Yeah you probably wouldn't want the KV that big, with current model runtimes. Large, full context windows start to reduce quality and performance. Usually around 80% they start to degrade in my experience (i don't have a reference for quality vs ctx size).

RAG is probably best suited?