Claude reads PDF using Visual Mode even if citations are disabled

[cedric-fauth]

I wanted Anthropic to handle PDFs parsing for my Agent. I followed the official docs.

I wanted to use Text extraction only ("Converse Document Chat (Original mode - Text extraction only)") which is Automatically used when citations are not enabled without visual mode ("Claude PDF Chat (New mode - Full visual understanding)")

So I disabled citations in each document object. This is what my tracing looks like:

{
    "role": "user",
    "content": [
        {
            "type": "document",
            "source": {
                "type": "base64",
                "media_type": "application/pdf",
                "data": "@@@langfuseMedia:type=application/pdf|id=HX8_5o4Ap6brS1uXGy89pK|source=base64_data_uri@@@"
            },
            "title": "root/DE/nature/nature_park/natureparkdata.pdf:1-1",
            "context": "You have access to this file until a system-reminder tells you otherwise.",
            "citations": {
                "enabled": false
            }
        },
        {
            "type": "text",
            "text": "what is this pdf about? use text extraction only. not visual mode."
        }
    ]
}

However with a completely "scanned pdf" (just one image) without OCR I get this response:

The user is asking me to extract text from a PDF file. Looking at the document content provided, I can see the text content that was already extracted from the PDF

The tracing shows, that all text was extracted from the image in the pdf.

[
    {
        "id": "toolu_01CW9ShJRYNpPzR7424jPjHv",
        "input": {},
        "name": "think",
        "type": "tool_use",
        "index": 0,
        "partial_json": {
            "thought": "The user is asking me to extract text from a PDF file. Looking at the document content provided, I can see the text content that was already extracted from the PDF. Let me analyze what's in there:\n\nFrom the PDF content:\n- Title: \"Map Preview\"\n- Text: \"No files available\"\n"
        }
    }
]

It seems like Anthropic extracted all data from the image using visual mode. This also corresponds to token usage which is quite high (too high for just reading text). I would like to know how I can prevent this from happening. It looks like disabling citations doesn't do the trick.

[marwq]

PDF is already an image like format, so there's no way to extract data from PDF without OCR. It will do OCR even in without-visual-understanding mode. Difference is it wouldn't recognize illustrations/images/graphs in the file.