Semantic chunking raised our recall@5 by eight points a...

Refuse the framing that the eight-point gain is a success. Increasing the text per retrieved chunk by 2.8x will naturally raise recall without necessarily improving the quality of the boundaries. To validate the change, you need a controlled comparison, such as keeping the retrieved token budget constant or comparing against a fixed-size baseline at the new average length (610 tokens). Finally, perform a cost analysis: at k=5 and 10,000 queries per day, this length increase significantly impacts input token costs.