Snap's SMART boosts ad conversions by 27.6% with hybrid LLM retrieval
New hybrid retrieval system cuts LLM costs by 90% while boosting ad conversions
Snap’s engineering team has developed SMART (SeMantic-aware Adaptive ReTrieval), a hybrid retrieval system designed to optimize dynamic product ads. The system addresses a core challenge in ad targeting: balancing retargeting (re-surfacing known user interests) and prospecting (discovering new categories). While LLMs excel at semantic intent, their deployment is costly and prone to lexical mismatches.
SMART combines two retrieval approaches: rule-generated queries for retargeting on a lexical BM25 index and LLM-generated queries for prospecting on a dense ANN index. A lightweight quality gate identifies users who would benefit most from semantic prospecting, routing only ~10% of users to the LLM path. This gated approach captures most of the semantic prospecting gains while reducing LLM inference costs by 90%. In a two-week online A/B test at Snap, SMART improved ad conversion rates by +27.6% over a strong embedding-based baseline.
- SMART reduces LLM inference costs by 90% by routing only 10% of users to LLM-generated queries
- Hybrid system combines lexical BM25 for retargeting and dense ANN for prospecting, improving ad conversion rates by +27.6%
- Published in ACM Conference on Recommender Systems (recsys'26), validated in a two-week Snap A/B test
Why It Matters
Revolutionizes ad targeting by cutting LLM costs while boosting conversion rates, proving hybrid retrieval can scale for real-world applications.