Late Chunking: Balancing Precision and Cost in Long Context Retrieval