The content recognition is incomplete.

[IceBear-huang]

When using the model to recognize Markdown content containing both tables and text in images, it only returns the table content. I deployed the project by cloning it from GitHub and building a Docker image. I'm currently using the model hosted_vllm/Qwen/Qwen2.5-VL-7B-Instruct-AWQ. Is this issue caused by incorrect prompting or an unsuitable model? Should I switch to hosted_vllm/Qwen/Qwen2.5-VL-32B-Instruct-AWQ instead?

[mandalsouvik3333]

Can you try the markdown model we have released? It should give better results. We have trained it explicitly for this kind of tasks. https://nanonets.com/research/nanonets-ocr-s/

[IceBear-huang]

How can I use nanonets-ocr-s in docext? Do I simply specify it using the --model_name parameter, and if so, what should the exact model name be?

[mandalsouvik3333]

Yeah. python -m docext.app.app --model_name hosted_vllm/nanonets/Nanonets-OCR-s

You can check the full documentation here

[IceBear-huang]

hello，Could you please provide me with an explanation?

[IceBear-huang]

Does Docext only provide a web interface to access Nanonets-OCR-s? If I only need to call the deployed Nanonets-OCR-s via API, can I skip deploying Docext altogether?

[IceBear-huang]

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

<img>SOFT BULLET GUN</img>

Why does the Nanonets-OCR-s model keep outputting the same label repeatedly for certain images? This significantly affects efficiency, as it's too slow and inaccurate. Here's the image link https://oss.tradeunity-ru.com/seller/7/shop/goods_img/product_Ji2xniFCF7.jpg
. Is it due to poor Chinese character recognition?

[mandalsouvik3333]

Its mainly trained on documents.

[IceBear-huang]

Is the image recognition effect not good? If the image is a screenshot of an Excel or Word document, will the effect be better? So far, I have tried using images of Excel table types and haven't found any issues.

[mandalsouvik3333]

If the image is a screenshot of an Excel or Word document

this will work because it will be similar to a document. But the one you have shared is not like document.