Catalog mechanics

How the catalog works

This page explains the three pieces people usually ask about first: pricing, image generation, and product generation. It is the short version of the system behind the store, written for demos and non-technical review.

Pricing

How item prices are produced and why they vary.

Broad price bands

Products are assigned to tiered price bands, then nudged inside that band with a deterministic seed. That gives us a real spread without making the numbers random from one refresh to the next.

Search intentwhat the user typed
Personanormal, luxury, bargain, minimalist
Price seedquery + item id

Deterministic per item

The same search and same item ID always produce the same price on the same build. That keeps demos stable while still making each product feel distinct.

Visible price variety

The goal is to avoid a shelf where every item clusters around the same amount. Premium, mid-range, and budget-like products should all appear naturally in the same search.

Image generation

How a query turns into a product illustration.

Query wordssignal terms from the search
Shape familyphone, audio, package, capsule, and more
Variant seeditem-specific visual offset

Shape families, not just colors

Images are generated as inline SVGs. The SVG picks a shape family from the item type and a second layout from the item seed, so cards differ in structure instead of only changing palette.

Keyword-aware routing

Queries like phone or headphones route into different silhouette families. The system also avoids accidental substring matches, so headphones do not look like phones.

Stable per item

Every product gets a deterministic SVG URL, so the same product stays visually consistent across cache hits while still looking different from its neighbors.

Product generation

How we turn a search phrase into a shelf.

Normal searchtypos are corrected against English
Intergalactic modekeeps the original wording
Persona + marketshapes copy, reviews, and ranking

Intent first

The generator starts with the user’s query, applies correction when needed, then picks a persona and product family. That yields titles, descriptions, reviews, and explanations that stay aligned with the search.

Generated once, reused often

The first request generates a full collection, stores it, and future requests reuse the same shelf until the cache version changes. That keeps demos fast and predictable.

Different voices, same engine

Normal mode sounds like a regular store. Intergalactic mode shifts the tone, reviews, and product naming, while still using the same generation pipeline underneath.