SysPad › Lessons › AI and machine learning lessons › Amazon Bedrock

How Amazon Bedrock works

Serverless foundation-model inference (Claude, Llama, Titan…). No servers to size: billed per input/output token, throughput capped by the model RPM quota, latency driven by output length. Set token sizes + rates below.

You want to add a large language model to your product, but you do not want to rent GPUs or host a multi-gigabyte model. Bedrock hands you top foundation models, including Anthropic's Claude, behind one serverless API. You send a prompt, you get a response, and you pay by the token. Nothing to run.

What you will learn

Foundation models behind one API

Bedrock offers foundation models from several providers, including Anthropic Claude, Meta Llama, and others, through a single serverless API. There are no servers, GPUs, or model weights to manage: you call the API and a managed model responds.

You pay per token

On-demand, Bedrock bills per token, counting both the input (your prompt) and the output (the generated text). A token is roughly a short word-piece. So cost scales with how much text goes in and comes out, and longer prompts or responses cost more.

⚓ Chat assistants, summarisation and extraction, RAG over company docs, content generation, and agentic features inside SaaS products.

On-demand vs Provisioned Throughput

On-demand needs no commitment and suits variable or bursty traffic. Provisioned Throughput reserves dedicated model capacity for a steady, high volume at a predictable rate, and is required for some customised models.

Knowledge Bases, Agents, Guardrails

Bedrock is more than raw model calls. Knowledge Bases wire up retrieval (RAG) so a model can answer from your documents; Agents let a model call tools and APIs to complete multi-step tasks; Guardrails filter unwanted content and topics. They save you building those pieces yourself.

Flashcards

What does Bedrock provide, and what do you manage?
Foundation models (incl. Claude) behind one serverless API; you manage no infrastructure.
How is on-demand Bedrock billed?
Per token, counting both input (prompt) and output (generation).
On-demand vs Provisioned Throughput, when each?
On-demand for variable/bursty traffic, no commitment; Provisioned Throughput for steady high volume with reserved, guaranteed capacity.
What do Knowledge Bases, Agents, and Guardrails do?
Knowledge Bases = RAG over your docs; Agents = tool/API calling for multi-step tasks; Guardrails = content/topic filtering.
Bedrock vs SageMaker?
Bedrock = managed foundation models via API; SageMaker = build, train, and host your own models.

Sources

Open the SysPad canvas · Official Amazon Bedrock documentation

Other AI and machine learning lessons