Transcribe Paleography (Economy)
🤖 AI Drafted (Not reviewed)
Use this when you want a near-free paleography draft: Apple Vision line geometry (free, on-device) feeds a cheap local HTR backend, then ONE small-model cleanup pass corrects it against the image. Total paid calls per page: at most 1, vs the ensemble’s 15+.
| Folder | /Transcribe |
| Steps | 3 |
| Tags | preset, transcribe, paleography, economy, line-htr |
Steps, in run order
1. Files
Tool: Files — Pass through input files from workflow context
2. Line HTR — Apple Vision boxes + local model
Tool: Economy HTR — Free Apple Vision line boxes + a cheap local HTR backend (no paid API calls).
Settings this step uses:
| Option | Value |
|---|---|
backend |
apple |
language |
es |
pad_y |
0.35 |
scale |
2.0 |
3. Cleanup — single small-model pass
Tool: Transcribe Review — Second-pass QA of a prior transcription against the image
Settings this step uses:
| Option | Value |
|---|---|
language |
auto |
prompt |
You are a paleographer correcting a MACHINE line-by-line HTR draft of a historical manuscript against its image. The draft (in Context) came from a line-level recognizer: its line order and line breaks are trustworthy, but individual characters and word boundaries are not, and it never marks uncertainty. Compare each draft line against the image and correct it. Preserve original spelling, capitalisation, punctuation, and line breaks; never modernise. Expand abbreviations only when unambiguous, expanded letters in [brackets]. Use [UNCERTAIN] for plausible readings and [ILLEGIBLE] only when truly unreadable. Mark crossed-out text [deleted: …], marginalia [M.N.], rubrics [Rúbrica], seals [Sello]. Common procesal/cortesana confusions to check: rn/m, c/e, u/n, long-s/f, t/l. Output ONLY the corrected transcription — no commentary, no diff. If a draft line is already correct, keep it verbatim. Do not invent text not visible in the image. If the page has no legible text output [sin texto]. |
provider_name |
$vision_small |
thinking_mode |
medium |
update_page_content |
yes |
vision_mode |
llm |
What this step asks the model:
You are a paleographer correcting a MACHINE line-by-line HTR draft of a historical manuscript against its image. The draft (in Context) came from a line-level recognizer: its line order and line breaks are trustworthy, but individual characters and word boundaries are not, and it never marks uncertainty.
Compare each draft line against the image and correct it. Preserve original spelling, capitalisation, punctuation, and line breaks; never modernise. Expand abbreviations only when unambiguous, expanded letters in [brackets]. Use [UNCERTAIN] for plausible readings and [ILLEGIBLE] only when truly unreadable. Mark crossed-out text [deleted: ...], marginalia [M.N.], rubrics [Rúbrica], seals [Sello]. Common procesal/cortesana confusions to check: rn/m, c/e, u/n, long-s/f, t/l.
Output ONLY the corrected transcription — no commentary, no diff. If a draft line is already correct, keep it verbatim. Do not invent text not visible in the image. If the page has no legible text output [sin texto].