Modernizing the Facebook Groups Search to Unlock the Power of Community Knowledge

Meta engineering team describes how they modernized Facebook Groups Search by replacing keyword-only retrieval with a hybrid architecture combining Facebook's Unicorn inverted index and a 200M-parameter dense embedding model (SSR) with Faiss ANN search. The system uses a Multi-Task Multi-Label ranking model optimizing for clicks, shares, and comments simultaneously. Automated offline evaluation uses Llama 3 as a judge to grade search results at scale, replacing manual labeling. The new system improves search engagement and relevance without increasing error rates. Future plans include applying LLMs directly in ranking and adaptive retrieval strategies.