Model Invocation

• Converse API Request Structure
• Streaming with ConverseStream
• InvokeModel (Provider-Specific)
• Cross-Region Inference
• Prompt Caching
• Service Tiers
• Prompt Management
• max_tokens Quota Mechanics
• Throttling & Retry Strategy

Converse API Request Structure

The Converse API is the unified interface. Key fields:

Content block types in messages:

Refer to the latest AWS documentation on Bedrock Converse API for supported content types and fields.

Security note: For workloads handling PII or sensitive data, use guardrailConfig to apply content filtering to both prompts and responses, and guardContent blocks to selectively evaluate only user input while excluding system prompts. See guardrails reference for configuration details and the PII logging compliance gap.

Streaming with ConverseStream

Events arrive in strict order:

stopReason values:

• end_turn — model finished naturally
• tool_use — model wants to call a tool, process toolUse blocks
• max_tokens — hit maxTokens limit, response may be truncated
• stop_sequence — model generated one of your custom stop sequences
• guardrail_intervened — a guardrail blocked the response, check trace for details
• content_filtered — model’s built-in safety filtered the response

Additional values exist for edge cases (malformed_model_output, malformed_tool_use, model_context_window_exceeded). Refer to the latest AWS documentation on Bedrock Converse stopReason for the full current list — new values are added as features launch.

InvokeModel (Provider-Specific)

Use InvokeModel ONLY for provider-specific features not available in Converse. For streaming with InvokeModel, use InvokeModelWithResponseStream — it returns the same provider-specific response format but as a stream. Each provider has a different request body format:

Anthropic Claude: anthropic_version required, messages format differs from Converse. Meta Llama: Uses prompt string with max_gen_len and temperature. Llama 2 uses [INST]...[/INST] prompt wrapping; Llama 3+ uses <|begin_of_text|><|start_header_id|>user<|end_header_id|>...<|eot_id|><|start_header_id|>assistant<|end_header_id|> special tokens. Amazon Titan: Uses inputText, textGenerationConfig. Amazon Nova: Uses Converse-compatible format but with Nova-specific parameters.

For detailed format examples, parameter names, and common mistakes per provider, see prompt engineering by model.

Refer to the latest AWS documentation on Bedrock InvokeModel for current request body formats per provider. The Converse API eliminates the need to know these formats for most use cases.

Cross-Region Inference

Model ID format determines how requests are routed:

• In-region (base model ID): e.g., anthropic.claude-3-haiku-20240307-v1:0 — single-region invocation, only for models with In-Region availability in your region
• Geo cross-region (inference profile): e.g., us.anthropic.claude-sonnet-4-6 — routes within a geography (US, EU, APAC). Required for many newer models, even for standard on-demand invocation
• Global cross-region (inference profile): e.g., global.anthropic.claude-sonnet-4-6 — routes to any commercial region where the model is available, for maximum throughput
• Provisioned throughput: ARN format arn:aws:bedrock:<region>:<account-id>:provisioned-model/<id>

Common errors from using the wrong ID format:

• Using a base model ID for a model without In-Region support: ValidationException: "on-demand throughput isn't supported" — use an inference profile ID instead
• Using a cross-region prefix from an unsupported source region: ResourceNotFoundException or AccessDeniedException

Verify the Correct ID format:

• For foundation models: `aws bedrock get-foundation-model --model-identifier`````
• For inference profiles: aws bedrock list-inference-profiles --region <region> - see Supported inference profiles (opens in a new tab)

Prompt Caching

Insert cachePoint blocks in the content to mark cache boundaries:

Placement rules:

• Place after large, reusable content (system prompts, few-shot examples, documents)
• Content before the cachePoint is cached; content after is not
• Supported on select models — refer to the latest AWS documentation on Bedrock prompt caching for current model support and availability
• Reduces latency and cost for repeated prompts with shared prefixes

Service Tiers

Set via serviceTier object in Converse API request: "serviceTier": {"type": "priority"}. If omitted, Bedrock routes to the Standard tier (API value "default").

Refer to the latest AWS documentation on Bedrock service tiers for current pricing, latency benchmarks, and model availability per tier.

Prompt Management

When using a managed prompt, pass the prompt ARN as modelId:

Critical restrictions when using managed prompts:

• MUST NOT include inferenceConfig — baked into the prompt definition
• MUST NOT include system — baked into the prompt definition
• MUST NOT include toolConfig — baked into the prompt definition
• MUST NOT include additionalModelRequestFields
• If you include messages, they are appended after the prompt’s messages, not replacing them
• promptVariables field: JSON with keys matching {{variable}} placeholders in the prompt
• Pin version in production: use :1 suffix, not DRAFT
• guardrailConfig still works — applied to the entire prompt + appended messages

max_tokens Quota Mechanics

Bedrock reserves quota at request start based on total input tokens (including cache read/write tokens) + max_tokens. Three stages:

1. Initial reservation: InputTokenCount + CacheReadInputTokens + CacheWriteInputTokens + max_tokens — determines if request is throttled
2. Dynamic adjustment: Bedrock releases unused reserved tokens as output is generated
3. Final settlement: InputTokenCount + CacheWriteInputTokens + (OutputTokenCount × burndown rate) — CacheReadInputTokens do not count toward final settlement

Burndown rate: Anthropic Claude 3.7+ models have a 5x burndown rate for output tokens — 1 output token = 5 quota tokens at settlement. All other models: 1x.

Impact of unset max_tokens (Claude Sonnet example): With 500 input tokens:

• max_tokens=1000: reserves 1,500 tokens → ~1,333 concurrent requests from 2M TPM
• max_tokens unset (defaults to model max): reserves based on model’s max output — e.g. 8,192 for Claude 3.5 Sonnet v2, up to 64K for Claude 3.7 Sonnet/4.x with extended thinking → as few as ~31 concurrent requests from 2M TPM
• Massive difference in concurrent capacity from one parameter (up to 43x with 64K models)

Right-size max_tokens to your expected output length. Use CloudWatch OutputTokenCount metrics to calibrate.

Model invocation logging: If model invocation logging is enabled, full prompts and responses are captured to CloudWatch Logs and/or S3. This is disabled by default but when enabled, logs contain complete text of every request and response. For PII-sensitive workloads: encrypt log destinations with KMS, restrict access, or disable invocation logging entirely. See the parent skill’s Critical Warnings section for the guardrails PII logging gap.

Throttling & Retry Strategy

Two types of 429 ThrottlingException:

• RPM (requests per minute): Too many requests. Quota refreshes on 60-second windows.
• TPM (tokens per minute): Too many tokens reserved. Affected by max_tokens (see above).

Use adaptive retry mode — it handles both types:

For sustained throttling:

• Right-size max_tokens (biggest impact)
• Check current limits: aws service-quotas get-service-quota --service-code bedrock --quota-code <code> --region <region>
• Request quota increase through AWS Service Quotas
• Consider provisioned throughput for predictable high-volume workloads
• Use batch inference for non-real-time processing (discounted pricing)