to generalise better to new or complex domains than dense embeddings, are ridiculously data-efficient and are even better suited to efficiently being trained on non-English languages with low amount of data! Unfortunately, most of those new approaches aren't very well known, and are much harder to use than dense embeddings.

This is where RAGatouille comes in: RAGatouille's purpose is to bridge this gap: make it easy to use state-of-the-art methods in your RAG pipeline, without having to worry about the details or the years of literature! At the moment, RAGatouille focuses on making ColBERT simple to use. If you want to check out what's coming next, you can check out our broad roadmap!

If you want to read more about the motivations, philosophy, and why the late-interaction approach used by ColBERT works so well, check out the introduction in the docs.

Want to give it a try? Nothing easier, just run pip install ragatouille and you're good to go!

⚠️ Running notes/requirements: ⚠️

Get Started

RAGatouille makes it as simple as can be to use ColBERT! We want the library to work on two levels:

<!-- (more on [components](https://ben.clavie.eu/ragatouille/components)). -->

In this section, we'll quickly walk you through the three core aspects of RAGatouille:

➡️ If you want just want to see fully functional code examples, head over to the examples⬅️

🚀 Training and fine-tuning

If you're just prototyping, you don't need to train your own model! While finetuning can be useful, one of the strength of ColBERT is that the pretrained models are particularly good at generalisation, and ColBERTv2 has repeatedly been shown to be extremely strong at zero-shot retrieval in new domains!

Data Processing

RAGatouille's RAGTrainer has a built-in TrainingDataProcessor, which can take most forms of retrieval training data, and automatically convert it to training triplets, with data enhancements. The pipeline works as follows:

This is all handled by RAGTrainer.prepare_training_data(), and is as easy as doing passing your data to it:

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)

ColBERT prefers to store processed training data on-file, which also makes easier to properly version training data via wandb or dvc. By default, it will write to ./data/, but you can override this by passing a data_out_path argument to prepare_training_data().

Just like all things in RAGatouille, prepare_training_data uses strong defaults, but is also fully parameterizable.

<!-- Check out the [Data Processing](https://ben.clavie.eu/ragatouille/data-processing) section of the docs! -->

Running the Training/Fine-Tuning

Training and Fine-Tuning follow the exact same process. When you instantiate RAGTrainer, you must pass it a pretrained_model_name. If this pretrained model is a ColBERT instance, the trainer will be in fine-tuning mode, if it's another kind of transformer, it will be in training mode to begin training a new ColBERT initialised from the model's weights!

from ragatouille import RAGTrainer
from ragatouille.utils import get_wikipedia_page

pairs = [
    ("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 ..."),
    # You need many more pairs to train! Check the examples for more details!
    ...
]

my_full_corpus = [get_wikipedia_page("Hayao_Miyazaki"), get_wikipedia_page("Studio_Ghibli")]


trainer = RAGTrainer(model_name = "MyFineTunedColBERT",
        pretrained_model_name = "colbert-ir/colbertv2.0") # In this example, we run fine-tuning

# This step handles all the data processing, check the examples for more details!
trainer.prepare_training_data(raw_data=pairs,
                                data_out_path="./data/",
                                all_documents=my_full_corpus)

trainer.train(batch_size=32) # Train with the default hyperparams

When you run train(), it'll by default inherit its parent ColBERT hyperparameters if fine-tuning, or use the default training parameters if training a new ColBERT. Feel free to modify them as you see fit (check the example and API reference for more details!)

🗄️ Indexing

To create an index, you'll need to load a trained model, this can be one of your own or a pretrained one from the hub! Creating an index with the default configuration is just a few lines of code:

from ragatouille import RAGPretrainedModel
from ragatouille.utils import get_wikipedia_page

RAG = RAGPretrainedModel.from_pretrained("colbert-ir/colbertv2.0")
my_documents = [get_wikipedia_page("Hayao_Miyazaki"), get_wikipedia_page("Studio_Ghibli")]
index_path = RAG.index(index_name="my_index", collection=my_documents)

You can also optionally add document IDs or document metadata when creating the index:

document_ids = ["miyazaki", "ghibli"]
document_metadatas = [
    {"entity": "person", "source": "wikipedia"},
    {"entity": "organisation", "source": "wikipedia"},
]
index_path = RAG.index(
    index_name="my_index_with_ids_and_metadata",
    collection=my_documents,
    document_ids=document_ids,
    document_metadatas=document_metadatas,
)

Once this is done running, your index will be saved on-disk and ready to be queried! RAGatouille and ColBERT handle everything here:

Curious about how this works? Check out the Late-Interaction & ColBERT concept explainer

<!-- or find out more about [indexing](https://ben.clavie.eu/ragatouille/indexing)! -->

🔎 Retrieving Documents

Once an index is created, querying it is just as simple as creating it! You can either load the model you need directly from an index's configuration:

from ragatouille import RAGPretrainedModel

query = "ColBERT my dear ColBERT, who is the fairest document of them all?"
RAG = RAGPretrainedModel.from_index("path_to_your_index")
results = RAG.search(query)

This is the preferred way of doing things, since every index saves the full configuration of the model used to create it, and you can easily load it back up.

RAG.search is a flexible method! You can set the k value to however many results you want (it defaults to 10), and you can also use it to search for multiple queries at once:

RAG.search(["What manga did Hayao Miyazaki write?",
"Who are the founders of Ghibli?"
"Who is the director of Spirited Away?"],)

RAG.search returns results in the form of a list of dictionaries, or a list of list of dictionaries if you used multiple queries:

# single-query result
[
    {"content": "blablabla", "score": 42.424242, "rank": 1, "document_id": "x"},
    ...,
    {"content": "albalbalba", "score": 24.242424, "rank": k, "document_id": "y"},
]
# multi-query result
[
    [
        {"content": "blablabla", "score": 42.424242, "rank": 1, "document_id": "x"},
        ...,
        {"content": "albalbalba", "score": 24.242424, "rank": k, "document_id": "y"},
    ],
    [
        {"content": "blablabla", "score": 42.424242, "rank": 1, "document_id": "x"},
        ...,
        {"content": "albalbalba", "score": 24.242424, "rank": k, "document_id": "y"},
    ],
 ],

If your index includes document metadata, it'll be returned as a dictionary in the document_metadata key of the result dictionary:

[
    {"content": "blablabla", "score": 42.424242, "rank": 1, "document_id": "x", "document_metadata": {"A": 1, "B": 2}},
    ...,
    {"content": "albalbalba", "score": 24.242424, "rank": k, "document_id": "y", "document_metadata": {"A": 3, "B": 4}},
]

I'm sold, can I integrate late-interaction RAG into my project?

To get started, RAGatouille bundles everything you need to build a ColBERT native index and query it. Just look at the docs! RAGatouille persists indices on disk in compressed format, and a very viable production deployment is to simply integrate the index you need into your project and query it directly. Don't just take our word for it, this is what Spotify does in production with their own vector search framework, serving dozens of millions of users:

Statelessness: Many of Spotify’s systems use nearest-neighbor search in memory, enabling stateless deployments (via Kubernetes) and almost entirely removing the maintenance and cost burden of maintaining a stateful database cluster. (Spotify, announcing Voyager)

Integrations

If you'd like to use more than RAGatouille, ColBERT has a growing number of integrations, and they all fully support models trained or fine-tuned with RAGatouille!