How Yahoo enhances search retargeting using Amazon Bedrock

Yahoo DSP implemented Amazon Bedrock to modernize its Search Retargeting (SRT) keyword expansion. The legacy system used Word2Vec with locality-sensitive hashing (LSH), which suffered from outdated vocabulary and limited semantic understanding. The new generative AI workflow uses Amazon Bedrock to access multiple LLMs, including Amazon Nova, Meta Llama, and Anthropic's Claude. After evaluation, Anthropic's Claude 3.5 Sonnet v2 was selected for its balance of expansion ratio and semantic similarity. To reduce hallucinations, expanded keywords are verified by generating embeddings and comparing similarity scores against the original keywords; only those above a threshold are retained. The system also applies sensitive keyword filtering before and after inference. The enhanced solution launched in production in Q1 2025, achieving up to 600 times higher keyword expansion rates and up to 5 times growth in addressable audience reach. Yahoo used Amazon SageMaker Studio for prototyping and Amazon Bedrock for serverless access to foundation models, avoiding infrastructure management. Amazon Bedrock provides flexible inference options (batch and real-time) and straightforward model upgrades.