Operator note

Google confirms MUVERA-style retrieval in Search - what it means for rankings

At Search Central Live, Gary Illyes reveals Google's live MUVERA-style retrieval while ruling out Graph Foundation Models. Learn the ranking impact.

Minimalist illustration of Google search shift with magnifying glass funneling result tiles and puzzled analyst reacting

At Google’s Search Central Live Deep Dive Asia on 25 June 2024, Search Analyst Gary Illyes told attendees that Google already runs a production retrieval system “similar to MUVERA.” He added that, to his knowledge, Graph Foundation Models (GFM) are not yet part of Google Search. Illyes made the remarks during the live Q&A, which is summarized below and in a related LinkedIn post.

Illyes briefly joked about the MUVERA name before confirming that Google’s search stack uses a comparable approach under a different internal label. He did not share rollout dates or specify which Search features rely on the method.

Key details from the event

  • Venue: Search Central Live Deep Dive Asia, streamed 25 June 2024
  • Speaker: Gary Illyes, Analyst, Google
  • Statement: Google employs a MUVERA-style retrieval technique in production
  • Naming: The internal system has a different name
  • Follow-up: No timeline or product list provided

How MUVERA works

Google Research introduced Multi-Vector Retrieval via Fixed-Dimensional Encodings (MUVERA) on 11 June 2024. The technique compresses sets of token embeddings into single fixed-length vectors, enabling fast Maximum Inner Product Search. A secondary Chamfer similarity stage re-ranks the results, regaining multi-vector accuracy while using less compute than the PLAID baseline.

In internal tests, MUVERA achieved higher recall with fewer retrieved documents. Full details are available in the Google Research MUVERA announcement.

Graph Foundation Models remain experimental

When asked whether Graph Foundation Models power Search, Illyes said he believed they were “not in production” and noted he does not review every research post. Google unveiled GFM on 6 June 2024 as an architecture that converts relational tables into graphs and learns general patterns across them. Early internal tests showed three- to forty-fold precision gains in certain ad-spam tasks, but Google has not announced broader use. See the Google Research GFM announcement for technical specifics.

Source citations

  • Search Central Live Deep Dive Asia, Q&A session, 25 June 2024
  • Google Research. “MUVERA: Multi-Vector Retrieval via Fixed-Dimensional Encodings,” 11 June 2024
  • Google Research. “Graph Foundation Models: Unlocking Graph-Based Learning for Relational Databases,” 6 June 2024

Keep reading

Related articles

AI powered shopping cart protocol illustration with funnel price tag alert loyalty user tapping toggleInside Google's Universal Commerce Protocol that lets AI agents tap carts, catalogs and loyalty pricing2 min readMinimalist illustration of AI checkout hub with Cart Catalog Identity cards and user tapping settingsGoogle quietly upgrades AI shopping protocol: what Cart, Catalog and Identity Linking change next2 min readMinimalist tablet health UI privacy risk toggle character adjusting shield and prescription funnelGoogle and DocMorris Launch AI Health Companion for Europe - What Changes Next2 min readMinimalist site health dashboard illustration with 404 410 toggle funnel filtering errors into green checksWorried About Endless 404 Reports In Search Console? John Mueller Reveals What They Really Mean3 min read