Even Oldridge

Gabriel de Souza P. Moreira, Radek Osmulski, Mengyao Xu et al.

Text embedding models have been popular for information retrieval applications such as semantic search and Question-Answering systems based on Retrieval-Augmented Generation (RAG). Those models are typically Transformer models that are fine-tuned with contrastive learning objectives. One of the challenging aspects of fine-tuning embedding models is the selection of high quality hard-negative passages for contrastive learning. In this paper we introduce a family of positive-aware mining methods that use the positive relevance score as an anchor for effective false negative removal, leading to faster training and more accurate retrieval models. We provide an ablation study on hard-negative mining methods over their configurations, exploring different teacher and base models. We further demonstrate the efficacy of our proposed mining methods at scale with the NV-Retriever-v1 model, which scores 60.9 on MTEB Retrieval (BEIR) benchmark and placed 1st when it was published to the MTEB Retrieval on July, 2024.

Zhenrui Yue, Sara Rabhi, Gabriel de Souza Pereira Moreira et al.

Recently, large language models (LLMs) have exhibited significant progress in language understanding and generation. By leveraging textual features, customized LLMs are also applied for recommendation and demonstrate improvements across diverse recommendation scenarios. Yet the majority of existing methods perform training-free recommendation that heavily relies on pretrained knowledge (e.g., movie recommendation). In addition, inference on LLMs is slow due to autoregressive generation, rendering existing methods less effective for real-time recommendation. As such, we propose a two-stage framework using large language models for ranking-based recommendation (LlamaRec). In particular, we use small-scale sequential recommenders to retrieve candidates based on the user interaction history. Then, both history and retrieved items are fed to the LLM in text via a carefully designed prompt template. Instead of generating next-item titles, we adopt a verbalizer-based approach that transforms output logits into probability distributions over the candidate items. Therefore, the proposed LlamaRec can efficiently rank items without generating long text. To validate the effectiveness of the proposed framework, we compare against state-of-the-art baseline methods on benchmark datasets. Our experimental results demonstrate the performance of LlamaRec, which consistently achieves superior performance in both recommendation performance and efficiency.

Haidong Rong, Jiashu Yao, Matthias Langer et al.

Traditional GPU hash tables preserve every inserted key -- a dictionary assumption that wastes scarce High Bandwidth Memory (HBM) when embedding tables routinely exceed single-GPU capacity. We challenge this assumption with cache semantics, where policy-driven eviction is a first-class operation. We introduce HierarchicalKV (HKV), the first general-purpose GPU hash table library whose normal full-capacity operating contract is cache-semantic: each full-bucket upsert (update-or-insert) is resolved in place by eviction or admission rejection rather than by rehashing or capacity-induced failure. HKV co-designs four core mechanisms -- cache-line-aligned buckets, in-line score-driven upsert, score-based dynamic dual-bucket selection, and triple-group concurrency -- and uses tiered key-value separation as a scaling enabler beyond HBM. On an NVIDIA H100 NVL GPU, HKV achieves up to 3.9 billion key-value pairs per second (B-KV/s) find throughput, stable across load factors 0.50-1.00 (<5% variation), and delivers 1.4x higher find throughput than WarpCore (the strongest dictionary-semantic GPU baseline at lambda=0.50) and up to 2.6-9.4x over indirection-based GPU baselines. Since its open-source release in October 2022, HKV has been integrated into multiple open-source recommendation frameworks.

Yauhen Babakhin, Radek Osmulski, Ronay Ak et al.

We introduce llama-embed-nemotron-8b, an open-weights text embedding model that achieves state-of-the-art performance on the Multilingual Massive Text Embedding Benchmark (MMTEB) leaderboard as of October 21, 2025. While recent models show strong performance, their training data or methodologies are often not fully disclosed. We aim to address this by developing a fully open-source model, publicly releasing its weights and detailed ablation studies, and planning to share the curated training datasets. Our model demonstrates superior performance across all major embedding tasks -- including retrieval, classification and semantic textual similarity (STS) -- and excels in challenging multilingual scenarios, such as low-resource languages and cross-lingual setups. This state-of-the-art performance is driven by a novel data mix of 16.1 million query-document pairs, split between 7.7 million samples from public datasets and 8.4 million synthetically generated examples from various open-weight LLMs. One of our key contributions is a detailed ablation study analyzing core design choices, including a comparison of contrastive loss implementations, an evaluation of synthetic data generation (SDG) strategies, and the impact of model merging. The llama-embed-nemotron-8b is an instruction-aware model, supporting user-defined instructions to enhance performance for specific use-cases. This combination of top-tier performance, broad applicability, and user-driven flexibility enables it to serve as a universal text embedding solution.

Gabriel de Souza P. Moreira, Ronay Ak, Benedikt Schifferer et al.

Ranking models play a crucial role in enhancing overall accuracy of text retrieval systems. These multi-stage systems typically utilize either dense embedding models or sparse lexical indices to retrieve relevant passages based on a given query, followed by ranking models that refine the ordering of the candidate passages by its relevance to the query. This paper benchmarks various publicly available ranking models and examines their impact on ranking accuracy. We focus on text retrieval for question-answering tasks, a common use case for Retrieval-Augmented Generation systems. Our evaluation benchmarks include models some of which are commercially viable for industrial applications. We introduce a state-of-the-art ranking model, NV-RerankQA-Mistral-4B-v3, which achieves a significant accuracy increase of ~14% compared to pipelines with other rerankers. We also provide an ablation study comparing the fine-tuning of ranking models with different sizes, losses and self-attention mechanisms. Finally, we discuss challenges of text retrieval pipelines with ranking models in real-world industry applications, in particular the trade-offs among model size, ranking accuracy and system requirements like indexing and serving latency / throughput.

Gabriel de Souza P. Moreira, Ronay Ak, Mengyao Xu et al.

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.

Mengyao Xu, Gabriel Moreira, Ronay Ak et al.

Motivated by the growing demand for retrieval systems that operate across modalities, we introduce llama-nemoretriever-colembed, a unified text-image retrieval model that delivers state-of-the-art performance across multiple benchmarks. We release two model variants, 1B and 3B. The 3B model achieves state of the art performance, scoring NDCG@5 91.0 on ViDoRe V1 and 63.5 on ViDoRe V2, placing first on both leaderboards as of June 27, 2025. Our approach leverages the NVIDIA Eagle2 Vision-Language model (VLM), modifies its architecture by replacing causal attention with bidirectional attention, and integrates a ColBERT-style late interaction mechanism to enable fine-grained multimodal retrieval in a shared embedding space. While this mechanism delivers superior retrieval accuracy, it introduces trade-offs in storage and efficiency. We provide a comprehensive analysis of these trade-offs. Additionally, we adopt a two-stage training strategy to enhance the model's retrieval capabilities.

Mengyao Xu, Wenfei Zhou, Yauhen Babakhin et al.

We present Omni-Embed-Nemotron, a unified multimodal retrieval embedding model developed to handle the increasing complexity of real-world information needs. While Retrieval-Augmented Generation (RAG) has significantly advanced language models by incorporating external knowledge, existing text-based retrievers rely on clean, structured input and struggle with the visually and semantically rich content found in real-world documents such as PDFs, slides, or videos. Recent work such as ColPali has shown that preserving document layout using image-based representations can improve retrieval quality. Building on this, and inspired by the capabilities of recent multimodal models such as Qwen2.5-Omni, we extend retrieval beyond text and images to also support audio and video modalities. Omni-Embed-Nemotron enables both cross-modal (e.g., text - video) and joint-modal (e.g., text - video+audio) retrieval using a single model. We describe the architecture, training setup, and evaluation results of Omni-Embed-Nemotron, and demonstrate its effectiveness in text, image, and video retrieval.

Gabriel de Souza P. Moreira, Sara Rabhi, Ronay Ak et al.

Session-based recommendation is an important task for e-commerce services, where a large number of users browse anonymously or may have very distinct interests for different sessions. In this paper we present one of the winning solutions for the Recommendation task of the SIGIR 2021 Workshop on E-commerce Data Challenge. Our solution was inspired by NLP techniques and consists of an ensemble of two Transformer architectures - Transformer-XL and XLNet - trained with autoregressive and autoencoding approaches. To leverage most of the rich dataset made available for the competition, we describe how we prepared multi-model features by combining tabular events with textual and image vectors. We also present a model prediction analysis to better understand the effectiveness of our architectures for the session-based recommendation.