Nemotron ColEmbed V2: Top-Performing Late Interaction embedding models for Visual Document Retrieval

2026-02-03

AI Summary

This work addresses the limitations of visual document retrieval in RAG systems, which suffer from reliance on OCR and insufficient preservation of visual information in dense retrieval. To overcome these challenges, the authors propose a Late Interaction embedding approach based on pretrained vision-language models. By integrating cluster-based sampling, hard negative mining, bidirectional attention mechanisms, and model fusion strategies, the method achieves substantial performance gains on the ViDoRe benchmark. Leveraging NVIDIA Eagle 2 and Qwen3-VL as backbone architectures, the authors release three model variants (3B, 4B, and 8B), with the 8B version achieving state-of-the-art results on the ViDoRe V3 leaderboard, attaining an average NDCG@10 of 63.42. The study also investigates low-dimensional embeddings to balance storage efficiency and retrieval accuracy.

Technology Category

• Computer Vision: Image and Video Retrieval
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Retrieval-Augmented Generation (RAG) systems have been popular for generative applications, powering language models by injecting external knowledge. Companies have been trying to leverage their large catalog of documents (e.g. PDFs, presentation slides) in such RAG pipelines, whose first step is the retrieval component. Dense retrieval has been a popular approach, where embedding models are used to generate a dense representation of the user query that is closer to relevant content embeddings. More recently, VLM-based embedding models have become popular for visual document retrieval, as they preserve visual information and simplify the indexing pipeline compared to OCR text extraction. Motivated by the growing demand for visual document retrieval, we introduce Nemotron ColEmbed V2, a family of models that achieve state-of-the-art performance on the ViDoRe benchmarks. We release three variants - with 3B, 4B, and 8B parameters - based on pre-trained VLMs: NVIDIA Eagle 2 with Llama 3.2 3B backbone, Qwen3-VL-4B-Instruct and Qwen3-VL-8B-Instruct, respectively. The 8B model ranks first on the ViDoRe V3 leaderboard as of February 03, 2026, achieving an average NDCG@10 of 63.42. We describe the main techniques used across data processing, training, and post-training - such as cluster-based sampling, hard-negative mining, bidirectional attention, late interaction, and model merging - that helped us build our top-performing models. We also discuss compute and storage engineering challenges posed by the late interaction mechanism and present experiments on how to balance accuracy and storage with lower dimension embeddings.

Problem

• Visual Document Retrieval
• Retrieval-Augmented Generation
• Dense Retrieval
• Embedding Models
• Late Interaction

Innovation

• late interaction
• visual document retrieval
• vision-language model
• hard-negative mining
• embedding compression