How Extraction Works

Reading a property document well is harder than it looks — scanned certificates, photographed leases, decades-old solicitor formatting. Proprietas handles them with a layered pipeline: OCR to recover text from scans and images, machine-learning classification to identify the document and route it, and large-language-model extraction to pull the structured fields — each one grounded in the source text and confidence-scored. It’s built so you never have to take the output on faith. Three principles make the extraction trustworthy: every figure is cited, confidence is surfaced, and anything uncertain waits for a human.

​

Page Citations on Every Figure

Each extracted value links back to the page it came from. Open a lease and every term date, break clause and rent figure carries a “see page 3” reference; click it and the source wording is highlighted in the PDF. Your legal team can spot-check a 30-page lease faster than it takes to make tea. This isn’t decoration — a value whose source wording can’t be found in the document is treated as a hallucination and dropped rather than shown.

​

Confidence Is Surfaced, Never Hidden

Every extraction carries a confidence score, and it’s calibrated — the raw model is over-confident, so Proprietas derates the score against observable signal quality (was the text clean or a noisy scan? is the address plausible?). You see a realistic number, not a flattering one.

​

The 0.85 Statutory Bar

Statutory-compliance dates have a hard rule: below a confidence of 0.85 they never auto-file — they route to the Inbox for a human to confirm. This is a fixed business policy, tuned for legal and liability reasons, not a model setting. The corollary: a confident-but-wrong date is a bug we fix by improving the parser, never by lowering the bar.

​

Clean Scans, Automatically

If a document is a poor scan, Proprietas falls back to OCR and — when confidence is low or an address looks garbled — re-reads it with OCR forced, so a bad photocopy doesn’t quietly produce a bad date.

​

You Never Pay Twice

Every extraction is cached by the document’s content and the schema it was read against. Upload the same PDF again and the result is served from cache — the same file is never re-charged against your AI usage.