What causes "No message content blocks with 'text' key found in response" in huggingface/smolagents?

Check response['output']['message']['content'] and handle toolUse blocks before expecting text (use the model's tool-calling flow rather than plain generate)

huggingface/smolagents · error · KeyError

No message content blocks with 'text' key found in response

Error message

No message content blocks with 'text' key found in response

What it means

KeyError raised when parsing the Bedrock Converse response: response['output']['message']['content'] contained no content blocks with a 'text' key. This happens when the model only returned non-text blocks (e.g. toolUse, or only reasoning/thinking blocks) or the response shape differs from expectations.

Source

Thrown at src/smolagents/models.py:2046

        if response_format is not None:
            raise ValueError("Amazon Bedrock does not support response_format")
        completion_kwargs: dict = self._prepare_completion_kwargs(
            messages=messages,
            tools_to_call_from=tools_to_call_from,
            custom_role_conversions=self.custom_role_conversions,
            convert_images_to_image_urls=True,
            **kwargs,
        )
        self._apply_rate_limit()
        # self.client is created in ApiModel class
        response = self.retryer(self.client.converse, **completion_kwargs)

        # Get content blocks with "text" key: in case thinking blocks are present, discard them
        message_content_blocks_with_text = [
            block for block in response["output"]["message"]["content"] if "text" in block
        ]
        if not message_content_blocks_with_text:
            raise KeyError("No message content blocks with 'text' key found in response")
        # Keep the last one
        content = message_content_blocks_with_text[-1]["text"]
        if stop_sequences is not None and not self.supports_stop_parameter:
            content = remove_content_after_stop_sequences(content, stop_sequences)
        return ChatMessage(
            role=response["output"]["message"]["role"],
            content=content,
            tool_calls=response["output"]["message"]["tool_calls"],
            raw=response,
            token_usage=TokenUsage(
                input_tokens=response["usage"]["inputTokens"],
                output_tokens=response["usage"]["outputTokens"],
            ),
        )


AmazonBedrockServerModel = AmazonBedrockModel

View on GitHub (pinned to 30bb116109)

Solutions

Check response['output']['message']['content'] and handle toolUse blocks before expecting text (use the model's tool-calling flow rather than plain generate)
If thinking blocks are the cause, adjust model/inference config so a text block is produced
Upgrade smolagents — response parsing for mixed content blocks has been improved across versions
Retry the request: occasional empty completions from Bedrock can produce this

Defensive patterns

Strategy: try-catch

Try / catch

from smolagents.models import BedrockModel

try:
    msg = model.generate(messages, tools_to_call_from=tools)
except KeyError as e:
    if "text" in str(e):
        # model returned only tool-use/thinking blocks; inspect raw response or retry
        msg = model.generate(messages)  # retry without tools

Prevention

Prefer the agent's tool-calling flow over raw generate with tools on Bedrock
Pin a smolagents version tested with your Bedrock model
Add retry with backoff for occasional empty completions

When it happens

Trigger: Calling BedrockModel.generate with tools_to_call_from where the model responds solely with a tool call block; or with extended thinking/reasoning enabled so only thinking blocks are returned; or an unexpected/empty response body.

Common situations: Using Claude models with tool use on Bedrock; enabling thinking/reasoning configuration; model returning an empty or truncated response; mismatches between Bedrock API versions and response schemas.

Related errors

AI-assisted analysis of huggingface/smolagents@30bb116109 (2026-08-28). Data as JSON: /api/errors/ccf6cd2d69012a80. Report an issue: GitHub.