All work
RAG
Serverless RAG on AWS
Document Q&A inside your own AWS account: hybrid search, Bedrock reranking and cited answers, all defined in CDK.
The problem
Companies want a document assistant inside their own cloud account, with no servers to babysit and answers they can verify.
How it works
S3 events drive idempotent ingestion: unchanged files are skipped by hash, re-uploads replace chunks atomically and deletions leave the index. Queries combine pgvector HNSW and PostgreSQL full-text search with Reciprocal Rank Fusion, rerank with the Bedrock Rerank API and answer with Claude using native citations. Follow-up questions are rewritten before retrieval.
Pipeline
- 01S3 upload
- 02Ingestion Lambda
- 03Hybrid search
- 04Bedrock rerank
- 05Claude + citations
Where it's used
- Companies that need data to stay inside their own AWS account
- Internal assistants over contracts, policies and technical docs
- Teams already on AWS that want a serverless, pay-per-use setup
Highlights
- Hybrid retrieval with RRF and graceful fallback when reranking is unavailable
- Character-level citations back to the source passage
- API keys, dead-letter queue, alarms, dashboard and X-Ray tracing
- Entire stack in AWS CDK, with an optional fully private network
Stack
AWS CDKLambdaBedrock · ClaudeTitan EmbeddingspgvectorAPI Gateway
Want something like this for your company?
I build production versions of these systems on your data and your infrastructure.
Related service: Knowledge assistants
Contact me