Key collisions in multi-file collections. For formats without explicit IDs I generate keys using hash(filename:row_index). Works great for single-file collections. Immediately broke with multiple files because I was hashing just the basename, not the full path. 2 files named embeddings.npy in different uploads produce identical keys for every row.
The fix is straightforward -- include enough context in the hash input to be unique. But the failure mode was subtle. No error, just silently overwriting vectors from the first file with vectors from the second. Collection looks healthy, result count is right, half your data is just gone.