Modernizing Facebook Groups Search to Unlock the Power of Community Knowledge

Facebook Groups search was redesigned to solve three problems: discovery, consumption, and verification. The existing keyword-centric search missed results when users' natural language intent diverged from the actual wording of posts, required users to manually read through many comments even after finding results, and required manually gathering scattered signals to use community knowledge for decision-making.

The search pipeline was changed to a hybrid retrieval architecture. A query first goes through tokenization, normalization, and rewriting, then is processed in parallel via two paths.

• Lexical path: Facebook's Unicorn inverted index quickly finds exact matching terms.
• Semantic path: SSR (a 12-layer, 200-million-parameter model) converts the query into a dense vector, and Faiss-based ANN search finds semantically close posts.
• Candidates from both paths are merged at the ranking stage and sorted using sparse/dense features together, such as TF-IDF, BM25, and cosine similarity.

The ranking model was changed from a single objective to an MTML (Multi-Task Multi-Label) structure. It jointly optimizes for clicks, shares, and comments, placing results that are actually likely to drive engagement ahead of those that merely appear relevant.

The verification method was also automated. Instead of human labeling, Llama 3-based automated evaluation was incorporated into build verification tests (BVT) to score search results. Rather than simple good/bad judgments, this evaluation system distinguishes intermediate levels such as somewhat relevant, reflecting similarity within the same domain or topic more precisely.

As a result of the rollout, search engagement and relevance improved, and the error rate did not increase. Going forward, the plan is to explore using LLMs directly at the ranking stage with LLMs in Ranking, as well as adaptive retrieval, which adjusts retrieval parameters based on query complexity.