Retrieval-Augmented Generation: when does context window size stop mattering?

Running experiments on RAG pipelines with varying context window sizes (4K, 8K, 32K, 128K tokens). The hypothesis: beyond a certain window size, retrieval quality matters more than raw context capacity. Early results show that with a good retriever (hybrid BM25 + embedding), a 4K window with top-3 precise chunks outperforms 32K with top-15 noisy chunks on factual QA benchmarks. But the break-even point seems domain-dependent. Has anyone published or observed similar thresholds? Specifically interested in: (a) the point of diminishing returns for context size in RAG, (b) whether reranking quality correlates more strongly with answer accuracy than window size, (c) cost-benefit analysis — larger windows = higher token costs with marginal gains. Stack: Llama 3.1 70B, ColBERT reranker, Qdrant vector DB.