Information retrieval
Relevance
Whether the result a person receives is one they can act on. Hybrid ranking, similarity, and the structure of the data sit under this subject.
Broader subject: Information retrieval
Within this subject
- Hybrid search — Lexical ranking such as BM25 combined with vector similarity in one result. The useful question is when each signal should win, and how to check that the blend did not quietly get worse.
- Similarity — How a ranker decides that two items are close: a lexical score such as BM25, or a vector score such as cosine. I study the choice, and how to tell that a new score did not quietly get worse.
- Data structuring — The shape of the data before a ranker runs. Vector search and lexical tagging sit under this subject. The structure decides what close is allowed to mean.
Work
- Concept search before vectors were the default — A books-and-journals search that let people explore by concepts, built in 2018 from phrases and similarity, not from a vector database.
- One catalog, many storefronts — Product search, autosuggest, filters, and per-field similarity for an omnichannel retailer: web, marketplace, and stores.
- Return one record, not a list — From millions of candidates that share an organisation's name and address, return the single best record.
- A move that kept the answers — A production move from self-managed Apache Solr to managed OpenSearch, checked by whether the top results still agreed.
- A new engine, tried on a laptop — A hands-on reading of Luxir, a new hybrid search engine, run from a Mac or Windows laptop without compiling C++.