i’ve been experimenting with building a fully local rag pipeline: weaviate for vectors + hybrid search, node.js scripts, qwen 3.5 on ollama

what i found is that most of the challenges live in retrieval and chunking, not the LLM, and a good chunking strategy + the right balance in hybrid search is more effective than using a bigger and more expensive model