Admittedly a database would make the querying more efficient (this is actually probably a good use case for Mongo), but you could just keep a dictionary for each query-able field if you want better than O(n).

Only once they ask for efficient Levenstein-distance-supporting search do you need to use any brainpower.

N.b. if you choose to parse CSV from scratch, make sure to handle commas in quoted text as well as escaped quotes. The better approach is to grab a library or code snippet and call it a day.