Getting better answers from Ask CentProof.

Ask CentProof turns a plain-English question into a safe, read-only query against your local database and shows you the answer with the receipts underneath. The 0.2.1 release makes it noticeably better at three things people actually ask for. They're in beta, so they're off until you switch them on.

First, turn it on

Open Preferences (⌘ ,) → AI → the "AI Search (beta)" section, and flip on whichever of the three you want:

• Smart merchant & category matching
• Conversational follow-ups
• Multi-step questions

With them off, Ask CentProof behaves exactly as it did before. Nothing else changes, and nothing leaves your Mac either way.

Say it how you remember it

You don't always remember exactly how a merchant shows up on a statement. With smart merchant & category matching on, CentProof matches your phrasing to a real name in your data using on-device similarity:

• "spending at the warehouse club this year" → matched to Costco
• "how much on the streaming service?" → matched to Netflix
• "rideshare last month" → matched to your tagged Uber / Lyft

CentProof tells you what it matched ("warehouse club → Costco") so you can trust it — and the dollar figures stay exact. Only the matchinggets more forgiving; the math doesn't change. This works best when the merchant or category already exists in your tagged data.

Refine instead of re-typing

Real questions come in layers. With conversational follow-ups on, your next question builds on the last one instead of starting over:

• "top 5 merchants this year"
• → "just the Apple Card"
• → "only over $100"

A small "Next question refines …"note shows what you're building on, and New question clears it whenever you want a fresh start.

Compare two things in one question

This is the big one. Some questions are really two questions stitched together — and they used to need two separate searches. With multi-step questions on, CentProof recognizes a comparison, runs each part as its own query, and lays the results side by side:

• "compare Costco vs Whole Foods this year"
• "did I spend more on groceries this month or last month?"
• "dining vs entertainment in 2025"

You get a one-line summary at the top (for two-number comparisons, it even computes the difference) and a card per part underneath, each with its own count and a button to see the transactions behind it. Every figure is computed locally from your data — the AI decides which queries to run, never the numbers themselves.

See the exact receipts behind any number

Every answer already shows its source rows, but the new View transactionsbutton opens them in a focused pop-up — including for grouped answers like "top 5 merchants" and for each part of a comparison. The pop-up shows precisely the transactions that produced the number (not a loose re-search), and every row links straight to its original PDF page, with the line highlighted.

So when an answer looks surprising, you're one click from the receipt — the same "trace any number to its source" principle that runs through the rest of CentProof.

What doesn't change

• It's still local. The model and the matching both run on your Mac. Nothing about your transactions leaves the machine.
• The numbers are still exact. The AI plans and explains; the figures always come from a read-only query against your own data.
• It's opt-in. Off by default, on when you say so, and reversible any time in Preferences → AI.

One honest note: these features lean on the data you've already tagged. The more you've named your merchants and categories, the better the matching and comparisons get.