Meta's Hybrid AI Architecture Revolutionizes Facebook Groups Search for Community Knowledge
Meta has given a fresh look to Facebook Groups Search by introducing a new hybrid retrieval system that combines precise keyword searches with deeper understanding. This upgrade aims to improve how users find groups, cut down on endless scrolling, and ensure that reliable community knowledge is easy to access. The new framework uses two methods: the Unicorn inverted index for accurate matches and a sophisticated 12-layer semantic retriever model with 200 million parameters for more nuanced queries. A Multi-Task Multi-Label supermodel sorts results based on how much they engage users. Additionally, Llama 3 helps refine search results without increasing errors. There are plans to further integrate LLMs for better ranking and adaptable search methods. A related technical paper has been published on arXiv.
Key facts
- Meta implemented a hybrid retrieval architecture for Facebook Groups Search
- The system addresses discovery, consumption, and validation friction points
- Parallel retrieval uses Unicorn inverted index and a 200-million-parameter semantic model
- Results are ranked by an MTML supermodel optimizing for clicks, shares, and comments
- Automated evaluation uses Llama 3 as judge with nuanced relevance categories
- The deployment improved search engagement metrics without increasing error rates
- Future work includes LLMs in ranking and adaptive retrieval strategies
- A technical paper was published on arXiv detailing the approach
Entities
Institutions
- Meta
- Engineering at Meta
- arXiv