MUVERA

Overview

MUVERA (MUlti-VEctor Retrieval Algorithm) is a retrieval mechanism which reduces multi-vector similarity search to single-vector similarity search. MUVERA's innovation is to take whole groups of multi-vectors and compress them into a single, easier-to-handle vector called a Fixed Dimensional Encoding (FDE).

Key Innovation

MUVERA asymmetrically generates Fixed Dimensional Encodings (FDEs) of queries and documents, which are vectors whose inner product approximates multi-vector similarity. FDEs give high-quality ε-approximations, thus providing the first single-vector proxy for multi-vector similarity with theoretical guarantees.

Performance Benefits

Retrieval Efficiency

• FDEs achieve the same recall as prior state-of-the-art heuristics while retrieving 2-5× fewer candidates
• Achieves an average of 10% improved recall with 90% lower latency
• Consistently good end-to-end recall and latency across diverse BEIR retrieval datasets

Data-Oblivious Design

A key advantage of MUVERA is that the FDE transformation is data-oblivious, meaning it doesn't depend on the specific dataset. This makes it:

• Robust to changes in data distribution
• Suitable for streaming applications
• Adaptable to new data without retraining

Theoretical Foundations

MUVERA provides theoretical guarantees for approximation quality, making it the first single-vector proxy for multi-vector similarity with such guarantees.

Production Deployment

The research was presented at NeurIPS 2024 and has been implemented in production systems. Weaviate added MUVERA support in version 1.31, demonstrating its practical value for real-world vector search applications.

Use Cases

• Large-scale semantic search requiring efficiency
• Production systems needing reduced latency
• Applications with streaming data requirements
• Systems requiring both high recall and low latency

Pricing

Free research from Google, with open implementations available.