One thing that wasn't mentioned that works pretty well - if you have a RAG process running async rather than in a REPL loop, you can retrieve documents then per

One thing that wasn't mentioned that works pretty well - if you have a RAG process running async rather than in a REPL loop, you can retrieve documents then perform a pass with another LLM to do summarization/extraction first. This saves input token costs for expensive LLMs, and lets you cram more information in the context, you just have to deal with additional latency.