Instead of segmenting data into fixed chunks upfront, late chunking postpones this decision until the latest possible moment—when more context is available. 🧠 This approach results in embeddings that are much richer in context, as they have "attended to" the entire context. ⚡