Manage the context window in an applied LLM system

An applied AI product keeps hitting the model's context limit. How do you decide what goes into the context on each call? Discuss retrieval, trimming and summarizing history, ordering, the cost and latency of long contexts, and how you would test that answers did not get worse.