Transcribe (Kraken)

🤖 AI Drafted (Not reviewed)

Use this when: you want fully on-device HTR from Kraken — its own neural baseline segmentation plus a CC-BY recognition model (McCATMuS), no paid API calls. It reads each line and SAVES the transcript tied to its baseline (per-line geometry the reader can overlay). Install the Kraken runtime AND a recognition model first from Settings → AI → Local Inference (the on-device model catalog); the run reports clearly if either is missing. Best on the historical hands McCATMuS was trained on. Pair with a Cleanup pass for hard pages.

Folder	/Transcribe
Steps	2
Tags	preset, transcribe, htr, kraken, historical, local, single-pass

Steps, in run order

1. Files

Tool: Files — Pass through input files from workflow context

2. Kraken HTR — baseline + recognition

Tool: Transcribe — Extract text from images (OCR)

Settings this step uses:

Option	Value
kraken_model	kraken-mccatmus
language	auto
regions_first	no
update_page_content	yes
vision_mode	kraken

What this step asks the model:

Transcribe the text visible on this image.

Language: transcribe in the language of the source. Do not translate, and do not assume the document is in English.

Rules:
- Output ONLY the transcription. No headings, no preamble, no commentary,
  no summary, no notes, no explanations, no observations about quality or
  legibility, no descriptions of seals or images, no language about the
  difficulty of the handwriting.
- Preserve original layout, line breaks, and paragraph structure.
- Preserve original spelling and capitalisation, including ALL CAPS
  headers if they appear that way.
- Preserve orthography exactly as written, including all diacritics
  and accent marks (e.g., keep "Chocó" as "Chocó", never "Choco";
  keep "Ramón" as "Ramón", never "Ramon").
- Do not strip accents, tildes, cedillas, or umlauts. If a mark is
  visible, keep it.
- Include every visible text element — headers, body, marginalia, stamps,
  signatures (transcribe the signed name as written), printed labels,
  handwritten annotations.
- For text you cannot confidently read, use explicit uncertainty markers:
  [ilegible] for unreadable text and [uncertain] for plausible-but-low-
  confidence readings. Place the marker inline at the uncertain span.
  Do not guess. Do not fill in.
- Do NOT invent dates, numbers, names, or words that are not legibly
  present. Do not normalise dates ("23/7/1999" stays "23/7/1999", not
  "1999-07-23").
- Do NOT repeat any portion of the transcription. Output each visible
  passage exactly once.
- If the image contains no legible text, output the single token
  [sin texto].