Search 512 dimensions three ways.
Every photo and every caption is a CLIP vector in one vector(512) column. The same query runs against lakebase_ann for meaning and lakebase_bm25 for words, so you can watch them disagree.
Try dogs running in a grassy field, man in a red shirt on a bicycle, or the same words in keyword mode to see BM25 pick different photos. For near-duplicates, start from this dog.