HS Code Classifier

Step 1 of 4 — How it works

The ML pipeline

Your product description is converted into a 384-dimensional vector using a multilingual embedding model (multilingual-e5-small). A K-Nearest Neighbors classifier then finds the closest training examples in that embedding space and votes on the most likely HS code. The confidence score reflects how strongly the neighbors agree.

Visualization

The scatter plot uses UMAP (Uniform Manifold Approximation and Projection) to project the high-dimensional embeddings down to 2D, so you can see how product categories cluster together. Points that are close in the plot are semantically similar.

Step 2 of 4 — Classify

Describe a product

Type or paste any product description in the text box on the left. The model supports multiple languages including English, Thai, Vietnamese, and Chinese. You'll get up to 5 predictions ranked by confidence, plus the nearest training examples used to make the decision.

⌘/CtrlEnter to classify

Step 3 of 4 — Examples

Try the example chips

Click any of the example chips below the input to instantly classify pre-built product descriptions across four languages. This is the quickest way to see the model in action and understand what the results look like.

Step 4 of 4 — Explore

Explore the latent space

The UMAP visualization on the right shows all {{ metadata.get('n_examples_full_dataset', metadata.get('n_examples', '')) }} training examples projected into 2D. When you classify a product, your query appears as a diamond and the nearest neighbors light up. Filter by chapter or language, and click any point for details.

⌘/CtrlK to search points

Classification