Multimodal embeddings beyond a single vector

Back to Devoured - October 08, 2026

DEVOURED

Multimodal embeddings beyond a single vector

Perplexity launched PPLX-EMBED-V2-LATE, a multi-vector embedding model that aims to improve multimodal retrieval by retaining token-level data.

What: The new 0.6B and 9B parameter embedding models utilize multi-vector representations to enhance performance on benchmarks like ViDoRe(V3) by enabling richer semantic retrieval across text and images.

Why it matters: Transitioning from single-vector to multi-vector embeddings is a critical step for improving retrieval accuracy in RAG systems, especially when dealing with complex, multimodal documents.

Decoder

Embedding: A numerical representation of data (like text or images) in a high-dimensional vector space where similar concepts are mathematically closer together.
ViDoRe(V3): A benchmark suite used to evaluate how well document retrieval systems perform across different formats and layouts.

Original article

The Perplexity team launched PPLX-EMBED-V2-LATE, providing multi-vector, multimodal embeddings that enable richer retrieval across text and images. These models retain token-level vectors and utilize a shared embedding space, leading to improved retrieval on benchmarks like ViDoRe(V3). Available in two sizes, 0.6B and 9B, they offer industry-leading performance for various retrieval tasks while maintaining efficiency.

Explore all 59 articles in today's digest