Towards Mixed-Modal Retrieval for Universal Retrieval-Augmented Generation

TL;DR

URAG addresses retrieval for mixed-modal queries and documents to improve vision-language generation. The authors present Nyx, a unified mixed-modal retriever trained via a four-stage NyxQA data pipeline and a two-stage training regime that combines Matryoshka Representation Learning with VLM-guided feedback. Nyx achieves strong results on both traditional text-only RAG benchmarks and more general URAG tasks, outperforming strong baselines and demonstrating better alignment with downstream generation. NyxQA provides a realistic, diverse URAG benchmark, and the approach facilitates scalable, modality-agnostic retrieval for real-world multimodal information needs.

Abstract

Retrieval-Augmented Generation (RAG) has emerged as a powerful paradigm for enhancing large language models (LLMs) by retrieving relevant documents from an external corpus. However, existing RAG systems primarily focus on unimodal text documents, and often fall short in real-world scenarios where both queries and documents may contain mixed modalities (such as text and images). In this paper, we address the challenge of Universal Retrieval-Augmented Generation (URAG), which involves retrieving and reasoning over mixed-modal information to improve vision-language generation. To this end, we propose Nyx, a unified mixed-modal to mixed-modal retriever tailored for URAG scenarios. To mitigate the scarcity of realistic mixed-modal data, we introduce a four-stage automated pipeline for generation and filtering, leveraging web documents to construct NyxQA, a dataset comprising diverse mixed-modal question-answer pairs that better reflect real-world information needs. Building on this high-quality dataset, we adopt a two-stage training framework for Nyx: we first perform pre-training on NyxQA along with a variety of open-source retrieval datasets, followed by supervised fine-tuning using feedback from downstream vision-language models (VLMs) to align retrieval outputs with generative preferences. Experimental results demonstrate that Nyx not only performs competitively on standard text-only RAG benchmarks, but also excels in the more general and realistic URAG setting, significantly improving generation quality in vision-language tasks.

Figures (7)

• Figure 1: An illustration of the input patterns of "mixed-modal" content in the URAG scenario.
• Figure 2: The proposed four-step automated NyxQA construction pipeline.
• Figure 3: Overview of the Nyx architecture and its training paradigm.
• Figure 4: Impact of training data scale on NyxQA accuracy when training Nyx with varying sample sizes.
• Figure 5: Impact of (a) the number of in-context documents and (b) feedback-based retriever fine-tuning on downstream generation performance. Results are shown on NyxQA using InternVL3 models of varying sizes, respectively.