Blogs›Shopify›Product Categorization System
Shopify has developed and enhanced its product categorization system by introducing an AI multi-agent system to continuously evolve the taxonomy. This system goes beyond static classification by actively improving taxonomy labels themselves, keeping the system agile and future-proof. It integrates actual merchant product data for real-world grounding, employs specialized agents for structural consistency and product-driven insights, synthesizes these insights intelligently, and detects complex e. This post details the use of multimodal LLMs to organize and augment product data across Shopify, including data curation, model fine-tuning, training, and the infrastructure for large-scale multimodal LLM inferences, impacting search, recommendations, and conversational commerce.
6/17/2026
This post details the application of LLMs and ANN retrieval for intra-store product clustering, focusing on a 'core value proposition' framework to distinguish product variants from distinct products. It also outlines a two-stage LLM pipeline with pre-chunking to handle large catalogs and a singleton detector to optimize the process by identifying deterministically clustered products.
10/9/2025
This post details the development of an AI multi-agent system for evolving Shopify's product taxonomy. The system employs specialized agents for structural analysis and product-driven analysis, which are then intelligently synthesized. A key innovation is the detection of complex equivalence relationships where specific categories equal broader categories filtered by attribute values. Automated quality assurance is performed by specialized AI judges. The system integrates merchant product data for grounding and aims to keep the taxonomy agile and future-proof.
7/16/2025
This post details the engineering behind Shopify's Global Catalogue initiative, which uses multimodal Large Language Models (LLMs) to organize and augment product data. It covers data curation, model fine-tuning (moving from LlaVA 1.5 7B to LLaMA 3.2 11B to Qwen2VL 7B), selective field extraction for training to improve generalization and reduce latency/GPU usage, automated annotation pipelines using LLM agents and human expertise, and the infrastructure for making 40 million multimodal LLM-powered inferences daily. The four integrated layers of the Global Catalogue are described: product data foundation, product understanding (classification, attribute extraction, image understanding, title standardization, description analysis, review summarization), product matching (candidate generation, discriminator models, graph-based clustering), and reconciliation (canonical product record construction).
5/8/2025
This post details the evolution of Shopify's product understanding system, moving from early basic classification with logistic regression and TF-IDF to a multi-modal approach combining image and text data. The current generation is built on Shopify's Standard Product Taxonomy and Vision Language Models (VLMs). It introduces the technical deep dive into inference optimization, including FP8 quantization, in-flight batching, and KV cache optimization. The pipeline architecture is described, highlighting a two-stage prediction process for categories and attributes using a Dataflow pipeline and a Kubernetes cluster with NVIDIA GPUs. The post also elaborates on the robust training data system, which uses a multi-LLM annotation system with arbitration and human validation. The impact is quantified with an 85% acceptance rate of predicted categories and doubled hierarchical precision and recall. Future directions include incorporating new VLM architectures, expanding attribute prediction, and migrating to a DAG taxonomy.
9/8/2021
This post details the evolution of Shopify's product categorization system by introducing a new model that incorporates rich image and text data. It describes the use of multi-lingual BERT for text embeddings and MobileNet-V2 for image embeddings. The model architecture is framed as a multi-task, multi-class classification problem, where each level of the Google Product Taxonomy (GPT) is a separate classification task, with outputs feeding into subsequent levels. Training is optimized using data parallelization on Google Cloud Platform. Inference logic is enhanced to enforce hierarchical consistency, ensuring accurate and reliable predictions for consumers. The new model achieved an 8% increase in leaf precision and doubled coverage.
4/30/2020
This post details the implementation of a product categorization system at Shopify. It addresses the challenges of classifying billions of products into a large, hierarchical taxonomy (Google Product Taxonomy). The solution involves text featurization using HashingTF and PySpark, opting for simplicity and scalability over more complex methods like Word2Vec. The core modeling approach uses Kesler's Construction to transform multi-class classification into a binary classification problem, allowing for a single Logistic Regression model. This approach scales to thousands of categories, leverages taxonomy structure, reduces computational resources, and maintains simplicity. The inference process uses a greedy traversal of the taxonomy to predict the most relevant category path. The post also touches upon evaluation metrics for hierarchical classification.