Charlie Hull on search in the AI era

Recorded on-site at Berlin Buzzwords 2026.

Search picked up a new user. An AI agent running a multi-step loop of queries is not the two-keyword human the engines were designed for, and the field has not fully answered what that means for load or design. AI brought useful tools for query understanding and recall expansion. It did not remove the older problems, most of which are about process, evaluation, and measurement rather than the underlying technology.

A few things that come up:

• BM25 is back as a benchmark, and edit-distance spelling correction still beats LLMs when the misspelled form is not in the training data.
• Inverted indexes are debuggable. Embeddings are not. The explainability gap shows up the moment a weird result appears.
• Reaching for OpenAI embeddings by default ignores whether the model actually understands your domain. Grocery retail has its own vocabulary, and generic web-scale training may not carry it.
• Most vector-search adoption starts with an executive who read a Forbes article or an engineer who wants a new toy. Neither begins with the problem being solved, and both end up owning the infrastructure bill, the API dependency, the salaries, and a bus factor of one.
• Language coverage is the field’s blind spot. English is well covered. European languages get less attention, and rare languages spoken in parts of South America or Africa get almost none.