Processing

The Processing screen is where you pick the models and tune how documents are enriched.

Models

Two model slots, each chosen from your connected providers:

• OCR model: reads scanned documents. Used only when OCR is enabled. Must be vision-capable.
• Language model: extracts the title, tags, correspondent and date.

If you haven’t connected a provider yet, the picker links you to the API Keys screen.

The pipeline

OCR

• Run OCR before extraction: when on, each original is re-read with your OCR model and the result is written back to paperless’s content. When off, paperless’s own (free) Tesseract text is reused. Disabled until you pick an OCR model.
• Skip OCR above … pages: larger files reuse paperless’s text instead of paying for OCR, and extraction sees only their first and last pages instead of the whole original. Blank means no limit.

Extraction

Extraction reads more than text: when the language model is vision-capable, the original document is attached to the call as visual evidence (layout, letterheads, stamps and signatures), with the recognised text as an aid. Files over the OCR page cap attach only their first and last pages; models that can’t take files run on text alone.

• Run extraction: when off, the pipeline stops after OCR (OCR-only mode). The recognised text is written back, the trigger tag is dropped, and the document’s metadata is left untouched. No language model is required then. Turning both extraction and OCR off leaves the pipeline nothing to do, so jobs fail with a configuration error.
• Create new tags: let the AI create tags that don’t exist yet.
• Create new correspondents: let the AI create correspondents that don’t exist yet.
• Skip extraction above … pages: larger files are skipped entirely (no OCR, no extraction). Blank means no limit.

Two caps, two jobs

The extraction cap is the outer gate (over it → skip everything). The OCR cap is the inner gate (over it → reuse paperless text, but still extract) and doubles as the attachment limit: over it the extraction model sees only the document’s first and last pages instead of the whole file. So with OCR ≤ 20pp and extraction ≤ 100pp, a 50-page file extracts on paperless’s text plus its first and last pages, and a 1000-page book is skipped. See How processing works.

Apply

• Auto-apply suggestions: when on, changes are written to paperless immediately and the trigger tag is dropped. When off, every result is queued for review.

General

• Poll interval (seconds): how often the poller checks paperless. Floored at 15s (PATCH/upload are CPU-heavy on paperless).
• Output language: the language the model writes metadata in. auto uses the document’s own language.
• Correspondent blacklist: one name per line; these are never assigned or created as a correspondent.

Defaults

Setting	Default
Poll interval	60s (minimum 15s)
Auto-apply	on
Create new tags	off
Create new correspondents	on
Run extraction	on
Skip extraction above	100 pages
Run OCR	on (once a model is picked)
Skip OCR above	20 pages
Output language	auto
Max attempts per job	3