crewAIInc/crewAI · error · ValueError

Bedrock does not support URL references - resolve to bytes f

Error message

Bedrock does not support URL references - resolve to bytes first

What it means

The Bedrock formatter raises ValueError('Bedrock does not support URL references - resolve to bytes first') when the resolved file is a UrlReference. The Bedrock Converse API only accepts inline bytes or S3 locations; it cannot fetch arbitrary public URLs, so the library refuses UrlReference rather than sending a request Bedrock would reject.

Source

Thrown at lib/crewai-files/src/crewai_files/formatting/bedrock.py:83

        Returns:
            Content block dict or None if not supported.
        """
        content_type = file.content_type

        if isinstance(resolved, FileReference):
            if not resolved.file_uri:
                raise ValueError("Bedrock requires file_uri for FileReference (S3 URI)")
            return self._format_s3_block(content_type, resolved.file_uri, name)

        if isinstance(resolved, InlineBytes):
            return self._format_bytes_block(content_type, resolved.data, name)

        if isinstance(resolved, InlineBase64):
            file_bytes = base64.b64decode(resolved.data)
            return self._format_bytes_block(content_type, file_bytes, name)

        if isinstance(resolved, UrlReference):
            raise ValueError(
                "Bedrock does not support URL references - resolve to bytes first"
            )

        raise TypeError(f"Unexpected resolved type: {type(resolved).__name__}")

    def _format_s3_block(
        self,
        content_type: str,
        file_uri: str,
        name: str | None,
    ) -> dict[str, Any] | None:
        """Format block with S3 location source.

        Args:
            content_type: MIME type.
            file_uri: S3 URI.
            name: File name for documents.

View on GitHub (pinned to 754d7323be)

Solutions

1. Resolve/download the URL to bytes before formatting: use a resolver mode that produces InlineBytes/InlineBase64 instead of UrlReference.
2. Or download the file to S3 and pass FileReference(file_uri="s3://...") which Bedrock supports.
3. In a shared pipeline, branch on the target provider: keep UrlReference for OpenAI-style providers, force byte resolution for Bedrock.
4. Pre-fetch the URL yourself (requests.get(...).content) and attach the raw bytes.

Example fix

# before
resolved = UrlReference(url="https://example.com/doc.pdf", content_type="application/pdf")
block = bedrock_formatter.format_block(resolved)  # raises

# after
import requests
resp = requests.get("https://example.com/doc.pdf", timeout=30)
resolved = InlineBytes(data=resp.content, content_type="application/pdf")
block = bedrock_formatter.format_block(resolved)

if isinstance(resolved, UrlReference) and provider_uses_bedrock:
    resolved = InlineBytes(data=fetch(resolved.url), content_type=resolved.content_type)

def needs_resolution_for_bedrock(resolved: ResolvedFileType) -> bool:
    return isinstance(resolved, UrlReference)

try:
    block = bedrock_formatter.format_block(resolved)
except ValueError as e:
    if "URL references" in str(e) and isinstance(resolved, UrlReference):
        data = requests.get(resolved.url, timeout=30).content
        block = bedrock_formatter.format_block(InlineBytes(data=data, content_type=resolved.content_type))
    else:
        raise

When it happens

Trigger: Attaching a file by URL (UrlReference with url="https://...") to a crew whose LLM is a Bedrock model; a resolver configured to keep remote files as URL references (no download) combined with BedrockFormatting.

Common situations: User pastes a public https URL for an image/PDF and assumes every provider can read it; a pipeline shared across OpenAI (which accepts URLs) and Bedrock agents; resolvers upgraded to lazily pass URLs through to avoid downloading.

Related errors

• Bedrock requires file_uri for FileReference (S3 URI)
• URL scheme must be 'http' or 'https'
• Invalid URL scheme: {self.url}
• Gemini requires file_uri for FileReference
• Unsupported content type for Responses API: {content_type}

AI-assisted analysis of crewAIInc/crewAI@754d7323be (2026-08-15). Data as JSON: /api/errors/94e873bace9fc70d. Report an issue: GitHub.