Should I create a tfidf on a subset of a data set or use the whole corpus?

My goal in this project is to see if businesses on a list are currently customers within my organization. One piece of this involves producing a similarity score using cosine similarity on the names of the businesses

My initial process uses TfidfVectorizer from sklearn.feature_extraction.text to tokenize the names across all of the customers.

Is it incorrect to subset my organization's customers to those that are in the same state as the customers in the new list?

The next question is should I produce a tfidf across all of the customers within the organization, store the data, and build a new one at runtime that incorporates the names from the new list (as described in this answer)? Or should I only consider the customer names in my organization that are in the same state as those on the new list?