colpali

ColPali: Visual Retriever based on PaliGemma-3B with ColBERT strategy

ColPali is a model based on a novel model architecture and training strategy based on Vision Language Models (VLMs) to efficiently index documents from their visual features. It is a PaliGemma-3B extension that generates ColBERT- style multi-vector representations of text and images. It was introduced in the paper ColPali: Efficient Document Retrieval with Vision Language Models and first released in this repository

<p align="center"><img width=800 src="https://github.com/illuin-tech/colpali/blob/main/assets/colpali_architecture.webp?raw=true"/></p>

Model Description

This model is built iteratively starting from an off-the-shelf SigLIP model. We finetuned it to create BiSigLIP and fed the patch-embeddings output by SigLIP to an LLM, PaliGemma-3B to create BiPali.

One benefit of inputting image patch embeddings through a language model is that they are natively mapped to a latent space similar to textual input (query). This enables leveraging the ColBERT strategy to compute interactions between text tokens and image patches, which enables a step-change improvement in performance compared to BiPali.

Model Training

Dataset

Our training dataset of 127,460 query-page pairs is comprised of train sets of openly available academic datasets (63%) and a synthetic dataset made up of pages from web-crawled PDF documents and augmented with VLM-generated (Claude-3 Sonnet) pseudo-questions (37%). Our training set is fully English by design, enabling us to study zero-shot generalization to non-English languages. We explicitly verify no multi-page PDF document is used both ViDoRe and in the train set to prevent evaluation contamination. A validation set is created with 2% of the samples to tune hyperparameters.

Note: Multilingual data is present in the pretraining corpus of the language model (Gemma-2B) and potentially occurs during PaliGemma-3B's multimodal training.

Parameters

All models are trained for 1 epoch on the train set. Unless specified otherwise, we train models in bfloat16 format, use low-rank adapters (LoRA) with alpha=32 and r=32 on the transformer layers from the language model, as well as the final randomly initialized projection layer, and use a paged_adamw_8bit optimizer. We train on an 8 GPU setup with data parallelism, a learning rate of 5e-5 with linear decay with 2.5% warmup steps, and a batch size of 32.

Usage

Using Sentence Transformers

ColPali can be used as a multi-vector (ColBERT-style late interaction) retriever directly with Sentence Transformers via the MultiVectorEncoder.

pip install "sentence-transformers[image]>=6.0.0"

from sentence_transformers import MultiVectorEncoder

model = MultiVectorEncoder("vidore/colpali")

queries = [
    "What is the variable represented on the y-axis of the graph?",
    "Total outlay is maximum in which year?",
]
images = [
    "https://huggingface.co/datasets/sentence-transformers/example-documents/resolve/main/doc1.jpg",
    "https://huggingface.co/datasets/sentence-transformers/example-documents/resolve/main/doc2.jpg",
    "https://huggingface.co/datasets/sentence-transformers/example-documents/resolve/main/doc3.jpg",
    "https://huggingface.co/datasets/sentence-transformers/example-documents/resolve/main/doc4.jpg",
]

query_embeddings = model.encode_query(queries)
document_embeddings = model.encode_document(images)
print(f"Query 0 shape:    {tuple(query_embeddings[0].shape)}")
print(f"Document 0 shape: {tuple(document_embeddings[0].shape)}")
# Query 0 shape:    (23, 128)
# Document 0 shape: (1030, 128)

# MaxSim late-interaction scoring (rows = queries, columns = images)
scores = model.similarity(query_embeddings, document_embeddings)
print(scores)
# tensor([[17.3789, 17.1055, 15.4727, 15.4082],
#         [ 8.3750, 12.3047,  8.5898,  9.0957]])

Using ColPali Engine

[!WARNING] Note: current colpali-engine no longer sends the query prefix and trailing newline that this checkpoint was trained with. The trailing newline went in 0.3.11 (illuin-tech/colpali#280) and the prefix in 0.3.13 (illuin-tech/colpali#339). The Sentence Transformers configuration in this repository reproduces the original training-time format, so its embeddings differ slightly from current colpali-engine output. Release 0.3.4 had already changed the prefix from Question: to Query: (illuin-tech/colpali#125), which this checkpoint predates. The Sentence Transformers configuration also sends token_type_ids to the model, which on transformers 5.x is what makes PaliGemma build an explicit attention mask at all. Without it no mask is materialized and the shorter queries in a batch attend to their own padding.

For best performance, newer models are available (vidore/colpali-v1.2)

# This model checkpoint is compatible with version 0.1.1, but not more recent versions of the inference lib
pip install colpali_engine==0.1.1

import torch
import typer
from torch.utils.data import DataLoader
from tqdm import tqdm
from transformers import AutoProcessor
from PIL import Image

from colpali_engine.models.paligemma_colbert_architecture import ColPali
from colpali_engine.trainer.retrieval_evaluator import CustomEvaluator
from colpali_engine.utils.colpali_processing_utils import process_images, process_queries
from colpali_engine.utils.image_from_page_utils import load_from_dataset


def main() -> None:
    """Example script to run inference with ColPali"""

    # Load model
    model_name = "vidore/colpali"
    model = ColPali.from_pretrained("vidore/colpaligemma-3b-mix-448-base", torch_dtype=torch.bfloat16, device_map="cuda").eval()
    model.load_adapter(model_name)
    processor = AutoProcessor.from_pretrained(model_name)

    # select images -> load_from_pdf(<pdf_path>),  load_from_image_urls(["<url_1>"]), load_from_dataset(<path>)
    images = load_from_dataset("vidore/docvqa_test_subsampled")
    queries = ["From which university does James V. Fiorca come ?", "Who is the japanese prime minister?"]

    # run inference - docs
    dataloader = DataLoader(
        images,
        batch_size=4,
        shuffle=False,
        collate_fn=lambda x: process_images(processor, x),
    )
    ds = []
    for batch_doc in tqdm(dataloader):
        with torch.no_grad():
            batch_doc = {k: v.to(model.device) for k, v in batch_doc.items()}
            embeddings_doc = model(**batch_doc)
        ds.extend(list(torch.unbind(embeddings_doc.to("cpu"))))

    # run inference - queries
    dataloader = DataLoader(
        queries,
        batch_size=4,
        shuffle=False,
        collate_fn=lambda x: process_queries(processor, x, Image.new("RGB", (448, 448), (255, 255, 255))),
    )

    qs = []
    for batch_query in dataloader:
        with torch.no_grad():
            batch_query = {k: v.to(model.device) for k, v in batch_query.items()}
            embeddings_query = model(**batch_query)
        qs.extend(list(torch.unbind(embeddings_query.to("cpu"))))

    # run evaluation
    retriever_evaluator = CustomEvaluator(is_multi_vector=True)
    scores = retriever_evaluator.evaluate(qs, ds)
    print(scores.argmax(axis=1))


if __name__ == "__main__":
    typer.run(main)

Limitations

• Focus: The model primarily focuses on PDF-type documents and high-ressources languages, potentially limiting its generalization to other document types or less represented languages.
• Support: The model relies on multi-vector retreiving derived from the ColBERT late interaction mechanism, which may require engineering efforts to adapt to widely used vector retrieval frameworks that lack native multi-vector support.

License

ColPali's vision language backbone model (PaliGemma) is under gemma license as specified in its model card. The adapters attached to the model are under MIT license.

Contact

• Manuel Faysse: manuel.faysse@illuin.tech
• Hugues Sibille: hugues.sibille@illuin.tech
• Tony Wu: tony.wu@illuin.tech

Citation

If you use any datasets or models from this organization in your research, please cite the original dataset as follows:

@misc{faysse2024colpaliefficientdocumentretrieval,
  title={ColPali: Efficient Document Retrieval with Vision Language Models}, 
  author={Manuel Faysse and Hugues Sibille and Tony Wu and Bilel Omrani and Gautier Viaud and Céline Hudelot and Pierre Colombo},
  year={2024},
  eprint={2407.01449},
  archivePrefix={arXiv},
  primaryClass={cs.IR},
  url={https://arxiv.org/abs/2407.01449}, 
}