Hexagonal Architecture Makes RAG Pipelines Swappable Across AI Providers

A developer built a Retrieval-Augmented Generation (RAG) pipeline using hexagonal architecture to decouple core logic from specific technology providers. The project was inspired by a Machine Learning Mastery tutorial, where the developer noticed several APIs had already changed or been deprecated. By relying on interfaces, abstractions, and lazy imports, the design allows components like LLMs, vector stores, and document sources to be swapped without altering core business logic. The demo stack includes LangChain, Llama.cpp, Mistral.ai, Qwen Embeddings, PostgreSQL, Telegraf, and Prometheus, all running on Docker. The full implementation has been published on GitHub under the repository yoga1290/rag.