Researchers have developed SMART, a novel retrieval system designed to optimize Dynamic Product Ads (DPAs) by intelligently combining keyword-based and LLM-based search methods. This approach addresses the high costs and lexical mismatch issues associated with using LLMs for large-scale product catalog retrieval. SMART adaptively routes users to LLM-powered semantic prospecting only when initial keyword searches show gaps, significantly reducing LLM inference costs while maintaining retargeting performance. In a live A/B test at Snap, SMART demonstrated a substantial improvement in ad conversion rates. AI
IMPACT This hybrid retrieval system could significantly reduce the cost of deploying LLMs for large-scale recommendation and advertising systems.
RANK_REASON Publication of a research paper detailing a new system and its performance metrics. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
