AWS Bedrock

Provides a single standardized API endpoint to invoke foundation models from six different vendors (Anthropic Claude, Meta Llama, Mistral, Cohere, Stability AI, Amazon Titan) without requiring separate API keys, authentication flows, or vendor-specific SDKs. Bedrock abstracts vendor differences through a unified request/response schema, allowing developers to switch models or run multi-model inference with minimal code changes. Authentication is handled via AWS IAM, integrating with existing AWS identity infrastructure.

Enables creation of enterprise knowledge bases that automatically chunk, embed, and index documents (PDFs, web content, structured data) using Bedrock's managed embedding models, then retrieves relevant context during inference to augment LLM prompts. The system handles vector storage, similarity search, and context injection without requiring separate vector database infrastructure. Supports hybrid retrieval combining semantic similarity with metadata filtering.

Provides built-in tools and best practices for prompt engineering, including prompt templates, variable substitution, and prompt versioning. Enables testing multiple prompt variations against a dataset to measure performance differences. Integrates with model evaluation framework to quantify impact of prompt changes. Supports prompt chaining (multi-step prompts) and dynamic prompt generation based on context.

Implements end-to-end encryption for all data processed through Bedrock. Data in transit is encrypted using TLS 1.2+ (HTTPS). Data at rest is encrypted using AWS KMS (Key Management Service) with customer-managed keys (CMK) or AWS-managed keys. Supports encryption of knowledge base documents, fine-tuning datasets, and inference logs. Integrates with AWS CloudHSM for hardware-backed key management in highly regulated environments.

Implements AWS IAM-based access control for all Bedrock operations, enabling fine-grained permission policies at the action level (e.g., bedrock:InvokeModel, bedrock:CreateKnowledgeBase) and resource level (specific models, knowledge bases). Supports resource-based policies, cross-account access, and temporary credentials via STS. Integrates with AWS Organizations for centralized policy management across multiple AWS accounts.

Provides two agent frameworks: Amazon Bedrock Agents (guided, lower-code) and Amazon Bedrock AgentCore (flexible, framework-agnostic). Agents decompose user requests into multi-step reasoning chains, dynamically invoke tools (APIs, Lambda functions, databases), interpret results, and iterate until reaching a goal. Built on ReAct (Reasoning + Acting) pattern with native support for function calling via OpenAI-compatible schema format. Handles tool invocation orchestration, error recovery, and context management across steps without requiring manual prompt engineering.

Implements configurable guardrails that intercept model inputs and outputs to block harmful content, enforce compliance policies, and validate response accuracy. Uses automated reasoning checks (symbolic logic, pattern matching, and LLM-based classification) to identify policy violations before responses reach users. Supports custom guardrail policies (e.g., 'block financial advice', 'redact PII', 'enforce brand voice'). Claims to block up to 88% of harmful content and identify correct responses with up to 99% accuracy using multi-stage filtering.

Enables fine-tuning of select Bedrock models (Claude, Llama) using custom training data to adapt models to domain-specific tasks, terminology, or style. Handles data preparation, training orchestration, and deployment of fine-tuned models as new Bedrock endpoints. Supports both supervised fine-tuning (SFT) for task adaptation and continued pre-training for domain adaptation. Fine-tuned models are versioned and can be A/B tested against base models.

Supports streaming model responses token-by-token via bedrock:InvokeModelWithResponseStream API, enabling real-time output display in user-facing applications. Responses are streamed as a sequence of events (text chunks, metadata, stop reasons) over HTTP chunked transfer encoding or WebSocket connections. Allows applications to display partial results immediately rather than waiting for full response completion, improving perceived latency and user experience.

Provides tools to evaluate and compare model outputs across multiple Bedrock models using custom evaluation metrics, human feedback, and automated scoring. Enables A/B testing of models, prompts, and configurations on a dataset of test cases. Generates evaluation reports showing performance metrics (accuracy, latency, cost) and side-by-side output comparisons. Supports both automated evaluation (using reference answers or scoring functions) and human-in-the-loop evaluation.

Enables deployment of Bedrock inference within AWS VPC using VPC endpoints, ensuring model requests and responses never traverse the public internet. Supports both interface endpoints (for API calls) and gateway endpoints (for S3 data access). Integrates with AWS security groups, network ACLs, and IAM policies to enforce network-level access control. Audit logging via CloudTrail tracks all API calls with source IP, IAM principal, and request details.

Supports batch processing of multiple inference requests asynchronously, optimized for high-throughput, latency-tolerant workloads. Batch jobs accept a dataset of prompts, process them in parallel across Bedrock infrastructure, and return results in bulk. Pricing is typically discounted compared to on-demand inference. Useful for offline analysis, report generation, and bulk content creation where real-time response is not required.

Enables deployment of Bedrock models across multiple AWS regions with automatic failover if a region becomes unavailable. Applications can be configured to retry requests in alternate regions if the primary region fails. Supports both active-active (load balancing across regions) and active-passive (failover on failure) architectures. Integrates with AWS Route 53 for DNS-based routing and health checks.