wondarchive
EN
Log in Create my free account

wondarchive › How it works

From photo to text, in six steps

You prepare nothing. You deposit the folder exactly as the camera wrote it, and you only step in where the machine hesitates.

Deposit a whole folder

Drag in the research trip’s folder, subfolders included, whatever the depth. Every file gets a fingerprint computed in your browser: what’s already online isn’t sent again, and an interruption resumes where it stopped. The original JPEG is kept intact, never recompressed.

Straighten, on your own machine

While the files are uploading, your browser decides each page’s reading direction — a fraction of a second per image, nothing sent. When it’s unsure, it shows you the four poses large, with a magnifier, and you click the right one. Nothing happens on its own, nothing blocks the upload.

Transcribe

Every page is read from the original, at full resolution. In the same pass, the tool notes the nature of the item, its language, its date, dates mentioned, proper names and figures, and drafts a short description. A handwritten page is flagged as such: no text is invented for it.

Translate into your language

Page by page, with the end of the previous page and the start of the next as context, and a glossary of the terms of the fonds so that “Lampenhaus” is translated the same way throughout. Figures cut out of the page stay in place within the text.

Read

Two panes: the photo on the left, the text on the right, original or translated with one click. An endless scroll across the whole repository, in the order the photos were taken. A click on the photo opens the original JPEG, at full resolution.

Organise

The fonds is organised as in an archive: a repository, shelfmarks, folios assigned once and for all. Pages from a trip can stay “to be sorted” then move in batches to their shelfmark. The shelfmark printed at the foot of the page is read, and suggested for the pages that follow.

What it costs in time

The upload runs at the speed of your connection: the tool sends three files at once and never stops to wait for a calculation. Transcription and translation happen on the server, in batches, within minutes or hours — you close the tab, they carry on. By the next morning, the fonds is readable.

How we choose the reading engine

The engine depends on the document and the task: each page is read by the model that scored best on our test bench for that type of document, handwriting goes through a reinforced second reading, and translation and the index each have their own model. We compared around thirty engines on the same reference archive photographs — typewritten, printed, Fraktur, handwritten, shot at an angle — measured the percentage of words recovered, and keep the best. A weekly watch flags each new model, which goes through the same bench before it replaces the one it outperforms.

Engines compared, all usable commercially:

  • Google Gemini (Vertex AI): 3 Flash, 3.1 Flash-Lite, 3.1 Pro, 3.5 Flash, 3.5 Flash-Lite, 3.6 Flash, 3.7 Flash, 3.8 Flash
  • Google Cloud Vision
  • Mistral OCR, Mistral Medium, Mistral Small (OVHcloud AI Endpoints)
  • Qwen (OVHcloud AI Endpoints)
  • OpenAI GPT-5.6, GPT-6
  • Anthropic Claude Haiku, Opus, Fable
  • Microsoft Azure Document Intelligence
  • Datalab (Marker/Surya)
  • Open models: olmOCR 2, GLM-OCR, dots.ocr, PaddleOCR-VL 1.5
  • Kraken, the engine behind eScriptorium, with the models: Party, PP-OCRv6, Manicule français moderne, german_print
  • PyLaia (Belfort, POPP, NewsEye), TrOCR large, Calamari (gt4histocr)

See it on your own pages

A hundred pages transcribed and translated, for a year, no card required.

Create my free account See pricing