Thanks! Vector databases store the semantic vector representation of each record and compare it to the query for retrieval, which would give results close to the meaning of the text, but not necessary the text surface. A lexical search, i.e. BM25 and levenshtein distance, seems to work better as translation examples in this case