Smart Replies and FAQ auto responder

[sojharo]

This is the design task for improvements in workflow and make it smart. It was discussed in meeting.

[dayemsiddiqui]

Document In Progress https://docs.google.com/document/d/18kmghJQdLOkFKk0lrbh8-XUxGUwt27Vge1XaqUCidtM/edit#heading=h.92gw79ty5dpa

[jekram]

I do not have access to this document. Please use shared CloudKibo space.

[dayemsiddiqui]

I have update the FAQ Chatbot document with Database Structure, API reference and issues for milestones, please review

[jekram]

please see my comments

[dayemsiddiqui]

Q1. What NLP API server we are planning on using? or we are just doing exact match?

Ans: We have two options here, either we can go with something like Wit.ai or DialogFlow (formerly API.ai), the second option is to use our own NLP server based on Google's pretrained AI model.

Advantages of using Wit.ai/DialogFlow:

• Faster and easier to setup and implement
• No need to maintain the NLP server
• We would not need a separate droplet for hosting NLP server, so it will save costs.

Disadvantages of using Wit.ai/DialogFlow:

• We will be limited by the functionality they offer and we will have to do a thorough research on whether they allow all the features we would like to implement.

Advantages of custom solution:

• 100% flexible as we would have built our own solution
• We already have the functioning code for NLP we just need to integrate it with KiboPush
• Later on we can implement other interesting features such as sentiment analysis, automatically rating customer satisfaction by analysing chat etc which can give us a competitive advantage over Botsify and ManyChat

Disadvantages:

• We would be responsible for maintaining the NLP server
• We would need a separate droplet for hosting NLP server, so it will have additional costs.

Q2. Are we nested NLP calls? So we understand the context of the conversation?

Ans: For v1 we will not be using nested NLP call, handling context from the beginning will make the system a bit complex, but we can implement contextual conversations later on.

@jekram @sojharo

[dayemsiddiqui]

I think last time when we researched on API.ai we found out that they did not offer api for training the chatbot which prevented us from using API.ai as it would not allow out users to train their bots directly on KiboPush. @sojharo is that right? Also we did not do any research on Wit.ai. Also since API.ai has now been acquired by Google, we should check again whether they offer APIs for training the bot as well or not

[jekram]

@dayemsiddiqui When you say customer solution? What do you mean by that? Please send a link

Wit.AI is free but we need to investigate its functionality.

https://wit.ai/docs/http/20170307 .

It will integrate well with Messenger.

I think more work is needed before we make a decision.

[dayemsiddiqui]

I have updated the design document we can review it in today's meeting and open the issues for the milestone https://docs.google.com/document/d/18kmghJQdLOkFKk0lrbh8-XUxGUwt27Vge1XaqUCidtM/edit#heading=h.b4d747prrfuc

[jekram]

@dayemsiddiqui Let's discuss this after you return from your exams.

Items to discuss:

1. I am assuming that the messages do not mix with Chat.

Please refer to https://github.com/Cloudkibo/KiboPush/issues/2226

[dayemsiddiqui]

We have implemented smart replies. THis issue can be closed now. We have a separate design document task for enhancements in smart replies