Problem
In a multi-stage ads recommender, the retrieval stage cuts tens of millions of ad candidates to a few thousand for the ranking models. It sees about three orders of magnitude more ads than later stages and has a tight latency budget. Automated campaign tools and generative AI creative tools keep adding ads, and the number of creatives is expected to grow significantly. The older Meta retrieval system used isolated model stages and many rule-based heuristics, which allowed only limited personalization and blocked end-to-end optimization.
Example from the article: in one month more than a million advertisers used Meta's generative AI tools to create more than 15 million ads.
Solution in brief
Andromeda replaces the staged, rule-heavy retrieval with one deep neural network over a hierarchical ad index. The index and the model are trained jointly, so a request visits only the most relevant nodes and inference cost grows sub-linearly with the number of ads. That headroom is spent on a much larger model.
Questions this case answers
- How to scale the candidate-generation stage of a multi-stage recommender when the candidate pool grows rapidly and latency is fixed
- How to increase retrieval model capacity while keeping inference cost sub-linear in the number of candidates
- How to keep ads retrieval fast and accurate when the number of candidate ads grows several times over?
- How to add model capacity to the candidate-retrieval stage without a matching rise in serving cost?
Tools and methods
- Hierarchical ad index — multi-layer index in which each request visits only the most relevant nodes, trained together with the retrieval model.
- Two-stage recommender — retrieval narrows the pool, then larger ranking models pick the final ads.
- Model elasticity — segment-aware model complexity that adapts to available resources.
Google or a one-time email code. No password.