Notes on search, e-commerce, and the engineering behind them.
SAP Commerce Cloud ships with Solr search out of the box, yet the experience is often disappointing. Drawing on 15+ years of tuning Hybris search, here are three practical, opinionated fixes: turning off fuzzy and prefix-wildcard matching, keeping field boosts sane, and using ValueResolvers instead of the long-deprecated ValueProviders.
Read article →A side project that brings several search techniques together: hybrid (lexical + vector) retrieval, cross-encoder re-ranking, and an agentic LLM shopping assistant — all toggleable on the fly, on both Solr and OpenSearch. Plus a few honest learnings about weighting, self-hosted models, and where the real effort goes.
Read article →A cross-encoder re-ranks the top results of an existing search by looking at the query and each product together. Tested on 1.4M Amazon products, re-ranking the top 50 gave a noticeable uplift in about two out of three queries — and never made things worse. An easy, high-impact first step beyond lexical search.
Read article →Solr 9.9.0 brings end-to-end semantic vector search built right in — but you're limited to a few cloud-hosted model providers. Here's how to connect your own custom embedding model on any endpoint, plus two extras: combining multiple fields into one vector, and lazy vectorization that only re-embeds when relevant fields change.
Read article →Precise product search and the ability to refine results with facets are integral to any online shop — but calculating every facet for every search is inefficient, and maintaining them by hand on the category tree is tedious. Here's how to make faceting dynamic and tailored to each individual search with Solr.
Read article →