AWS Bedrock — Library index
Canonical reference for Amazon Bedrock — the AWS-managed gateway to Anthropic Claude, Meta Llama, Mistral, Cohere, AI21, Amazon Nova / Titan, and Stability AI, plus the value-added stack on top: Converse API, Bedrock Agents, Knowledge Bases for RAG, Guardrails, Prompt Management, Flows, and the cross-region inference profile machinery. House style and source URLs match the rest of the Compute library; this sub-library is the AWS-specific counterpart to
Claude/,OpenAI/, andGoogleGemini/(the model-provider sub-libraries).
Bedrock sits at an awkward but useful place in the AI stack — it is the only major hyperscaler offering that exposes third-party foundation models behind a uniform AWS-IAM-controlled API. The cost is real (per-token markup vs the model provider’s own API, less feature parity, lag on new releases — Anthropic’s newest Claude lands on the Anthropic API days-to-weeks before Bedrock), but the wins are also real: IAM-native auth, VPC-native networking, KMS-native encryption, CloudWatch-native observability, and a single bill. If your org is already deep into AWS, Bedrock is the path of least resistance to production LLMs even if the per-token price is 5-10% higher than calling the provider directly.
This sub-library lives inside Math-and-Compute/Compute/ as a sibling of Claude/, OpenAI/, GoogleGemini/, Vercel/, and CloudflareWorkers/. Notes are version-pinned to 2026-05-25 — Bedrock evolves quickly (new model families and feature releases on a roughly-quarterly cadence at AWS re:Invent and a smaller “summer release” in July), so verify model IDs, pricing, and regional availability against docs.aws.amazon.com/bedrock/ before quoting.
See also
- _index — the model-provider deep dive for Claude (Bedrock exposes Claude but is not the canonical reference for it)
- prompt-engineering-agent-systems — vendor-neutral discipline overview
- rag-embeddings-vector-search — pairs with bedrock-knowledge-bases-rag for RAG architecture
- observability-stack — CloudWatch is the native observability target for Bedrock invocations
- auth-authz — IAM is how every Bedrock call is authorized
Library contents
Foundations
| File | What it covers |
|---|---|
| bedrock-models-and-providers | The full model catalog — Anthropic Claude (Opus 4/4.5/4.7, Sonnet 4/4.5/4.6, Haiku 3.5/4.5), Meta Llama 3.1/3.2/3.3/4, Mistral Large 2, Cohere Command R/R+, AI21 Jamba, Amazon Nova (Pro/Lite/Micro/Canvas/Reel), Amazon Titan, Stability AI SD3 — pricing per million tokens, regional availability matrix, cross-region inference profiles |
| bedrock-api-and-converse | The two API surfaces: legacy InvokeModel / InvokeModelWithResponseStream (model-specific request body) vs the modern Converse / ConverseStream (unified message format, unified tool spec, unified guardrails). Python boto3 + aws-sdk-js-v3 examples |
Higher-level products
| File | What it covers |
|---|---|
| bedrock-agents | Bedrock Agents — action groups (Lambda or schema-based), session state, agent memory (long-term + short-term), multi-agent collaboration (supervisor + collaborators), agent traces, advanced prompts |
| bedrock-knowledge-bases-rag | Knowledge Bases for RAG — vector store options (OpenSearch Serverless / Pinecone / Redis Enterprise / Aurora pgvector / MongoDB Atlas / Neptune Graph), chunking strategies, hybrid vs semantic search, Cohere Rerank 3 integration, ingestion pipelines |
| bedrock-guardrails | Guardrails — content filters (sexual / violence / hate / insults / misconduct / prompt-attack), denied topics, word filters, sensitive-info filters (PII redaction), contextual grounding check, automated reasoning checks |
| bedrock-prompt-management-and-flows | Prompt Management (versioned prompts with variables) + Bedrock Flows (visual workflow builder linking prompts, agents, knowledge bases, and Lambda) |
Operational
| File | What it covers |
|---|---|
| hidden-tricks-and-gotchas-aws-bedrock | Cross-region inference profiles for higher quotas, application inference profiles for cost tagging, hybrid-search nuances, agent memory size limits, guardrail latency overhead, model-specific tool-call format quirks, provisioned throughput vs on-demand crossover, PrivateLink VPC routing |
How to use this library
- Picking a model for a new project — start at bedrock-models-and-providers for the catalog, then bedrock-api-and-converse for the call shape, then hidden-tricks-and-gotchas-aws-bedrock for the cross-region quota tricks.
- Building a RAG application — bedrock-knowledge-bases-rag is the managed path; rag-embeddings-vector-search is the from-scratch path.
- Building an agent — bedrock-agents is the managed orchestrator; the Converse API with tool use in bedrock-api-and-converse is the lower-level building block.
- Compliance / safety — bedrock-guardrails covers what the managed safety layer can do.
- Cost optimization — pricing tables in bedrock-models-and-providers + cross-region profiles in hidden-tricks-and-gotchas-aws-bedrock + Batches / Provisioned Throughput tradeoffs.
Conventions used in this library
- Model IDs are quoted verbatim (e.g.
anthropic.claude-sonnet-4-6,meta.llama3-3-70b-instruct-v1:0,amazon.nova-pro-v1:0). On-demand vs cross-region inference profile IDs differ — the latter are prefixed with the source region (e.g.us.anthropic.claude-sonnet-4-6). - Pricing is in USD per million tokens (MTok) on-demand on-demand in the US East (N. Virginia) region as of 2026-05-25. Where multiple
\$symbols appear on the same line they are backslash-escaped to avoid Obsidian MathJax interpretation. - API examples use Python
boto3and TypeScript@aws-sdk/client-bedrock-runtimev3. Both target thebedrock-runtimeendpoint for inference andbedrock-agent/bedrock-agent-runtimefor Agents and Knowledge Bases. - “Per docs.aws.amazon.com 2026-05” tags a claim sourced from official Bedrock documentation as of the file’s
createddate. - Region matrix snapshots are us-east-1, us-west-2, eu-central-1, eu-west-3, ap-northeast-1, ap-southeast-2 unless otherwise noted. Bedrock now operates in 18+ regions; consult the per-model availability table on the AWS docs for current state.
Sources read while building this library (2026-05-25 pass)
docs.aws.amazon.com/bedrock/latest/userguide/models-supported.htmldocs.aws.amazon.com/bedrock/latest/userguide/conversation-inference.html(Converse API reference)docs.aws.amazon.com/bedrock/latest/userguide/agents.html+ sub-pagesdocs.aws.amazon.com/bedrock/latest/userguide/knowledge-bases.htmldocs.aws.amazon.com/bedrock/latest/userguide/guardrails.htmldocs.aws.amazon.com/bedrock/latest/userguide/prompt-management.htmldocs.aws.amazon.com/bedrock/latest/userguide/flows.htmldocs.aws.amazon.com/bedrock/latest/userguide/inference-profiles.htmldocs.aws.amazon.com/bedrock/latest/userguide/cross-region-inference.htmldocs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.htmldocs.aws.amazon.com/bedrock/latest/userguide/tool-use.htmlaws.amazon.com/bedrock/pricing/github.com/aws/aws-sdk-python(boto3bedrock,bedrock-runtime,bedrock-agent,bedrock-agent-runtimeclients)github.com/aws/aws-sdk-js-v3(@aws-sdk/client-bedrock-runtime,@aws-sdk/client-bedrock-agent-runtime)
Where official docs are silent but behavior is observable (cross-region failover specifics, undocumented header behavior in the runtime, exact CloudWatch metric dimensions), entries are labelled “Undocumented as of 2026-05-25 — observed behavior” so a future reader knows to verify.