Voice is the feature people ask the most careful questions about, so here's the whole story in one place, moved off the pricing page where it was asking shoppers to care about architecture before they cared about the product.
The words come from your phone or browser
The words themselves come from the speech service your phone or browser provides. It turns the sound into a sentence and hands Buhata the sentence, so the sound never reaches us. In Chrome and on most Android phones that service is Google's, and it can send the recording to Google to be recognized. In Safari and on Apple devices it's Apple's, and in Microsoft Edge it's Microsoft's. What happens to the recording there is covered by their own privacy policies.
Voice that learns instead of asking every time
A voice command isn't one big call to a language model. Your phone reads the sentence first and checks it against what you've already written. So a shop you've typed before is recognized as that shop, not as whatever it sounded like. A command it finishes there never reaches us unless you mark it with a thumb, which is covered further down.
What your phone cannot place goes to us, where it can still be answered without a language model.
A command you type into the command bar goes the same way. Your phone reads it first, and what it can't place comes to us. On Premium, that part can go on to the language model. A typed command is kept in the same 30-day record as a spoken one.
What leaves your phone is stored as a shape
A sentence shaped like ones other people have said is matched against a pool of patterns. The pool holds the shape and not the contents, with the name and the amount stripped out. A sentence you've said before is remembered against your account as a shape, with the name and the amount taken out. It's filed under a one-way code, not under the words. So the second time is instant, and nobody could read anything back from that store.
And one short record that does hold the sentence
The pattern pool isn't the only thing kept. When a command reaches us, the sentence itself goes into a log against your account, and it deletes itself after 30 days. I keep it because it's the only way to find the commands that were heard wrong and teach the recognizer the words you use. It never appears in the app, and it isn't used for advertising. Deleting your account erases it immediately. If you want yours cleared sooner, write to [email protected] and I'll do it. The privacy policy describes this same record in its voice section.
A thumb up or down on a command is kept with the rest of your Buhata data. The thumb, its sentence, the reasons you pick and whatever you type sit in your account's copy on the server while you're signed in. If the AI model or your own saved phrases read the command, the thumb reaches me when you press it, with the answered sentence. An up can save that reading as one of your phrases, and it counts toward that pattern joining the pool. A down forgets the reading but keeps the command's pattern, with the name and the amount taken out. A thumb down holds that pattern against your account until you've confirmed that kind of command twice, or you delete your account.
Ticking "Allow Buhata to use this request to improve commands" adds a separate record for the few people who fix commands. It holds the sentence, the reasons, which part of Buhata read it and the pattern it was read as. That copy deletes itself after 90 days. What you type about it isn't part of that record.
Only the awkward ones reach a language model
Only the genuinely awkward ones go to a language model, which is why the microphone doesn't have to be priced like an AI product.