Google Confirms Use of MUVERA-Like System

Google’s Gary Illyes recently confirmed that the company does make use of a system similar to Multi-Vector Retrieval via Fixed-Dimensional Encodings (MUVERA). However, he was less specific when asked about the use of Graph Foundation Models (GFMs).

These comments came during a Q&A session at the Search Central Live Deep Dive event held in Asia, where Illyes addressed questions on whether Google had adopted these advanced retrieval and modelling techniques.

 

MUVERA

Google has recently introduced MUVERA, a new retrieval method detailed in a blog post and academic paper. This approach simplifies the process of multi-vector search by converting it into a faster single-vector system. It does this by compressing token embeddings into fixed-length vectors that closely reflect their original relationships, allowing for quicker identification of potential matches. These are then re-ordered using more detailed comparisons to ensure accuracy.

MUVERA works by transforming sets of multiple vectors into Fixed Dimensional Encodings (FDEs), which effectively summarise the original data into a single vector form. These FDEs retain enough of the original detail to allow for reliable search results.

The system uses Maximum Inner Product Search (MIPS), a well-established method in retrieval, to efficiently scan for relevant items. Once a shortlist is created through this faster method, MUVERA refines the results using Chamfer similarity—a more precise, multi-vector comparison—to restore full accuracy.

Compared with previous retrieval systems like PLAID, MUVERA not only improves recall but also does so with fewer candidates and faster processing times. This makes it a strong choice for handling large-scale search tasks efficiently without sacrificing precision.

 

Google Confirms That They Use MUVERA

José Manuel Morgal (as seen on LinkedIn) brought up a question to Google’s Gary Illyes regarding MUVERA. In a light-hearted exchange, Illyes initially joked by asking what MUVERA was, before going on to confirm that Google uses a similar system—though not under the same name.

As José described the interaction:
“A research article and paper have been published by Google about MUVERA. I asked if it was currently in use in Search. Gary laughed and asked me what MUVERA was, then explained they do use a comparable method, just not referred to by that name internally.”

 

Does Google Use Graph Foundation Models (GFMs)?

Google has recently shared details of a new AI development known as the Graph Foundation Model (GFM) in a blog post.

This model is designed to understand and learn from relational databases by converting them into graphs—where each row is treated as a node and the relationships between tables are represented as edges. Unlike traditional machine learning models or graph neural networks (GNNs), which typically require training on each specific dataset, GFMs can process entirely new databases with varying structures and features without needing to be retrained.

GFMs work by leveraging a large-scale AI model to identify relationships across data tables, allowing them to detect complex patterns that other models might overlook. This has proven highly effective in areas such as spam detection across Google’s broader systems. With performance improvements ranging from three to forty times greater in average precision, GFMs mark a significant leap forward in handling structured data.

When José Manuel Morgal asked Google’s Gary Illyes whether GFMs were currently being used in Search, Illyes playfully responded by pretending not to know what GFMs were. José explained the interaction:
“I mentioned that Google Research had published an article about Graph Foundation Models for data, although there was no academic paper linked to it. I asked if this was being used in Search. Gary joked that he wasn’t familiar with the term and said he didn’t believe it was in production. He also pointed out that he doesn’t oversee what appears on the Google Research blog.”

According to Illyes, it appears that Graph Foundation Models are not currently implemented in Google Search—at least not to his knowledge.

 

Is GFM Ready For Scaled Deployment?

Google’s official announcement on the Graph Foundation Model (GFM) highlights that it was trialled on a real internal use case—specifically for detecting spam within advertisements. This indicates the model wasn’t simply tested in a controlled lab setting or with synthetic data, but rather with live internal systems and real-world datasets.

According to Google, operating at their scale requires handling vast graphs containing billions of nodes and edges. Their advanced JAX framework and scalable TPU infrastructure are particularly effective in managing this volume of data. The GFM was evaluated across various internal classification tasks, such as spam detection in advertising, which involved working with numerous large, interconnected relational tables.

Google noted that conventional tabular models, while scalable, often ignore the relationships between rows across different tables—missing crucial context that could lead to more accurate predictions. Their experiments with the GFM clearly demonstrated the model’s ability to close this gap and deliver better results.