Small decisions. Interesting possibilities.Submit contentSubmit

Cascade Search typed query filters

A search bar that knows when it doesn't know. cascade-search is a 27K-parameter model that parses queries like "open bugs from sam since last week" into typed filters, in your browser, in 0.25 ms. When it isn't sure about a word, it asks Jev. Only that word.

Added to Jevfast

How it uses Jev

A small in-browser model parses natural-language search queries into typed filters. When it is uncertain about a word, the search bar asks Jev about only that word.

App workflow

  1. Parse the search query

    The 27K-parameter model turns a query into typed filters in the browser.

  2. Detect an uncertain word

    When the parser is unsure about a word, it identifies that word for follow-up.

  3. Resolve only that uncertainty

    Jev is asked about the uncertain word, then the query can be represented as typed filters.

Why it is interesting

The system keeps parsing local and narrows escalation to an uncertain query component, rather than sending the entire request for interpretation.