Prompts

There are exactly two LLM calls with editable text, so the Prompts screen has two whole prompts you control completely:

• Extraction: the single call that returns the title, tags, correspondent and date.
• OCR: the vision-LLM prompt used when OCR runs. (The dedicated Mistral OCR endpoint ignores it.)

Each is shown as a complete prompt, so what you edit is exactly what the model receives. The output shape stays enforced by a strict schema regardless of the prompt text: you control the instructions, not the structure.

Both calls also carry the original document alongside the prompt when the model is vision-capable: OCR transcribes it, and extraction uses it as visual evidence next to the recognised text (files over the OCR page cap attach only their first and last pages).

Insertable variables

Click a variable to drop it at your cursor. At run time each is replaced with the document’s own values.

Extraction prompt

Variable	Replaced with
{{content}}	the recognized text (OCR output, or paperless’s text)
{{language}}	your configured output language, or auto
{{all_tags}}	every existing tag, comma-separated; once any tag has an AI hint (Tags page), hinted tags render as a Markdown table of tag → hint with the rest listed below
{{all_correspondents}}	every existing correspondent, comma-separated
{{tag_policy}}	a sentence reflecting your create new tags setting
{{correspondent_policy}}	a sentence reflecting your create new correspondents setting
{{title}}	the document’s current title
{{tags}}	the document’s current tags
{{correspondent}}	the document’s current correspondent, or (none)
{{created}}	the document’s current date, or (none)
{{filename}}	the original filename, or (none)

OCR prompt

Variable	Replaced with
{{language}}	your configured output language, or auto
{{filename}}	the original filename, or (none)

Note

Unknown placeholders are left untouched, so a typo like {{ contnt }} stays visible in the rendered prompt instead of silently disappearing.

Test on a real document

Pick a document from your paperless instance and hit Run test. It runs your current (even unsaved) prompt against that document and shows:

• the structured result: title, tags, correspondent, date (or the transcribed text, for OCR),
• the token usage, and
• the exact rendered prompt that was sent.

Testing never writes anything back to paperless.

Reset to default

Defaults live in code; the database stores only your override (one row per prompt). Reset to default is a plain delete that drops your override, and because defaults can keep evolving, you pick up improvements when you reset. There is no version history to manage.