The underlying linguistics of a "phrase" here are sort of narrow. What we did is retokenize the text so that entities and basic noun phrases are merged into a single token. "Netflix and chill" is analysed as multiple tokens, so it doesn't come up as a query result.