[Bug]: PDF as input type is not working for anthropic models via bedrock converse

[tanvithakur94]

Check for existing issues

• I have searched the existing issues and checked that my issue is not a duplicate.

What happened?

As per the docs: https://docs.litellm.ai/docs/completion/document_understanding, this payload should work:
https://docs.litellm.ai/docs/completion/document_understanding#base64

However, when we are trying to test out the same:

{
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "type": "file",
          "file": {
            // "file_id": "draconomicon.pdf",
            "file_data": "data:application/pdf;base64,JVBERi0xLjQKJf////amGUlRU9GCg=="
          }
        }
      ]
       }
  ]
}

I get error litellm.exceptions.BadRequestError: litellm.BadRequestError: BedrockException - {}

Steps to Reproduce

Shared above

Relevant log output

What part of LiteLLM is this about?

SDK (litellm Python package)

What LiteLLM version are you on ?

v1.80.10, v1.81.7

Twitter / LinkedIn details

No response

[krrishdholakia]

@jquinter can you try and validate this?

[jquinter]

Investigated this. The root cause is that the Bedrock Converse API requires a text content block alongside document blocks. Sending only a file block without any text prompt causes Bedrock to reject the request.

Reproducing the error (file-only payload, no text block):

import litellm

PDF_BASE64 = "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"

response = litellm.completion(
    model="bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0",
    messages=[{
        "role": "user",
        "content": [
            {"type": "file", "file": {"file_data": f"data:application/pdf;base64,{PDF_BASE64}"}}
        ],
    }],
)

Bedrock returns:

"A text block must be included when using documents. Add a text block and retry your request."

Working request (adding a text block):

response = litellm.completion(
    model="bedrock/us.anthropic.claude-3-5-sonnet-20241022-v2:0",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "What text does this PDF contain?"},
            {"type": "file", "file": {"file_data": f"data:application/pdf;base64,{PDF_BASE64}"}}
        ],
    }],
)
# Response: "The PDF contains only the text 'LiteLLM'"

This succeeds and correctly reads the PDF content.

Summary: This is a Bedrock API requirement — document blocks must be accompanied by at least one text block. The docs example should probably be updated to include a text block alongside the file block to avoid confusion.

@tanvithakur94 Could you try adding a {"type": "text", "text": "your prompt here"} entry to the content array alongside the file block and confirm this resolves your issue?

[jquinter]

@krrishdholakia from our side, we can work on improving error response parsing to surface actual error details instead of just {}, and even adding a pre-flight check or helpful error message when a document block is sent without any text block, since Bedrock requires both.

[tanvithakur94]

Thanks for checking this @jquinter. I tried with this payload:

{
  "messages": [
    {
      "role": "user",
      "content": [
        {"type": "text", "text": "What text does this PDF contain?"},
        {
          "type": "file",
          "file": {
            "file_data": "data:application/pdf;base64," "
          }
        }
      ]
       }
  ]
}

and still its just {} as the exception with status code 400.
I tried hitting bedrock directly as well with the transformed payload and it also gave me the same error {} and status code 400

{
    "messages": [
        {
            "role": "user",
            "content": [{"text": "What text does this PDF contain?"},{"document": {
                        "source": {
                            "bytes": ""
                        },
                        "format": "pdf",
                        "name": "DocumentPDFmessages_ae72287b9a3fede9_pdf"
                    }
                }
            ]
        }
    ],
    "additionalModelRequestFields": {},
    "system": [],
    "inferenceConfig": {}
}

[tanvithakur94]

One more quick question @jquinter : We used awscurl to directly hit bedrock and there also, we were getting no error message except {} with 400. Where did you find this exception message coming from?
"A text block must be included when using documents. Add a text block and retry your request."

[krrishdholakia]

unable to repro, validated locally as well