← All questions
RAGMedium· Retrieval· Context Window· Accuracy

You raised top-k from 5 to 40, retrieval recall went up...

A strong answer separates retrieval from utilization. The passage is now in the prompt more often but used less effectively. You can confirm this by holding k fixed and injecting distractors; if accuracy falls while the gold passage remains present, it confirms a selection/utilization issue rather than a token overflow or model confusion. At 512-token chunks, 40 passages is only 20,480 tokens, which is well within the capacity of modern context windows.

Code playground (Python)