Natural language interface

In a natural language interface the user says what they want and the system works out what that means. There is no field for the customer, no dropdown for the status, and no date range picker: there is a sentence, and behind it a parser that turns “unpaid invoices from Acme over 500 last quarter” into four constraints. The trade is exact and worth naming. The user is spared learning the interface, and in exchange they lose the guarantee a control gives, which is that every option they can pick is an option that exists. A sentence can always ask for something the system cannot do.

That is why the good ones show their work. The parse is echoed back as something the user can see and correct, usually structured chips or a restated query, so the interface answers two questions at once: here is what I did, and here is what I understood you to mean. Where the guess is shown, a wrong guess costs a tap to fix. Where it is hidden, a wrong guess produces a confidently empty result and the user has no way to tell a bad query from an empty database. Echoing the parse also teaches the vocabulary: a user who sees “amount over 500” become a chip learns that amounts are something they may ask about.

A natural language interface does not have to be a conversation. A single search box that accepts a sentence, parses it, and returns a result is a complete NLI with no turn taking, no history, and no assistant persona. The turn-taking pattern is a further thing layered on top, and confusing the two leads teams to build a chat window where a query field would have done. It is also distinct from its neighbours in the same corner of the vocabulary. A command palette matches commands by their names, so it rewards a user who already knows the word for what they want; an NLI accepts the description instead of the name. A prompt input is the composer that submits a message to a model and is the shape most assistants ship, but the shape is not the pattern: the pattern is that language is the input, whether the parser behind it is a model, a grammar, or forty regular expressions.

The pattern is old, and its history is a warning about scope. The natural language search engines of the nineties failed less because parsing was hard than because they invited questions of a system that could only answer a few of them. The remedy has not changed: narrow the domain, publish examples so the user can see the range of what is understood, keep the parse visible and editable, and leave the controls available for anyone who would rather point than type. A faceted search sidebar beside the sentence box is not a redundancy, it is the fallback for the query the parser gets wrong. Speech adds nothing conceptually and everything practically: dictated language is looser, so the parse has to be more forgiving and the echo more prominent, since a user who cannot see the transcript cannot see the mistake.