Show, don’t configure
Upload 1–5 representative documents. One works; varied samples improve validation.
Document extraction infrastructure
Teach it once.
Extract forever.
Turn any recurring document into a fast, typed API. Create one template from a few samples, then process thousands of files without paying for an LLM on every extraction.
~1 creditper extraction
No LLMat extraction time
99.1%sample confidence
extract.ts LIVE
REQUESTRESPONSE142ms
// One template. Any volume.
const result = await decomp.extract({
template_id: "tpl_invoice_01",
file: invoicePdf,
});✓ Extraction complete 0.991 confidence
{
"invoice_number": "INV-2048",
"vendor": "Northstar Labs",
"total": 18420.00,
"currency": "USD"
}01 OCR→02 Extractors→03 Consensus→04 JSON
BUILT FOR THE SYSTEMS THAT RUN THE WORLD
FINTECHLOGISTICSHEALTHCAREACCOUNTINGINSURANCE
THE DECOMP MODEL
Pay for intelligence once.
Keep the speed forever.
Traditional extraction runs an expensive model on every file. Decomp spends intelligence up front, turning your examples into deterministic extractors that run at software speed.
We build the template
OCR, schema discovery, extractor generation, consensus testing, and validation happen in the background.
Extract at any volume
Send a template ID and a document. Get typed JSON back, without LLM inference in the hot path.
BUILT FOR SCALE
The unit economics
finally make sense.
One expensive setup. Millions of inexpensive extractions. Your cost curve flattens instead of growing with model tokens.
See generous launch pricing →COST PER DOCUMENTVOLUME →
LLM extractionDecomp
100×cheaper extraction target
0LLM calls per extraction
START WITH 1,000 FREE CREDITS
Your documents already
know the schema.
Give Decomp a few examples. We’ll do the rest.