Amazon Bedrock¶
Last reviewed: October 2026
Sources: AWS's official Amazon Bedrock documentation (docs.aws.amazon.com); verify against it before an interview. Content was rephrased for compliance with licensing restrictions.
What's changed recently
- The Converse API is the default interface. It is one request shape
across models, with tool use and streaming. Newer models are usually called
through cross-region inference profiles (IDs prefixed
us.,eu.,global.and so on) rather than single-region model IDs. - More ways to pay. Priority, Standard, and Flex service tiers (November 2025) sit alongside batch inference and provisioned throughput.
- Amazon Nova is now Amazon's first-party model family (successor to the
Titan text models). The catalog also includes OpenAI's open-weight
gpt-ossmodels alongside Claude, Llama, Mistral, DeepSeek and others. - Guardrails added Automated Reasoning checks (GA August 2025). Knowledge Bases can store vectors in Amazon S3 Vectors (GA December 2025) and use Neptune Analytics for GraphRAG.
- Agents split in two. Bedrock Agents is the managed, configuration-driven option. AgentCore (GA October 2025) is the runtime for agents you build with your own framework.
Bedrock is AWS's managed service for foundation models — call multiple providers (Anthropic Claude, Meta Llama, Amazon Nova, Mistral, DeepSeek, etc.) through one API, with enterprise security, and add RAG, agents, and guardrails.
What Bedrock provides¶
flowchart TB
APP[Your app] --> API[Bedrock API]
API --> FM[Foundation models: Claude, Llama, Nova, Mistral, ...]
API --> KB[Knowledge Bases - managed RAG]
API --> AG[Agents - tool use / orchestration]
API --> GR[Guardrails - safety & PII]
KB --> VDB[(Vector store)]
| Feature | Purpose |
|---|---|
| Model access | One API across providers; no infra to manage |
| Knowledge Bases | Managed RAG: ingest → embed → retrieve |
| Agents | Multi-step tool use / orchestration |
| Guardrails | Content filtering, PII redaction, denied topics |
| Service tiers / batch | Priority, Standard, Flex tiers; batch inference for offline jobs |
| Provisioned throughput | Reserved capacity (and required for some custom models) |
| Custom models | Fine-tuning / distillation and Custom Model Import for open-weight models |
import boto3
rt = boto3.client("bedrock-runtime")
resp = rt.converse( # model-agnostic request/response shape
# A cross-region inference profile ID. Check the console for current models.
modelId="us.anthropic.claude-sonnet-4-5-20250929-v1:0",
messages=[{"role": "user", "content": [{"text": "Summarize RAG in 2 lines."}]}],
inferenceConfig={"maxTokens": 300},
)
print(resp["output"]["message"]["content"][0]["text"])
invoke_model with a provider-specific JSON body still works, but converse
lets you swap models without rewriting the payload.
Why enterprises pick Bedrock¶
Data stays in your AWS account, integrates with IAM/VPC/CloudWatch, no model hosting to manage, and you can swap models without rewriting the app.
Interview questions¶
What does Bedrock give you over calling an LLM API directly?
Managed multi-provider access under AWS security (IAM, VPC, logging), plus built-in Knowledge Bases (RAG), Agents, and Guardrails — no infrastructure.
How do Bedrock Knowledge Bases implement RAG?
They ingest your documents, chunk and embed them into a vector store, and retrieve relevant chunks at query time to ground the model — managed for you.
What are Guardrails for?
Policy enforcement: block denied topics, filter harmful content, and redact PII consistently across models.
Interview deep dive¶
60-second talking points¶
- "One API, many models, inside your AWS boundary." Swap Claude/Llama/Nova without rewriting; data stays in your account under IAM/VPC.
- "Knowledge Bases = managed RAG; Agents = managed tool use; Guardrails = policy." You assemble, AWS operates.
Scenario & system-design questions¶
Design an enterprise RAG assistant on AWS with governance.
Bedrock Knowledge Base (ingest → embed → vector store) for retrieval → Bedrock model for generation → Guardrails for PII/denied topics → front with API Gateway + Lambda; IAM roles, VPC endpoints, CloudWatch. No data leaves the account.
How do you choose/switch models cost-effectively?
Start with a smaller/cheaper model, measure quality on an eval set, upgrade only where needed. Use batch or the Flex tier for latency-tolerant work, Priority for latency-critical traffic, and provisioned throughput for steady high volume. Prompt caching cuts the cost of repeated prefixes. The Converse API makes swapping models low-effort.
What do Guardrails actually enforce?
Denied topics, harmful-content filters, and PII redaction applied consistently across models — a policy layer independent of the chosen model.
Pitfalls interviewers probe¶
- Thinking Bedrock hosts your fine-tuned model for free (it's managed FMs; custom work has its own path/cost).
- Skipping Guardrails on user-facing apps.
- Ignoring per-model token pricing differences.
Rapid-fire¶
| Q | A |
|---|---|
| Bedrock value vs raw API? | Managed multi-model under AWS security + KB/Agents/Guardrails |
| Knowledge Bases? | Managed RAG (ingest/embed/retrieve) |
| Guardrails? | Content filtering + PII redaction + denied topics |
| Steady high volume? | Provisioned throughput (or Priority tier for latency) |
| Cheap offline jobs? | Batch inference / Flex tier |
| Model-agnostic call? | Converse API (+ inference profile ID) |
How interviewers probe this¶
You're getting ThrottlingExceptions at peak. What are your options?
A strong answer covers: checking per-model, per-region quotas and requesting increases; cross-region inference profiles to spread load; retries with backoff and jitter; the Priority tier or provisioned throughput for critical traffic; moving offline work to batch or Flex; and a fallback model. It also names the trade-off: cross-region inference can process data in other regions within the geography, which matters for residency.
Bedrock Agents or AgentCore for a new agent?
Bedrock Agents is quick when a managed, configuration-driven agent with action groups and Knowledge Bases is enough. AgentCore fits when you build with a framework (LangGraph, Strands, CrewAI), need custom orchestration, MCP tools through Gateway, enterprise identity, or Policy controls. Teams often prototype on one and standardize on the other.
Security asks where prompts and outputs go. What do you tell them?
Bedrock doesn't use your prompts or outputs to train the base models, and traffic can stay on PrivateLink VPC endpoints with IAM-scoped access and KMS encryption. Model invocation logging to S3 or CloudWatch is opt-in, and you secure it like any sensitive log. Cross-region inference and the third-party model provider's terms are the details to check.
How do you pick between Knowledge Bases and building RAG yourself?
Knowledge Bases gives managed ingestion, chunking, embeddings, a choice of vector store (OpenSearch Serverless, Aurora pgvector, S3 Vectors, and others), reranking, and metadata filters with little code. Build it yourself when you need custom chunking or retrieval logic, hybrid scoring you control, or non-AWS components, and accept owning the pipeline and its evals.