Semantic cold start / 1m 28.12s

A new product finds its place.

A newly eligible product can receive useful product-to-product recommendations before it has accumulated purchase history.

Silent film with burned-in narrationOpen WebM directly

What You Will See

Prepare a new productCatalog content becomes complete, published and sellable.
Build semantic identityLocalized product text is converted into a local embedding.
Open the product pageThe new product has no need to wait for co-purchase history.
See similar productsAn eligible “Similar products” strip appears beneath the product.

Cold start should not make a new product isolated when its content already explains what it resembles.

Semantic similarity supplies a product-to-product strategy while business eligibility still filters the result.

Localized product textName, description, category and attributes

→

Local ONNX modelPinned quantized MiniLM artifact

→

pgvector similarityNearest eligible product embeddings

→

PDP similar stripSeed excluded, stock and diversity applied

The model runs locally. The database compares vectors; Recommendation applies serving policy before returning results.

A newly eligible product receives a visible similar-products strip.

SIMILAR_CONTENT, the explanation and operation without purchase history are source- and BDD-backed, not labels visible in the film.