ocr not extracting the text from the image

[whole ocean]

Hey,

I have a pipeline for parsing documents, and I'm testing Mistral OCR, but the result output is a bit weird.

I attached the image, that's the output from mistral OCR

# Company structure 

BD is structured to serve customers by providing unique solutions. The data below represents the company structure for FY 2019.

## Revenue by geography

(millions of dollars)
![img-0.jpeg](img-0.jpeg)

## Revenue by segment

(billions of dollars)
![img-1.jpeg](img-1.jpeg)
![img-2.jpeg](img-2.jpeg)
![img-3.jpeg](img-3.jpeg)

Values in this exhibit reflect rounded numbers in billions and include Bard.
![img-4.jpeg](img-4.jpeg)

[stable pasture]

Seeing the same thing in our testing; and it makes the API rather unusable for visually complex documents.

I'd expect the API to also return the extracted text fro the images it identifies w/o requiring recursive calls to OCR those images.

Not sure if this was intentional by Mistral, or an oversight?

[spiral sparrow]

Good solution to use the recursive call ! But indeed it will be pretty good to have directly the text when include_image_base64=False or with another param

In the meanwhile I'm using pixtral-large-latest by converting PDF into images

"Not sure if this was intentional by Mistral, or an oversight?" => It's seem to be intentional by reading the structured_ocr cookbook (https://colab.research.google.com/github/mistralai/cookbook/blob/main/mistral/ocr/structured_ocr.ipynb). There is a replace_images_in_markdown function which is used after getting the response

[finite sphinx]

im seeing the same thing where half the image isnt being OCR'ed and jsut classified as <image_1> etc.

[finite karma]

Same on my side. Any fixes?

[cold linden]

same. Do devs see an option to have that image incuded in parsing? location of it can affect quality of parsing.

[finite karma]

As example, I used that image without markdown in response.

[cold linden]

another problem which i see, is that i need to host somewhere this file, to be able to process it. It's fine but quite annoying tbh 😅

[cold linden]

problem in that case can be that this is "audio/mpeg", instead of image/

nvm, that looks similar to this problem, but at this case it's at least possibilityto process it after first "reprocessing"

[queen wave]

Same problem. Text extraction on Images with API (type: "image_url") is not working in almost all cases, however if I input the same image on the chat interface (lechat), i get the text

[river junco]

Same problem for me, also with pdf files

[south willow]

Facing the same issue, support ignoring this