I’ve also noticed how powerful actual keywords are when trying to retrieve data from a sea of candidates that have high cosine similarity to the query. My quick and dirty ranking mechanism is simply keyword match first, and only then similarity score.