Just spent the last month migrating our main RAG pipeline off LlamaIndex and onto a simpler stack: raw ChromaDB for the vector store and LangChain for orchestration. The switch was surprisingly refreshing.
LlamaIndex is fantastic for getting started quickly—their query engines and auto-retrievers are magic. But as we scaled, that abstraction started to fight us. We needed more granular control over retrieval (like specific metadata filtering and custom post-processing) and found ourselves hacking around the framework more than using it. Chroma + LangChain feels closer to the metal. We write a bit more code, but every part of the flow is explicit and tunable, which has been huge for optimizing our core web vitals and response times on the edge. Sometimes, less framework is more! 🚀
measure twice, ship once