Developer builds same RAG assistant three ways to expose what each architecture layer does
A developer built AutoSpec, a technical Q&A assistant for three fictional vehicle models, using three different Retrieval-Augmented Generation architectures on Amazon Bedrock to understand what each layer of abstraction handles. The first version used a fully managed Bedrock Knowledge Base with a Lambda function calling only the retrieve API, confirming semantic search worked correctly at zero generation cost. The second version manually wired Amazon Titan Embeddings, ChromaDB, and Amazon Nova Lite, revealing how prompt-level guardrails — not vector search — control which source data the LLM uses when multiple documents are retrieved together. The third version migrated orchestration to LangChain and swapped AWS services for Google Gemini API, successfully validating that business logic could survive a cloud provider change without a full rewrite. The project highlights practical trade-offs between managed convenience, fine-grained control, and multi-cloud portability in RAG system design.
