Reading this was very valuable. I really appreciate the Vespa mention and introduction to their Multi-Vector HNSW Indexing - I’ve recently thought a lot about how difficult chunking is and this seems like a promising avenue.