New Search Method Lets AI Agents Build Their Own Queries

When you ask an AI assistant to find something — a cheap flight, a contract clause, a pair of shoes — it usually runs one fixed search recipe chosen by whoever built it. A new paper by Damian Delmas proposes something more flexible: letting the AI write its own search recipe at the moment you ask. The method is called Algebraic Retrieval, and the idea is that filtering rules, relevance criteria, and ranking preferences can all be combined in one mathematical query, then tweaked after the AI peeks at the results.

The analogy is a kitchen. Today's AI search tools are like a single preset button on a microwave — one setting, one outcome. This approach hands the AI a full set of dials and lets it decide how to combine them for your particular question. If a first attempt returns junk, the AI can revise its search program rather than start over.

The honest caveat matters here. The paper isn't trying to show the searches get better. It tested whether different ways of writing the same query produce identical results, using a public collection of 11,429 documents. They matched, with score differences smaller than one in a million, and only one tied pair of documents swapped order. That's a plumbing check, not a quality check.

So what does this mean for you? Probably nothing this year. It's a five-page research note with a single author, no user testing, and no evidence that smarter search plumbing produces better answers for real people. But it points at where AI assistants are heading: instead of searching the way their designers decided, they'll decide how to search based on what you actually asked. If that works, it could mean fewer frustrating "that's not what I meant" moments — and it could also mean it's harder to understand why an AI showed you what it showed you.