01OCR and vision extraction, with a confidence score per page
Every scanned page gets extracted and scored, not just transcribed — so quality is a visible number, not an assumption.
confidence per page96% · 61% · 34%
Records & operations
Twenty years of paper your AI can't read
The oldest records are often the ones that matter most legally — and the ones sitting in filing cabinets or an unverified OCR dump from years ago. If your knowledge system can't trust what came out of a 1994 scan, its ground truth has a decades-long blind spot.
A digitization run over box 14 of a pre-2005 paper archive: OCR and classification per page, confidence scored, and anything below threshold — an unclassified correspondence at 34% — routed to a human instead of guessed at.
Sound familiar?
The most important records are the oldest ones
Original deeds, decades-old contracts, the first version of a policy that's been amended nine times since — all sitting in filing cabinets or unindexed scan dumps that nothing modern can search.
Nobody trusts what OCR pulled out of them
A scanned archive was run through OCR once, years ago, and the output is full of garbled text nobody has verified. Trusting it for anything that matters means re-checking it by hand anyway.
"Somewhere in box 14" is not a retrieval strategy
Without labels, doc types, or an index, finding a specific historical record means someone physically searching through boxes — the exact problem a knowledge system is supposed to solve, made worse by decades of backlog.
Why the usual fixes don't move the number
A one-time bulk OCR pass, years ago, never revisited
OCR quality varies wildly by scan quality and document age, and a bulk pass with no confidence scoring means nobody knows which pages came out clean and which came out garbage.
Manually re-keying the important documents
It scales to the handful of documents someone got around to this quarter, not to twenty years and dozens of boxes — the backlog only grows.
Treating scanned archives as out of scope for the AI system
The oldest records are often the ones with the most legal or historical weight — excluding them means the system's ground truth has a decades-long blind spot exactly where accuracy matters most.
A generic OCR API with no downstream classification
Extracted text with no doc_type, no labels, and no confidence score is just a pile of strings — it still needs the same taxonomy and permission structure as anything else being ingested.
How we solve it
02The same document taxonomy and classifier as any other source
A scanned deed from 1994 gets typed the same way a native PDF uploaded today does — one taxonomy, one classifier, no separate pipeline for old paper.
doc_type: Deed | Assessmentone taxonomy, all sources
03Low confidence routes to a human, not a guess
A page the classifier isn't sure about doesn't get silently mistyped — it gets flagged for a person to confirm, the same rule that applies to any low-confidence ingestion.
< 70% confidence → reviewnever silently guessed
What you get on every document
Full provenance on every digitized page
Extraction engine, content hash, and original source are recorded per document — so a digitized 1994 deed is exactly as auditable as a document uploaded this morning.
Confidence is visible, not assumed
A 34% confidence classification is flagged, not treated the same as a 98% one — the difference between trustworthy and unverified is a number anyone can see.
Decades of paper become searchable knowledge
Once labelled and typed, a 1994 deed is retrievable by what it is — not by which box it happens to sit in — the same way any other document in the system is found.
No separate pipeline to maintain
Scanned archives, native PDFs, and connector-synced documents all land through the same ingestion, classification, and permission model — one system, not a legacy-data side project.
