Unexpected citation format

[fatihyildiz-cs]

The response I get from Claude 4.5 references the citation documents with inline xml tags (<cite index=\\"1-17,1-18,1-19\\">) rather than populating the citations field in the TextBlock. Same problem was mentioned in #1031 and @pjuhos noted that it was fixed but I'm experiencing it with Anthropic 0.72.0 (without using the web fetch tool)

Here is the response i get (I instruct claude to respond with a json):

Message(
    id="msg_vrtx_01PuyNnGoNLoyfoxCfJVcS36",
    content=[
        TextBlock(
            citations=None,
            text='```json\n{\n  "result": [\n    {\n      "discussionPoint": "<cite index=\\"1-17,1-18,1-19\\">Heidi Alexander extended condolences on behalf of the Committee and Transport for London to the family of Pam Alexander, a former non-executive of the Crossrail Board between 2015 and 2018 and former member of the LLDC who chaired its planning committee, following her recent passing.</cite>",\n      "actionItems": []\n    }\n  ]\n}\n```',
            type="text",
        )
    ],
    model="claude-sonnet-4-5-20250929",
    role="assistant",
    stop_reason="end_turn",
    stop_sequence=None,
    type="message",
    usage=Usage(
        cache_creation=None,
        cache_creation_input_tokens=0,
        cache_read_input_tokens=0,
        input_tokens=10868,
        output_tokens=134,
        server_tool_use=None,
        service_tier=None,
    ),
)

Also, I see that the citation examples in the docs let Claude return free text rather than structured JSON. Is there anything particular I need to pay attention to when expecting a JSON response with citations? Thank you

[yashwantbezawada]

I did some investigating and I think I found what's going on here.

This looks like an API bug rather than an SDK issue. Here's what I found:

The SDK is working correctly

When you use streaming (client.messages.stream()), the API sends citations as structured data via citations_delta events, and the SDK handles them perfectly - they show up in the citations field like they should.

But with non-streaming (client.messages.create()), the API is sending back citations as XML-like tags embedded in the text instead of as structured JSON. The SDK just deserializes whatever the API sends, so it's not really an SDK bug.

Why the SDK can't parse the XML tags

I checked if the SDK could parse these tags client-side, but the citation format <cite index="1-17,1-18,1-19"> doesn't give us enough info to build proper Citation objects. We'd need to know what those numbers mean, which document they reference, what type of citation it is, etc. Without that info, any parsing would be incomplete or wrong.

Historical context

Looking at past issues:

• Issue The API returns citations that don't pass the API's own validation. #965 (June 2025) showed the API correctly returning structured citations as JSON
• Issue Getting citations in unexpected format when using web fetch tool #1031 (Sept 2025) reported the XML tag problem with web fetch tool - marked as fixed
• But your issue shows it's still happening without the web fetch tool

Seems like the fix for #1031 might have only addressed the web fetch case, but the general citation handling in non-streaming mode is still broken.

Bottom line

The API should be sending citations in non-streaming responses the same way it does in streaming - as structured JSON objects, not XML tags. This needs to be fixed on the API side.

For now, the workaround in the comments above (parsing the tags yourself) is probably your best option until the API team can fix it. Sorry there isn't a better answer!