Instacart uses LLMs to understand users’ search intent and improve product retrieval through RAG, output validation, query rewriting, and structured tag extraction. Its system increased query rewrite coverage from approximately 50% to more than 95%, with precision exceeding 90% across all three types of rewriting. Its real-time SRL model uses LoRA fine-tuning on Llama-3 8B, while caching and automatic GPU scaling mean only 2% of queries require real-time inference. A/B testing showed a 6% reduction in average scrolling depth for tail queries and a 50% decrease in related complaints.
Source:ByteByteGo Newsletter · Read Original
Leave a Reply