Methodology

Every spec traces back to the manufacturer. That's the whole point.

A database that can't show its work isn't worth querying — by a person or an agent. Here's exactly how a raw manufacturer page turns into a fact you can trust.

Three layers, each doing one job

Raw

Nothing is overwritten

Every value is collected and kept exactly as the manufacturer published it, tagged with where it came from. If a translation rule gets smarter later, it re-runs against this — nothing has to be re-collected.

Canonical

One schema, applied deterministically

A reviewed mapping dictionary turns each source's own vocabulary into a shared set of keys — the same key, the same unit, regardless of which manufacturer it came from.

Identity

The same part, tied together

GTINs, ETIM/eClass codes, and distributor SKUs cross-reference so the same physical product collected from different places resolves to one record, not duplicates.

What "manufacturer-verified" actually means

Distributor catalogs are a fine way to discover that a part number exists — that's just sourcing. But the source shown for a product's specs is never a distributor re-listing. It's the manufacturer's own product page, or the product isn't marked verified.

93% of the database meets that bar today. The remaining 7% — mostly brands not yet crawled directly from their own site — is shown, not hidden, and labeled as pending verification.

Every fact carries
value + unit
source_key (the raw label)
confidence score
source_url (manufacturer page)
fetched_at (timestamp)

The LLM designs the rule once. It doesn't guess the answer each time.

The distinction matters more than it sounds like it should.

  • What we don't do: ask a model to read a spec sheet and guess the current rating every time a product is imported. That's expensive, unreproducible, and a wrong answer is silent.
  • What we do instead: for each new source, use LLM assistance once to draft a translation rule — "this label, in this source, means this canonical key, parsed this way." A person reviews that rule.
  • Then it's deterministic: a plain parser applies the reviewed rule to every product from that source. Same input, same output, every time — and it's auditable, because the rule is a readable file, not a model's internal weights.

Change is a diff, not a surprise

Every import is versioned rather than overwritten in place. When a rating changes or a part is discontinued, it shows up as an old value, a new value, a manufacturer link, and two timestamps — a claim you can independently verify, not a notification you have to take on faith.

Ask about change tracking →

Ready to see it in action?

For developers, distributors, or manufacturers.