Support for Base64-Encoded Documents in ChatBedrock/ChatBedrockConverse

[itogaston]

Hello,

I'm evaluating the use of the ChatBedrock and ChatBedrockConverse classes and would like to know if they support sending documents encoded in base64. For context, with boto, we can send documents like this:

response = bedrock_runtime_client.converse(
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "document": {
                        "format": file_type,
                        "name": file_name,
                        "source": {"bytes": file_content},
                    }
                }
            ],
        }
    ],
    modelId="eu.anthropic.claude-3-5-sonnet-20240620-v1:0",
)

I haven't been able to locate any examples or documentation that show how to pass documents in this manner using either ChatBedrock or ChatBedrockConverse. Could you please confirm whether this functionality is supported? If it's not currently available, are there any plans to include it in the near future? Any additional guidance or example use cases would be highly appreciated.

Thank you in advance for your assistance!

[3coins]

@itogaston
Yes, ChatBedrockConverse does support sending documents. Here is some sample code taken from an earlier PR (#143) to fix this support. Let us know if this works for you.

from langchain_core.messages import HumanMessage
from langchain_aws import ChatBedrockConverse

llm = ChatBedrockConverse(
    model="eu.anthropic.claude-3-5-sonnet-20240620-v1:0"
)

# Convert document to bytes
doc_path = "documents/random-doc.pdf"
doc_bytes = None
with open(doc_path, "rb") as f:
    doc_bytes = f.read()

messages = [
    HumanMessage(
        content=[
            {
                "type": "document",
                "document": {
                    "format": "pdf",
                    "name": "random-doc",
                    "source": {
                        "bytes": doc_bytes
                    }
                }
            },
            "What is this document about?"
        ]
    )
]

llm.invoke(messages)