There are two ways to point AI at your records. Put the records into the model — and it answers from a handful and makes things up. Or let your model write the question, and our box answers over every record, cited. Here’s the difference on our demo data — then run it yourself with your own AI.
The same questions, the same records. Computed live in your browser over our demo job book — not a canned screenshot.
| Question | Records in the model (RAG) | Your AI writes the query · our box answers |
|---|
The right column is the exact result of the query over every record, each answer cited to the records that produced it — not a judgment of what’s “correct,” just the arithmetic over what’s there. “Records in the model” answers from a top‑5 sample — partial by construction on anything that spans the records, and blind to a table it never retrieved. Illustrative demo data; the method is the point.
Your model writes the query from field names alone (no records leave this page). Our box runs it over every record and answers, cited and signed.
It carries the field names and your question — and not one record.
Claude, GPT, Gemini — whatever you use. It writes the query from the field names. It never sees a record.
Drop in the JSON it produced and run it on our box.
Bring a slice of your own data and the questions you actually ask. Your AI writes the query; the answers stay exact, cited, and in your building.
Honest fence · answers are cited to the record and auditable — a good AI can also get a single lookup right; the win is on every question that spans your records, and on being able to prove it.