BERNARD uses a few basic concepts that are mostly instinctive but were defined a bit further with the idea of fitting with all the platforms at the same time.
Bots are a way to easily access a specific service. In that sense, they do translate a human interaction into an API call, gets the result and they translates it back into human language. It looks like this:
As you can see, BERNARD is at the center of the information flow, but the buisness-related intelligence is located in external APIs.
This means that although BERNARD is a Python framework, it can integrate with services written in any other language and seamlessly communicate with your existing infrastructure.
NLU stands for "Natural Language Understanding". It's considered to be the single most important feature of bots and "Artificial Intelligence" is a very hot topic at the moment. However, let's dial down a bit:
- "Artificial Intelligence" is very dumb. Very very dumb. Google is the top player in that field so don't expect to perform better than they do. And you can't really have a conversation with a Google Assistant, can you?
- In practice, we observe that about 3% of people use NLU to interact with bots. When there was a 3% market share for IE6, a lot of people already had stopped supporting it.
In short, don't try to pass the Turing test. Instead, create an experience that doesn't depend on understanding what your users say.
However, there is still a need to understand a few things. And also, if you're selling bots, your clients will ask for it and even if they don't they will expect it and if you don't plan ahead this is going to suck up all your energy. Which is not worth it for about 3% of users.
There is a built-in intents detection system based on very
simple techniques. There is also the
Iron Throne project that
proposes a very simple NLU engine which can be quite effective for
some uses. If none of those fit your needs, you can turn to external
services like wit.ai, dialogflow.com or luis.ai. And finally, if
you want something very specific and a lot of control over it, you can
try rasa NLU.
After analyzing all existing messaging platforms, there is 3 specific atoms that stood out and that are the base of BERNARD's API.
(Messaging) Platforms will give you very little information about your users. You will get the following:
- ID
- Name (at least some form of it)
- Language (most of the time)
- Time zone (sometimes)
As you can see, the only thing we can reliably count on is an ID. Please note that in some platforms like Messenger, IDs are bot-scoped (aka the same user will have different IDs in different bots).
Bots are not necessarily in a 1-1 conversation with users. Users might insert a bot within an existing conversation.
The interesting thing is that most platforms don't give you any information about conversations. In the end it boils down to:
- ID
- List of people (rarely)
Messages are the most difficult atom to model. Most of the time, they will include "rich" content like images, videos, carousels of items, action buttons and so on.
They can be broken down to this:
- Sender
- Recipient
- Content
The content is based on layers. By example an image with a caption
would be an Image layer plus a Text layer. A text with quick replies
(which is a Facebook-specific concept) would be a Text plus a
QuickRepliesList layer.
Example:
In this real-world example we can see that a text message is associated to quick replies. Internally, this is represented by BERNARD layers and then automatically translated into Facebook API calls.
WARNING: don't confuse different layers with different messages. If
you want to send two images, don't send a single message with two
Image layers, but rather two messages with a single Image layer in
each.
BERNARD bots are a way to translate human queries into API queries and display back the result. They use 3 atoms which are users, conversations and messages. Finally, messages are built using layers of content, which allow to use a simple API which covers all platforms.
Next step: Finite-state Machines