Open oy321 opened 2 months ago
+1 here. Charts / any images for that matter are being hallucinated as Markdown tables. Specifying instructions doesn't help either. Are you guys passing all images to a VLM with conversion to MD table as a instruction?
JSON return type doesn't work now either.
We currently don't support this very well because understanding blocks of color in a PDF is a difficult problem. You should try the gpt4o mode as it may give you a more accurate result.
I am trying to parse this image but the result is very wrong:
even though I asked it specifically in the prompt to pay attention to the relationship between text and chart elements.