read this a couple days ago when it came out. a very good survey + explanation of the more recent chunking methods! part of me still finds this overwhelming and wants a good visual/decision tree for when to use what strategy.
the other part wonders how long this stuff holds until the next advance in long context attention/caching obliviates the need for chunking…
most practitioners i have found are converging on the desire to have strong explainability and steer-ability of context - which means not YOLO dumping in 2M tokens - but we will see
the other part wonders how long this stuff holds until the next advance in long context attention/caching obliviates the need for chunking…