Amazon Bedrock - Amazon Bedrock is a managed AWS service for building generative AI applications with foundation models through APIs.

Invoke models from Amazon and other providers without provisioning model-serving clusters. Check the selected model's supported modalities, Region availability, access requirements, and quotas before integrating it.

Use Converse and ConverseStream for a consistent conversational interface across supported models. Keep model-specific settings and capability checks explicit when switching providers or model versions.

Build retrieval-augmented generation workflows using supported data sources and retrieval stores. Tune chunking, metadata filters, and retrieval settings, and evaluate whether retrieved passages actually support the generated answer.

Configure agents to retrieve information and invoke defined application actions. Validate tool arguments and enforce authorization in the action implementation rather than relying on the model to enforce business rules.

Apply supported checks for harmful content, denied topics, sensitive information, and grounding. Test both false positives and missed violations, and retain application-level controls for security-critical decisions.

Customize supported models using available techniques such as fine-tuning or continued pre-training. Supported methods, training requirements, and deployment options vary by model and Region.

Use supported evaluation workflows to compare model outputs or assess retrieval and generation quality. Combine automated measurements with representative test cases and human review where correctness is subjective or consequential.

Choose among supported on-demand, batch, and provisioned inference options. Evaluate expected traffic, latency requirements, and model eligibility before selecting a capacity and billing approach.

Control API access with IAM and use supported private endpoints and encryption settings. Combine CloudWatch metrics, optional invocation logging, and CloudTrail records to investigate performance and access issues.

Connect an assistant to approved manuals, policies, and operational documentation using retrieval-augmented generation. Enforce document access restrictions and expose source references so users can check answers.

Generate draft responses, summarize conversations, and retrieve relevant support articles. Require human review or deterministic approval checks before sending sensitive responses or changing customer accounts.

Extract fields, classify documents, or summarize long reports using models that support the required input format. Validate extracted values against schemas and business rules before writing them to downstream systems.

Use agents or application-managed tool calls to check inventory, retrieve order status, or prepare service requests. Run actions under narrowly scoped permissions and require confirmation for irreversible changes.

Process eligible workloads such as catalog enrichment, document labeling, or large-scale summarization through batch inference. Use job tracking, output validation, and retry logic to handle partial failures.

Capabilities such as guardrails, evaluation, and some knowledge base processing operations can add charges beyond model inference. Include embedding generation, vector storage, S3, Lambda, logging, and data transfer where applicable in workload cost estimates.

Eligible models support asynchronous batch jobs with pricing separate from interactive inference. Use this option for workloads that can tolerate job completion latency, and account for associated storage costs.

Customization can incur training charges, storage charges for customized models, and charges for subsequent inference or required hosting capacity. Billing units and supported deployment options vary by model and customization method.

Pay for model usage without reserving dedicated capacity. Text models commonly charge separately for input and output tokens, while other modalities may use units such as images or audio duration; rates depend on the model and inference option.

For eligible models, purchase provisioned model capacity billed over time, with available commitment terms depending on the offering. Compare expected utilization and throughput needs against on-demand inference before committing.

AWS Step Functions

Use Step Functions to orchestrate Bedrock calls alongside validation, retries, and downstream processing. For example, a document-processing state machine can invoke a model, validate the returned JSON in Lambda, and route invalid results for review. Step Functions provides supported Bedrock service integrations, while other operations can be called through supported SDK integrations or Lambda.

Amazon CloudWatch

Use CloudWatch metrics to monitor invocation activity, latency, errors, and throttling. For example, alarm on sustained throttling and correlate it with application request volume before requesting quota increases. Optional model invocation logging can capture supported request and response data, so restrict access and set retention policies before enabling it.

AWS Lambda

Lambda functions can invoke Bedrock runtime APIs and implement agent action groups. For example, an order-status action can validate an order identifier, check the caller's authorization, and query an order system. Keep action permissions narrow and make write operations idempotent where retries are possible.

Amazon S3

Use S3 as a document source for supported knowledge base ingestion workflows or as input and output storage for batch inference. For example, upload approved product manuals to a dedicated bucket prefix and synchronize the knowledge base after updates. Scope bucket permissions to the relevant service role and configure encryption and retention for the data.