OCR challenge
- Model
- paddleocr-vl-1.6OpenReads imagesOpen model, not part of Orion. Open-weight document OCR model: from one page image and a task prompt, it returns the page’s text, tables, formulas or chart data.
paddleocr-vl-1.6 is an open model that reads document pages. The page draws four made-up documents and sends each one as an image; the model returns the text it reads and the page compares it, field by field, with what the document says. Damage a document to see what still reads, or try an image of your own.
Live: this demo makes real calls to the API, with a demo key and a limit of calls per visit.
Pick a document and press Read the document: the model reads the image and the page compares what it read with the truth, field by field. Then damage the document (turn it, add noise or blur, fade it, change the font) and see what still reads.
No image chosen yet.
PNG or JPEG. Your browser scales it down to 1,600 pixels on its longer side and sends at most 1 MB.
Do not upload personal or sensitive data. The image is sent to the Serenity Edge API to answer and is not stored by the website.
No image chosen yet.
Damage applies to the four documents of the page, not to your own image.
Field by field
Press Read the document to see what the model reads.
Your image has no answer key, so there is no score: this is the text the model returned.
Near: at least 75 % alike once case, spaces and accents are set aside.
The answer reached the token limit and is cut.
Full text the model returned
Fields read
Read a document to see how many of its fields the model gets right.
- Invoice: not read yet
- Form: not read yet
- Table: not read yet
- Notice: not read yet
Summary
- Model calls
- Median per call
- Cost at scale
With the local simulation, no cost or milliseconds.
Next: Questions about an imageModel calls
The cost and the milliseconds appear with the first call.
Milliseconds of each call
In this preview nothing is sent to the API, so there is no cost or milliseconds to measure.
See the call
One request per reading to POST /v1/chat/completions: the page as an image, then a task prompt of those the documentation lists: Table Recognition: for the table, OCR: for the other documents and for your image.
- Route
POST /v1/chat/completions- Model
paddleocr-vl-1.6
The server adds the API key, which never reaches your browser.
In the cURL, the image is a placeholder: replace it with the data URL of your PNG or JPEG image.
Request
API response
Take it to your code
The same route reads scanned documents and answers questions about images, from your own code.
