When ‘Beta’ is just too ‘beta’
Why we made the change from Wit.ai to Api.ai
For those of you that have not been following my story with this exploration into the world of restaurant discovery and chatbots… learn about the beginning here.
Before beginning this article, I want to start with a quick disclaimer:
Any opinion I express here is purely related to the needs of our product, @where. That does not mean that either of these products is good or bad.
In my last article we were working with Wit.ai as our main Natural-Language Processor (NLP) for building out our chatbot. Here, I’m going to tell you why we switched to Api.ai by the end of the week.
Things were looking pretty good. Wit.ai got @where talking. And Nick was getting it hooked into Facebook Messenger.
As a designer, I appreciated that the experience of using Wit as an interface is incredibly intuitive. The use of Stories in Wit is a concept which was easy for me to wrap my mind around and work with from the beginning.
(To the contrary — trying to understand the nuanced differences between Entities, Intents, etc in Api.ai was quite the challenge from the start.)
But, as we continued to work with Wit, we ran into a few complications. The major ones were these:
- Manual integration with platforms like FB, Slack, etc.
- Struggle creating a form-based chatbot
Manual Integration
When working with Wit, you have to integrate with platforms — like Facebook, Slack, etc., by hand. Aka, look through the documentation and figure out how to Frankenstein things together. For a non-programmer, that’s a lot to wrap your head around in one sitting. Simply put, it made shit challenging for Nick and I.
But, with Api.ai…there are other options.
In fact, they provide quick integrations for many of your favorite platforms — from Facebook Messenger to Slack or even Cortana.
This made a huge difference in the speed for getting your Natural Language Processor (NLP) up and running with Facebook Messenger. For us, it definitely made Api.ai fairly attractive.
But, it wasn’t just about integration…
Form-Based Chatbots
@where is all about using conversation to obtain information, and send back restaurant suggestions. Just because I can ask “Where can I find good desserts in New York?” doesn’t mean I always will. What if I just randomly blurt out: “I’m craving cake.” What then? How does @where know where I am? Short of tracking down my GPS location, it just has to ask —
And then we have the data we need. Our NLP can identify cake as a cuisine or food type and translate Bushwick into a geo-location. We then pass those variables into our Google Places Search and voila — restaurant suggestions abound.
But here’s the problem. Either we aren’t using it right, or it’s just too beta — because Wit was having trouble figuring out how to properly identify when to prompt the user for more information.
After training it with 20+ stories, it just wasn’t happening.
But we tested Api.ai and by the end of the first 2 test runs — it was working like a charm. Maybe no “Where you at, homie?”…but definitely a solid “Where are you located?”
And that. Is why we switched.
I’ll keep you updated on how it goes from here on out. But at this moment, for our purposes — I’d throw my hat in with Api.ai any day.
