Honesty page
How the index works — exactly
What runs on your documents
Deterministic rules. No AI model, no cloud processing. When you ingest a document, fixed pattern rules read its text layer and extract three kinds of facts. Every fact is stored with evidence: the exact text span and the rule ("cue") that fired. That is why quality here is a number and not an adjective.
| Fact | How it is extracted |
|---|---|
| Deadlines | A date counts as a deadline only when an obligation cue sits just before it in the text — due by, expires, no later than, must be completed by, deadline is, an obligation verb + by, and kin. Invoice dates, effective dates, "scheduled for", and past references are recorded as mentions, never deadlines. |
| People | Names in labeled positions (From:, Bill To:, Attendees:, Landlord:, technician:), salutations, "between … and …" contract parties, and signature blocks. Organization names are filtered out by suffix rules. |
| Tags | Document type by cue scoring (invoice / contract / letter / memo / note / receipt), plus topic tags from a published keyword map. |
What it cannot do — printed on the box
- No OCR. A scanned invoice image is stored and tagged as a scan; its text is not read. The due date inside the picture is missed. We say so instead of pretending.
- No relative dates. "Net 30" produces no deadline, because there is no calendar date to extract. The document is still indexed by people and tags.
- Simple PDFs only. The reader pulls the text layer from ordinary PDFs; exotic layouts can scramble line order.
- Rules, not understanding. A contract that says "hereby" may earn a legal tag it did not need. The evidence span lets you see the mistake and its cause.
The measured scores
Test corpus: 24 synthetic documents (invoices, contracts, letters, memos, notes, receipts, 2 text-layer PDFs, 2 images) generated for this product, scored against a written answer key. Latest run of the shipped engine:
| Field | Precision | Recall | Counts |
|---|---|---|---|
| Deadlines (by date) | 1.00 | 1.00 | 13 found / 13 expected / 0 invented |
| Deadline kind | accuracy 1.00 on matched | 13 / 13 | |
| People | 1.00 | 1.00 | 26 / 26 / 0 |
| Tags | 0.91 | 0.96 | 51 right / 5 extra / 2 missed |
| Document type | accuracy 0.96 | 23 / 24 | |
Known misses, kept in the open: the plant-rota note is typed as a plain document; a follow-up letter is not tagged client; two documents earn a topical tag (payment, legal) the answer key did not assign. The regression job replays this corpus on every engine change.
Your documents
The MVP retains the extracted text and the index facts so browsing and search work. It does not retain your original file bytes, it never trains anything on your documents, and no third-party AI service sees them — there is no AI service in the loop.