Forty characters of supplier shorthand in. A full, cited specification out.
Tap a token to see the field it has to become.
A value ships only if its quote can be found again.
63% of filled fields cost no tokens. A regex cannot hallucinate.
Documents matched on the part number, never the wrong product.
Every value comes back with its supporting sentence.
Quote re-found: publish. Weak: a human. Missing: deleted, and counted.
A real quote proves no text was invented — not that it says what was claimed.
The model states a value and cites a sentence that does not exist.
The quote is real, and the inference drawn from it is not.
The synthetic set degrades a complete record into shorthand, so the original is ground truth — but the parser's author wrote the generator too. The 12 real-document SKUs answer that: manufacturers' own PDFs, ground truth read out by hand. Small, and that limits it — but it is the one benchmark the author did not write, and it found three defects 652 generated SKUs never did.
Two thirds of filled fields never reach a model — that is both the cost story and the
certainty floor. The rest is metered: 678 tokens per SKU, $625 per million.
Real documents cost about three times that, because a submittal sheet is a dense
two-page table. List rates, cold cache.
Manufacturer domains first, never a shopping site — marketplaces are refused by a
blocklist checked on the requested URL and again on the redirect target.
robots.txt is honoured and requests are paced. No source URL means the
field stays empty rather than plausible.
Three sets, never averaged — they measure different things.
Rules do the cheap two thirds. The gap is documents, not model capability.
Pick a part, or type your own. The last example is meant to refuse.