Troubleshooting
This section provides solutions to common issues when working with Amazon Nova models.
Authentication and setup
- Missing permissions
-
Symptoms: Unable to access Nova models or features
Solution:
Ensure your IAM role has AmazonBedrockFullAccess or appropriate permissions
Request specific model access through the Amazon Bedrock console
Verify permissions for model access and tool usage
Model access denied
- Symptoms: Model access request fails
-
Solution:
Request specific model access through the Amazon Bedrock console
Verify your account has been granted access to the requested model
Check regional availability of the model
Regional availability issues
- Symptoms: Feature not available in selected region
-
Solution:
-
Web Grounding is only available in US regions with US CRIS profiles
-
Verify the model and features are available in your selected region
-
Switch to a supported region if necessary
-
Timeout configuration
- Symptoms: Requests timing out before completion
-
Cause: Default timeout too short for complex operations
Solution: Configure extended timeout settings
from botocore.config import Config bedrock = boto3.client( 'bedrock-runtime', region_name='us-east-1', config=Config( read_timeout=3600 # 60 minutes ) )Note
Amazon Nova inference requests can take up to 60 minutes for complex operations.
API response issues
Understanding stop reasons:
- end_turn
-
Normal completion. No action needed.
- max_tokens
-
Token limit reached.
Solution: Increase
maxTokensparameter ininferenceConfig. - content_filtered
-
Content violated AWS Responsible AI policy.
Solution: Review and modify your input to comply with content policies.
- malformed_model_output
-
Invalid output format.
Solution: Check your output schema and constraints; verify JSON schema is properly formatted.
- malformed_tool_use
-
Invalid tool call format.
Solution: Verify tool definitions match expected schema; check tool input parameters are correctly formatted.
- service_unavailable
-
Built-in tool service unavailable.
Solution: Retry the request after a brief delay; check AWS service health dashboard.
- invalid_query
-
Invalid query to built-in tool.
Solution: Review query format and parameters; ensure query meets tool requirements.
- max_tool_invocations
-
Tool retries exhausted.
Solution: Simplify the task or break it into smaller steps; review tool error messages for specific issues.
Reasoning mode errors
- Truncated responses with high reasoning effort
-
Solution: For high reasoning effort, unset these parameters:
temperature,topP,maxToken. This allows the model to use optimal settings for complex reasoning tasks. - Insufficient tokens for reasoning
-
Error: "maxTokens is insufficient"
Solution: Automatically retry with increased limit
token_limits = { "low": 15000, "medium": 30000, "high": 50000 } try: response = client.converse( modelId="us.amazon.nova-2-lite-v1:0", messages=messages, inferenceConfig={ "maxTokens": token_limits[max_effort] }, additionalModelRequestFields={ "reasoningConfig": { "type": "enabled", "maxReasoningEffort": max_effort } } ) except Exception as e: if "maxTokens is insufficient" in str(e): higher_limit = int(token_limits[max_effort] * 1.5) # Retry with higher limit
Tool use issues
Schema validation failures
- Tool schema validation errors
-
Solution:
Limit JSON schemas to two layers of nesting for best performance
Ensure all required fields are properly defined
Validate schema against JSON Schema specification
- Model not using tools correctly
-
Solution:
Ensure tool name clearly describes its purpose
Provide detailed description of tool functionality
Explicitly define input schema with clear parameter descriptions
Include examples in the description when helpful
- Inconsistent tool calling behavior
-
Solution: Set temperature to 0 for tool calling:
inferenceConfig={ "temperature": 0, "maxTokens": 10000 }This enables greedy decoding for more reliable tool use.
- Tool choice conflicts
-
Problem: Error when using custom tools with web search or code interpreter
Solution: Do not include a custom toolSpec with name
nova_grounding- this conflicts with the system tool. Use the system tool configuration instead:# Correct - use system tool tool_config = { "tools": [{ "systemTool": {"name": "nova_grounding"} }] } # Incorrect - don't create custom tool with this name # tool_config = { # "tools": [{ # "toolSpec": {"name": "nova_grounding", ...} # }] # }
Web Grounding issues
- Access control problems
-
Problem: Web Grounding and Code Interpreter not working
Solution: Ensure your IAM policy includes:
{ "Statement": [ { "Effect": "Allow", "Action": ["bedrock:InvokeTool"], "Resource": ["arn:aws:bedrock::{YOUR_ACCOUNT_ID}:system-tool/amazon.nova_grounding"] } ] } - Service Control Policy issues
-
Problem: Web Grounding blocked by SCP
Solution: If you have Service Control Policies with
aws:requestedRegioncondition, update them to allow "unspecified" region for Web Grounding functionality.
Media processing limitations
- Poor understanding of multilingual content in images/videos
-
Limitation: Nova models have limited understanding of multilingual content in visual media
Workaround:
Provide text translations alongside images
Use text-based inputs for multilingual content when possible
- People identification
-
Problem: Model refuses to identify people in images
Expected Behavior: Models will refuse to identify or name individuals in images, documents, or videos for privacy and safety reasons
Workaround: Ask about general characteristics or context instead of specific identities
- Spatial reasoning limitations
-
Problem: Inaccurate localization or layout analysis
Limitation: Limited capabilities for precise spatial reasoning
Workaround:
Use bounding box detection for object localization
Provide clear reference points in your prompts
Break complex spatial queries into simpler components
- Small text in images/videos
-
Problem: Cannot read small text in media
Solution:
Crop images to focus on relevant text sections
Increase resolution of source media
Provide text separately if available
Document and file handling
- Unsupported content
-
Problem: PDF processing fails
Causes:
PDFs with CMYK color profiles
PDFs containing SVG images
Solution:
Convert PDFs to RGB color profile
Rasterize SVG images before including in PDFs
- Token estimation
-
Problem: Unexpected token usage with PDFs
Guideline: Estimate approximately 2,560 tokens per standard 8.5×11" PDF page
Solution: Adjust
maxTokensaccordingly based on document length