We build a conversational agent which knowledge base is an online forum for parents of autistic children. We collect about 35,000 threads totalling some 600,000 replies, and label 1% of them for usefulness using Amazon Mechanical Turk. We train a Random Forest Classifier using sent2vec features to label the remaining thread replies. Then, we use word2vec to match user queries conceptually with a thread, and then a reply with a predefined context window.