Tevatron 2.0: Unified Document Retrieval Toolkit across Scale, Language, and Modality

TL;DR

The paper introduces Tevatron 2.0, a comprehensive toolkit for scalable, multilingual, and multimodal document retrieval, featuring a unified dense retriever and the OmniEmbed embedding model for diverse data types.

Contribution

It presents a unified retrieval pipeline supporting multiple scales, languages, and modalities, along with the novel OmniEmbed model unifying various data modalities for retrieval tasks.

Findings

Achieves strong multilingual and multimodal retrieval performance.

Demonstrates cross-modality zero-shot retrieval capabilities.

Provides a versatile toolkit bridging academia and industry.

Abstract

Recent advancements in large language models (LLMs) have driven interest in billion-scale retrieval models with strong generalization across retrieval tasks and languages. Additionally, progress in large vision-language models has created new opportunities for multimodal retrieval. In response, we have updated the Tevatron toolkit, introducing a unified pipeline that enables researchers to explore retriever models at different scales, across multiple languages, and with various modalities. This demo paper highlights the toolkit's key features, bridging academia and industry by supporting efficient training, inference, and evaluation of neural retrievers. We showcase a unified dense retriever achieving strong multilingual and multimodal effectiveness, and conduct a cross-modality zero-shot study to demonstrate its research potential. Alongside, we release OmniEmbed, to the best of our knowledge, the first embedding model that unifies text, image document, video, and audio retrieval, serving as a baseline for future research.