TechPrufer
Get started
Instacart
Instacart logoInstacart · Feature trail

Semantic IDs for Product Understanding

2 posts · Jun 2026

What this trail covers

Instacart has rebuilt its ads retrieval system by moving from a scoring-based approach to a generative model that spells out product recommendations token by token. This new system leverages Instacart Semantic IDs (SIDs), a compressed product representation generated by an RQ-VAE, which allows for better coverage of the entire catalog, including new products, and enables the model to generalize based on semantic codewords rather than just memorizing specific product co-occurrences. The training process uses a contrastive loss regularized by catalog structure and hierarchical batch sampling. The system also incorporates a two-flavor strategy for precision (ESCI) and discovery (ESCI+Gemma) recommendations.

All Instacart trails
  1. From Scoring to Spelling: Rebuilding Ads Retrieval at InstacartThis post details the transition from a BERT-based sequence model for ads retrieval that scored products to a generative model that spells out product recommendations token by token. The core technical innovation is the introduction of Instacart Semantic IDs (SIDs), generated by an RQ-VAE, which compress the product vocabulary by 125x and allow for better generalization and coverage of new products. The training data construction was also updated to use a 'Context Template' which enriches historical shopping sessions with new context tokens, moving from memorizing co-occurrences to generalizing based on semantic concepts. This enables better contextual matching and discovery of products, including those with zero transaction history.
  2. Semantic IDs: Product Understanding at ScaleThis post details the implementation of Semantic IDs (SIDs) at Instacart, a system that compresses product embeddings into discrete codes using a Residual Vector Quantizer (RQ-VAE). It explains the contrastive training approach that leverages catalog structure for supervision, the hierarchical batch sampling strategy to ensure diverse training data, and the loss function combining reconstruction and contrastive terms. The post also introduces a two-flavor strategy for generating SIDs: ESCI for precision (substitution, search, reordering) and ESCI+Gemma for discovery, utilizing Gemini Flash for broader product understanding.