Comment by ajbd

is the bottleneck actually retrieval/search, or just that the underlying data is unstructured to begin with? have you seen a setup where context and decision don't get lost in threads over time, without relying on constant summarization?