// the find
aws-samples/aws-genai-llm-chatbot
A modular and comprehensive solution to deploy a Multi-LLM and Multi-RAG powered chatbot (Amazon Bedrock, Anthropic, HuggingFace, OpenAI, Meta, AI21, Cohere, Mistral) using AWS CDK on AWS
A CDK-based reference deployment for a multi-model RAG chatbot on AWS, covering Bedrock, SageMaker, OpenSearch, Aurora pgvector, and Kendra in one stack. It's aimed at teams that want a working enterprise chatbot scaffold without stitching together each AWS service from scratch. Expect to spend real time understanding what you're deploying before you trust it in production.
The CDK infrastructure is genuinely well-structured — separate constructs for RAG engines, auth, model interfaces, and API make it possible to pull out just the pieces you need. Multiple vector store backends (pgvector, OpenSearch, Kendra) are all supported with a consistent workspace abstraction, which is rare. Integration tests cover the full API surface including access control scenarios, not just happy paths. The SeedFarmer deployment wrapper is a nice touch for teams that need to manage multiple environment configurations without bespoke shell scripts.
LangChain is used as the glue layer for model interfaces, which means you're now on the hook for LangChain's own churn and deprecation cycle on top of AWS API changes — the adapter tree under lib/model-interfaces/langchain is already sprawling. The mixed Python/TypeScript split (CDK in TypeScript, Lambda handlers in Python) doubles the toolchain you have to maintain and makes cohesive refactors painful. The README is thin for a repo this large; the actual configuration surface is buried in magic-config.ts and default-config.json with minimal inline documentation. SageMaker self-hosted model support is clearly bolted on — the Llama2 and Falcon adapters look like one-off scripts that will break when those model conventions shift.