tested gemini 3.1 pro's huge context vs claude on a real client rag, didnt go how i expected

so client has this messy 400 page policy corpus and i figured lemme just dump the whole thing in gemini 3.1 pro since the context is massive and skip half my retrieval layer. worked kind of?? recall was actually great for "is X mentioned anywhere" type stuff. but the second the question needed it to reason across two sections far apart it started dropping the earlier one, classic lost in the middle just at a bigger scale. claude with an actual retrieval step and way less context in the window was more consistent on the multi hop ones. so big context didnt kill my rag it just moved where the pain is. anyone gotten the giant context to actually hold multi hop or is chunking still the move