My summarization creation functions over batches of 50 messages, and I don’t often lose important context any more. The loss comes from trying to stick a whole conversation in a single compaction request (at least in my case)