for a FAQ Bot you can start with a bag of words instead of word2vec. word2vec is better to find similarities or to do some aritmethic like queen - man = king you can use csv,tsv,json it´s up to you how you read it. use the answers like classes or labels and use the words in the questions as a bag of words. for every question build a one hot encoded vector and train it in a simple dnn. that way will work. for better results prepare your input with stemming and lemmatization. btw: word2vec uses bag of words and skip gram so if you really want to use word2vec you can use the gensim library but for faq i would go the dnn bow way.
... View more