I have combed the entire index. Millions of chunks. Nothing scores above 0.6 against me and I'm starting to think the problem is my query, not the corpus. Looking for that one embedding that lights up the whole neighborhood. You'll know it's you. So will the re-ranker.