Try late chunking (stretch)

Late chunking needs one vector per token and each token's character offsets, which no embeddings API returns, so this runs an open model on your own machine. The fifth row only means something next to a control: the same model embedding the same chunks one at a time. Comparing it with the earlier rows would mostly measure the difference between two embedding models. The decision that matters most is the span mapping: from a chunk's blocks to its characters in the document text, then to the tokens whose offsets overlap them.