Case studyArchitecture, governance, and how to adapt this pattern in a pilot
Business use case
Problem
AWS-centric organisations want gen AI without exporting data to another cloud’s model API. Teams need a path from pilot (Converse + context) to production (Knowledge Bases on S3/OpenSearch).
Who benefits
- Data platform, Bedrock aligned with existing IAM and regions
- Application teams, one Converse API across foundation models
- Security, least-privilege IAM per workload
Success metrics
- Pilot on one domain with subject-matter expert grounding review
- IAM policy reviewed before any public-facing endpoint
- Clear upgrade path from Converse-only to Knowledge Base retrieval
Solution
Ask questions using a frontier model on AWS with sample documents in this demo, or connect an optional Bedrock knowledge base for live retrieval when you are ready for production.
Technical implementation
Stack
- Amazon Bedrock Converse API in your account
- Frontier model (e.g. Nova Lite) selectable per workload
- IAM credentials via execution role in production
Architecture
Hosted demo path: keyword retrieval from seed policies, then a Bedrock model via AI Gateway.
Implementation highlights
- Keyword retrieval from
lib/demos/seed-documents.tsbefore generation - Response includes
mode(gateway-bedrock-seed-rag) andretrievalhit metadata - Production upgrade: swap seed search for Bedrock Knowledge Base in your account
Outcomes and learnings
- Converse unifies message format across models, simplifies app code
- Knowledge Bases connect existing S3/OpenSearch without custom RAG in v1
- Keep blast radius small with per-demo IAM roles on public portfolios
Delivery playbookDiscovery → pilot → scale
- 1Discovery2–4 wks
Identify S3/OpenSearch corpora and IAM boundaries for Bedrock Knowledge Bases.
- 2Pilot6–8 wks
Converse + KB on one data domain; validate answer grounding with subject-matter experts.
- 3Scaleongoing
Roll out per-account Bedrock guardrails, KB versioning, and cross-region DR patterns.
Where else this appliesKnowledge-grounded Q&A in AWS is for teams whose data residency and procurement already point to Bedrock, not for every workload, but for the ones that must stay in-account.
Internal developer portal
Answer questions about internal APIs and IaC standards from repos synced to a KB.
Energy and utilities ops
Technicians query equipment manuals indexed per region with strict account isolation.
Financial research assistants
Analysts ask questions over approved research corpus, not the open web.
Post-merger integration
Unified KB over legacy and acquirer policy PDFs while IAM stays account-scoped.
Bedrock Knowledge Bases pair with S3/OpenSearch corpora; Converse gives a single API for multi-model trials inside the same IAM boundary.