Skip to content

SYSTEM Cited by 1 source

SSR — Search Semantic Retriever (Meta)

SSR (Search Semantic Retriever) is the 12-layer, 200-million-parameter encoder model Meta uses to convert natural-language user queries into dense vector representations for the semantic-retrieval arm of Facebook Groups Scoped Search. Disclosed in the 2026-04-21 Meta Engineering post.

Role in the retrieval pipeline

  1. Pre-processed query arrives at SSR (post-tokenization/normalization/rewriting).
  2. SSR encodes the query into a dense vector in the same embedding space as the precomputed group-post index.
  3. Faiss performs approximate nearest neighbor search over the precomputed vector index of group posts.
  4. Semantic candidates flow into the L2 MTML ranker alongside Unicorn's lexical candidates, with cosine similarity scores as features.

The key contribution of SSR is resolving the natural-language-intent gap in keyword search:

"We needed a system where searching for an 'Italian coffee drink' effectively matches a post about 'cappuccino,' even if the word 'coffee' is never explicitly stated."

Model details

  • Layers: 12
  • Parameters: ~200 million
  • Input: pre-processed user query (tokenized, normalized, rewritten)
  • Output: dense vector in the group-post embedding space
  • Training data, loss function, backbone family — not disclosed in the post

Relation to the broader Meta embedding-model portfolio

Not the same as EnCodec (audio), MTIA-served recsys embeddings (systems/meta-adaptive-ranking-model), or Llama — SSR is a purpose-built retrieval-encoder for scoped community search. Serves a narrower task: query → single dense vector for ANN lookup, not generation or ranking.

Seen in

Last updated · 550 distilled / 1,221 read