Some preliminary findings: You need a model with 8B parameters or more for effective RAG. Lighter models ignore context or hallucinate i.e. make up answers. Using the 8B models to embed is ...