#Embeddings

May 21

AI Product Recommendations in Magento — Adobe Sensei vs Custom OpenAI Embeddings

Adobe Sensei Product Recommendations ships only with Adobe Commerce — Magento Open Source merchants do not get it. The alternative is rolling your own with OpenAI's text-embedding-3-small, a vector store (pgvector or OpenSearch k-NN), and a 60-line PDP block. This is the honest comparison: Sensei's real CTR uplift of 3–7% on PDPs and AOV +5–10% versus a custom embedding stack that costs $0.02 per 1,000 products to embed, $5/month for a Redis cache, and gives you full ranking transparency. Includes the cold-start playbook (category-centroid embeddings until behavioural signals arrive), the pgvector schema, the OpenAI batch call, and the break-even math — at 100,000 monthly sessions, Sensei needs 0.4% incremental revenue to pay for itself; the custom stack needs 0.04%.

May 21

AI Search for Magento Stores — Semantic Embeddings vs Keyword Matching

Three search stacks were wired into the same Magento 2.4.9 + Hyvä storefront over a 50k-SKU catalog, and the results were not what the vendor decks promise. OpenSearch k-NN with OpenAI text-embedding-3-small ran at ~$0.02 per million tokens and held query latency under 120 ms; Algolia shipped in a weekend but billed $0.50 per thousand searches at scale; Coveo nailed precision but cost north of $20k a year. The article walks through each stack with real config, the embedding-model trade-off, and the 60/40 hybrid score blend that beats all three on the queries customers actually type.