AI Commerce Development

Reference architecture

1. Catalogue & Inventory Feed
2. Enrichment + Embedding Pipeline
3. Vector + Attribute Index
4. Search / Ranking / Assistant API
5. Storefront Screens
6. Experiment & Revenue Tracking

A typical AI commerce system: catalogue and inventory data land in an enrichment and embedding pipeline that keeps a vector and attribute index in sync; search, ranking, and the shopping assistant read that index at request time and return results inside the storefront’s latency budget, while every impression and add-to-cart flows into experiment tracking so lift is measured against a holdout.

Frequently asked questions

• How do you make a conversational shopping assistant recommend only products we actually sell?+

  By grounding it in your catalogue instead of the model’s memory. The assistant retrieves candidate products from your live catalogue and inventory index, then reasons only over what it retrieved, with catalogue-scoped guardrails and a refusal path for when there’s no good match. That’s what stops it inventing a SKU, quoting a discontinued item, or promising stock you don’t have.

• We already use Algolia (or Elasticsearch). Do you replace it or build on top of it?+

  Usually build on top of it. If your keyword search and infrastructure already work, we add a semantic layer: embeddings, vector retrieval, and re-ranking, alongside the engine you run, with the keyword path as a fallback. We only recommend replacing search when the existing setup is the bottleneck, and we’ll say so plainly rather than rebuild for the sake of it.

• How do you keep search and merchandising fast enough for a storefront?+

  We treat latency as a hard budget. Embeddings and product attributes are precomputed into a vector and attribute index, so a query resolves against a warm index rather than calling a model on the hot path. Where a re-ranking model is worth the cost we cap it inside the budget and fall back to the base ranking the moment it would blow through. The shopper never waits on the model.

• Our product data is messy and half the attributes are missing. Can AI commerce still work?+

  That’s the normal starting point, and it’s where a lot of the value is. We build LLM-assisted enrichment: attribute extraction, category tagging, and description generation, to fill the gaps, with a human review step for the fields that matter. Better catalogue data is the fuel every search, ranking, and recommendation feature runs on, so we usually fix it early rather than paper over it.

• How do you prove any of this is lifting revenue and not just churning results around?+

  With holdouts. We wire experiment assignment and a persistent control group into the search, ranking, and recommendation features so you can read incremental lift on conversion and revenue per session against a real baseline, not a “before” that also changed. If a feature isn’t beating its holdout, you’ll see it, and that visibility ships as part of the build.