Shopify builds Vision Language Model product classification system processing 30M daily predictions
Shopify's early product classification systems struggled with increasing product complexity and diversity on the platform. By early 2023, the team identified unmet requirements including more granular product understanding, consistent taxonomy, attribute extraction, richer metadata, and content safety features that category-only classification could not satisfy.
The initial 2018 logistic regression with TF-IDF classifier was effective only for simple cases, and the 2020 multi-modal system still fell short because classifying categories alone was insufficient for comprehensive product understanding.
The VLM-based system achieves an 85% merchant acceptance rate for predicted categories, processes over 30 million predictions daily, and has doubled hierarchical precision and recall compared to the earlier neural network approach.
Frequently asked questions
What did this team achieve with this AI workflow?
The VLM-based system achieves an 85% merchant acceptance rate for predicted categories, processes over 30 million predictions daily, and has doubled hierarchical precision and recall compared to the earlier neural net…
What tools did this team use?
Vision Language Models, LlaVA 1.5 7B, LLaMA 3.2 11B, Qwen2VL 7B, Nvidia Dynamo, Kubernetes, Dataflow.
What results were reported?
Merchant acceptance rate of predicted categories: 85%; Daily predictions processed: over 30 million predictions daily; Hierarchical precision and recall vs earlier neural network: doubled compared to our earlier neural network approach; Historical products in system: billions of historical products (source-reported, not independently verified).
What failed first in this deployment?
The initial 2018 logistic regression with TF-IDF classifier was effective only for simple cases, and the 2020 multi-modal system still fell short because classifying categories alone was insufficient for comprehensive…
How is this ecommerce ops AI workflow structured?
Dynamic product request batching → Input validation → VLM category prediction → VLM attribute prediction → Consistency check and retry → Multi-LLM training annotation → Human review of edge cases → Output storage and notification → Continuous quality feedback.