SauerkrautLM-ColPali: Multi-Vector Vision Retrieval Models

Fork Notice: This repository is a fork of colpali-engine by Illuin Technology. We extend the original codebase with additional model architectures for document retrieval using vision language models.

Overview

SauerkrautLM-ColPali provides model implementations and processors for multi-vector vision retrieval based on the ColPali architecture. This package includes support for several VLM backbones:

• ColQwen3 - Based on Qwen3-VL (2B, 4B, 8B)
• ColLFM2 - Based on LargeFlamingoModel 2 (~450M parameters)
• ColMinistral3 - Based on Ministral-3B-Instruct with Pixtral vision encoder

Models

Supported Languages: English, German, French, Spanish, Italian, Portuguese

🎯 Why Visual Document Retrieval?

Traditional document retrieval relies on OCR + Text Search, which has significant limitations:

Visual Document Retrieval solves these problems by:

• ✅ Direct visual understanding - No OCR errors, preserves full document context
• ✅ Layout-aware - Understands tables, forms, multi-column layouts
• ✅ Multimodal - Combines text and visual elements naturally
• ✅ End-to-end - Single model for retrieval, no pipeline complexity

🏆 Benchmark Results

Our models achieve state-of-the-art performance on the ViDoRe (Visual Document Retrieval) benchmarks while maintaining a compact 128-dimensional embedding space for efficient retrieval.

Key Highlights

128-dim Models Comparison (Fair Comparison)

When comparing only models with the same 128-dimensional embedding space:

Rankings among 128-dim models only. ⭐ = Best in category. Bold = our models.

Size Category Comparison (128-dim Models)

Small Models (<1B):

Medium Models (1-3B):

Large Models (3-5B):

XLarge Models (5-10B):

⭐ = Best 128-dim model in category

Why Choose Our Models?

1. 🏆 #1 in 128-dim Class: Our ColQwen3-8b beats ALL other 128-dim models on ViDoRe v1 and v3

2. ⚡ Compact Embeddings: All our models use 128 dimensions - same as ColPali/ColQwen2/colSmol

   • No storage overhead compared to standard ColPali models
   • 2.5x smaller than tomoro (320 dim)
   • 16-24x smaller than llama-nemoretriever (2048-3072 dim)

3. 💰 Best-in-Class for Every Size:

   • Small (<1B): ColLFM2-450M beats colSmol-500M with 10% fewer parameters
   • Medium (1-3B): ColQwen3-2b beats colqwen2-v1.0 by +1.01 points
   • Large (3-5B): ColQwen3-4b achieves 90.80, only -0.20 behind much larger llama-nemo
   • XLarge (5-10B): ColQwen3-8b achieves the highest 128-dim score ever

4. 🌍 Multilingual: Trained on 6 languages (EN, DE, FR, ES, IT, PT)

5. 🔧 Easy Integration: MTEB-compatible for standardized evaluation

Training

Hardware

Training Datasets

Our models were trained on a diverse mix of public and proprietary datasets:

Special Training Techniques

Installation

# From source (recommended)
pip install git+https://github.com/VAGOsolutions/sauerkrautlm-colpali

# For ColMinistral3 models (requires transformers 5.0.0rc0)
pip install "sauerkrautlm-colpali[ministral]"

Note: ColMinistral3 requires transformers>=5.0.0rc0. Install with pip install "sauerkrautlm-colpali[ministral]" or manually install the RC version.

Quick Start

ColQwen3 Example

import torch
from PIL import Image
from sauerkrautlm_colpali.models import ColQwen3, ColQwen3Processor

model_name = "VAGOsolutions/SauerkrautLM-ColQwen3-2b-v0.1"

# Load model and processor
model = ColQwen3.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="cuda:0",
).eval()

processor = ColQwen3Processor.from_pretrained(model_name)

# Prepare inputs
images = [Image.open("document.png")]
queries = ["What is the main topic of this document?"]

# Process inputs
batch_images = processor.process_images(images).to(model.device)
batch_queries = processor.process_queries(queries).to(model.device)

# Get embeddings
with torch.no_grad():
    image_embeddings = model(**batch_images)
    query_embeddings = model(**batch_queries)

# Calculate similarity scores
scores = processor.score(query_embeddings, image_embeddings)
print(f"Similarity scores: {scores}")

ColMinistral3 Example

import torch
from PIL import Image
from sauerkrautlm_colpali.models import ColMinistral3, ColMinistral3Processor

model_name = "VAGOsolutions/SauerkrautLM-ColMinistral3-3b-v0.1"

model = ColMinistral3.from_pretrained(model_name)
model = model.to(dtype=torch.bfloat16, device="cuda:0").eval()

processor = ColMinistral3Processor.from_pretrained(model_name)

# Same usage pattern as ColQwen3...

ColLFM2 Example

import torch
from PIL import Image
from sauerkrautlm_colpali.models import ColLFM2, ColLFM2Processor

model_name = "VAGOsolutions/SauerkrautLM-ColLFM2-450M-v0.1"

model = ColLFM2.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="cuda:0",
).eval()

processor = ColLFM2Processor.from_pretrained(model_name)

# Same usage pattern as ColQwen3...

MTEB Integration

This package includes MTEB (Massive Text Embedding Benchmark) integration for standardized evaluation on the ViDoRe benchmark. See the mteb_integration/ folder for details.

# Example: Run MTEB evaluation
import mteb

model = mteb.get_model("VAGOsolutions/SauerkrautLM-ColQwen3-2b-v0.1")
tasks = mteb.get_tasks(tasks=["VidoreArxivQARetrieval"])
evaluation = mteb.MTEB(tasks=tasks)
results = evaluation.run(model, output_folder="results/")

Architecture

All models in this package follow the ColPali architecture:

1. Vision Encoder: Extracts patch embeddings from document images
2. Language Model: Processes visual tokens alongside text tokens
3. Projection Layer: Maps hidden states to 128-dimensional embedding space
4. Late Interaction: MaxSim scoring between query and document embeddings

Document Image → Vision Encoder → Visual Tokens → LLM → Projection → Multi-Vector Embeddings
                                                                              ↓
Query Text → Tokenizer → LLM → Projection → Multi-Vector Embeddings → MaxSim Score

Original ColPali

This package is based on the excellent work by Illuin Technology. For the original ColPali models (ColPali, ColQwen2, ColQwen2.5, ColSmol), please use the original colpali-engine package.

Citation

If you use this package, please cite both the original ColPali paper and our work:

@misc{sauerkrautlm-colpali-2025,
  title={SauerkrautLM-ColPali: Multi-Vector Vision Retrieval Models},
  author={David Golchinfar},
  organization={VAGO Solutions},
  year={2025},
  url={https://github.com/VAGOsolutions/sauerkrautlm-colpali}
}

@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}, 
}

📊 Benchmark Visualizations

ViDoRe v1 Benchmark (128-dim Models)

MTEB v1+v2 Benchmark (128-dim Models)

ViDoRe v3 Benchmark (128-dim Models)

Our Models vs High-dim Models

📋 Summary Tables

128-dim Models Comparison

Comparison vs High-dim Models

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contact

• VAGO Solutions: https://vago-solutions.ai
• GitHub: https://github.com/VAGOsolutions
• Email: info@vago-solutions.ai