Source-bound value
The proposed value exactly matches field-eligible, cited authoritative evidence from the current source revision.
Reviewable document extraction
Extract named fields. Keep uncertainty visible.
Select a folder, name up to eight fields, and process as many as fifty source files sequentially. FileNyani validates model proposals against cited evidence before assigning a trust state.
AI proposes. Evidence determines status. Human review stays separate from the original AI result.
Extraction result
Customer · Notice period · Price
| Source | Customer | Notice |
|---|---|---|
| agreement-a.pdf | Found Acme Corporation | Needs review 30 days |
| agreement-b.docx | Found Contoso Limited | Not extracted |
Trust states
“Found” has a deliberately high bar.
Exact text must match cited authoritative content selected for that requested field. Otherwise the value remains a candidate for review or is reported as not extracted.
Useful, not authoritative
A candidate exists, but evidence does not satisfy the Found boundary. OCR-only values always stay here.
No accepted proposal
The model supplied no usable value, or could not complete a valid structured response for that field.
Not extracted does not mean absent.
The field may be missing from the bounded evidence, outside the supported source content, or unavailable because the local model did not complete that field.
Human review overlay
Review decisions never rewrite the AI result.
Confirm a source-bound value, accept an original candidate, enter a correction, mark no value, or clear the overlay. Reviews remain revision-bound and export separately from AI output.
Name fieldsWrite explicit labels or reuse a saved, editable field list.
Extract sequentiallyProcess one source at a time with bounded evidence and cancellation.
Inspect and reviewOpen evidence and record a separate human decision.
Export locallyCreate deterministic CSV or portable JSON with both layers.
Reusable field lists
Save labels such as “Contract basics” and reuse them in another folder. Reuse fills the composer but never starts work automatically.
Current-source checks
Changed or missing sources disable stale review actions instead of silently binding old evidence to new content.
Portable output
CSV and JSON preserve field order and keep AI status/value separate from human review status/value.
Turn a document folder into a reviewable table.
No hidden workflow schema or cloud extraction service. You choose the scope, fields, model, and final review decisions.