Towards Mixed-Modal Retrieval for Universal Retrieval-Augmented Generation

TL;DR

This paper introduces Nyx, a novel mixed-modal retriever designed for universal retrieval-augmented generation, addressing the challenge of retrieving and reasoning over combined text and image data to enhance vision-language generation.

Contribution

The paper presents Nyx, a unified mixed-modal retriever and a new dataset NyxQA, along with a two-stage training framework, advancing retrieval-augmented generation for mixed-modal information.

Findings

Nyx outperforms existing methods on vision-language tasks.

Nyx achieves competitive results on standard RAG benchmarks.

The dataset NyxQA reflects real-world mixed-modal 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.