Retail / E-commerce platform
Shopify: Unifying and standardising product information
- Company
- Shopify
- Country
- Canada
- Adoption stage
- In operation
- Source published
- Date basis
- The date the source was published. It can differ from the date adoption started.
- How the source was checked
- Read the full source text
The work problem
Millions of merchants list products on Shopify, each in their own way. Most product information arrives as free-form sentences rather than fixed fields, and each merchant names attributes and values themselves, so the structure varies. Typos, empty values and wrong categories are mixed in, and some information exists only in images rather than text. As a result semantic search and filters work poorly and the same product is scattered as duplicates.
Technology and data
Shopify built a product information layer called Global Catalogue. The first layer takes in more than 10 million product updates a day as a stream. The second layer fine-tunes a vision large language model so that classification, attribute extraction, image understanding, title standardisation, description summarisation and review summarisation are all handled by one model. The third layer groups cases where different merchants sell the same product, and the fourth layer merges the attributes, descriptions and images of a grouped product into a representative product record. Instead of a commercial API it went through LlaVA 1.5 7B and LLaMA 3.2 11B and now fine-tunes Qwen2VL 7B. Training material is produced by several LLM annotation agents, and for test material people decide the final labels.
Results
Shopify said its inference base handles 40 million LLM calls a day. It explained that by changing the training and prompting approach so only the needed fields are extracted, it cut median latency from 2 seconds to 500 milliseconds and lowered GPU usage by 40 percent. The company said this data is used for real-time suggestions in the merchant admin screen, for search and recommendations, and for conversational commerce such as Shopify Sidekick.
Limits and open questions
The 40 million a day is throughput, not a search result or a revenue improvement rate. The source does not disclose accuracy figures for classification or attribute extraction.
Sources
- Leveraging multimodal LLMs for Shopify’s global catalogue: Recap of expo talk at ICLR 2025 - Shopifyshopify.engineering, Accessed
Compiled from public sources. These are not results from ATF Works customers.