How it works

Four steps. Every claim quotes your own data sheet.

GRADE reads an inbound inquiry, screens it against your line card with typed rules, and drafts a reply. A model reads; code decides; the rep approves before anything is sent.

01
Parse
The inquiry becomes typed constraints — pH as two numbers, VOC as a ceiling, certifications as booleans. A "-free" clause is read as an absence, by code, not guessed.
02
Screen
Every constraint is a SQL WHERE clause over your parsed catalogue. A VOC ceiling is a filter, never an embedding comparison. Vector similarity never enforces a hard spec.
03
Rank
Only products that survive every hard constraint move on. Each is placed in exactly one lane, and near-misses surface separately — never as a match.
04
Draft
The model sees survivors only, and drafts a reply that quotes your data sheet verbatim for each claim. The rep edits and approves. GRADE sends nothing itself.

What the quote guarantees, precisely. Every value is quoted verbatim from that product's own data sheet — never inferred, never carried across from another product's document. Where a producer publishes several grades on one sheet, a stated line may cover more than one of them; we show you the line and the sheet it came from, so you can see exactly what was read rather than take our word for it.

The rule that makes it defensible

Structured extraction first, retrieval second. If we can't say in one sentence why a product was ruled out, the screen is wrong.

How we test that rule →

What the model never does

  • Enforce a hard constraint by similarity.
  • See a product that failed screening.
  • Invent a grade that isn't in your catalogue.
  • Send anything to a customer on its own.

See it run

Paste an inquiry on the home page.

It screens a demo catalogue live — parsed constraints, a recommendation with receipts, and an honest no-match when nothing fits.

Talk to us about your line card → Try the live demo