Have a look at Speakeasy for a bot that is trained on reddit using deep learning, it produces very plausible responses to queries but has a few weaknesses.
Firstly it's not task directed, so if you ask it about changing the oil in a car it doesn't know any useful facts and it doesn't present them - it just fobs you off.
Secondly, type "you're" into it - it will be very rude to you... because... reddit.
So what's needed is something that can read technical documents and use them at the right time in a conversation. Deepdive (http://deepdive.stanford.edu/) gives you a bit of this capability, but when I used it (version 0.4) I struggled, maybe things have improved! Non the less; I do believe that this is state of the art for machine reading.
IBM did a tonne of work on question answering in Watson but I am now so confused about Watson that I literally don't know what to make of where their work has got to.
Hope that helps, would be interested if there are different opinions out there!
As with the brief flurry of the Semantic Web era, I think the lesson with chat bot VI/AI will again be as much in striking the balance between things like Deep Learning and things like Expert Systems, assuming Deep Learning will never be perfect and Expert Systems will never be easy/cheap (building a good one still requires a domain expert in that field to train it in whatever rules engine you are using to express the domain).
However I think that the cost of knowledge engineering is still massive and the challenge of this kind of project so sharp that it makes me quail to imagine execution! I think three or four years of full on tool development might make this more practical...