# What should I do if Amazon Bedrock returns an error message or error code in response to a model invocation request?

Source: https://zilliz.com/ai-faq/what-should-i-do-if-amazon-bedrock-returns-an-error-message-or-error-code-in-response-to-a-model-invocation-request
Author: zilliz

If Amazon Bedrock returns an error during a model invocation, start by analyzing the error code and message to identify the root cause. Errors typically fall into four categories: client-side issues (4xx), rate limits (429), server-side problems (5xx), or model-specific constraints. For example, a `400 BadRequestException` often indicates invalid input parameters, while a `429 TooManyRequestsException` signals throttling. Parse the error response to extract details like `error_code`, `message`, and `model_id`, which help narrow down the issue. Cross-reference this with Bedrock’s documentation for model-specific requirements, such as input formats or token limits.

Next, address common causes systematically. For client errors (4xx), validate the request structure. Ensure parameters like `modelId`, `contentType`, and `body` align with the target model’s API specifications. For instance, Anthropic’s Claude requires a `prompt` field in a specific format, while Amazon Titan uses `inputText`. Verify IAM permissions: the invoking role must have `bedrock:InvokeModel` access, and the model must be enabled in your AWS account (via the Bedrock console). For throttling (429), implement retry logic with exponential backoff. For example, use AWS SDK auto-retry configurations or a library like `retry` in Python. If the error is server-side (5xx), check the AWS Service Health Dashboard for outages and retry after a delay.

If basic fixes don’t resolve the issue, debug using tools like the AWS CLI or CloudWatch Logs. Run `aws bedrock invoke-model ...` with the `--debug` flag to inspect raw request/response data. Check CloudWatch for execution logs if the invocation is part of a Lambda function or other AWS service. Test with simplified inputs to isolate the problem—for example, reduce prompt length to test token limits. If the error persists, contact AWS Support with details like the exact error code, timestamp, and a minimal reproducible request. For model-specific quirks (e.g., guardrail violations), consult the model provider’s documentation or community forums.