• Introduction
• Demo
• Installation
  • Backend
  • Frontend
  • Configuration
• 30-Second Startup
  • Start the backend server
  • Start the front-end server
• Resources

Table of contents generated with markdown-toc

End-to-end framework to build automatic agents (chatbots) for task-oriented dialogs

• 💬 Use chat interfaces to collect dialog data and interact with agents
• 🤖 Build/deploy automated chatbot agents with your own ML/rule-based models
• 📈 Evaluate chatbots using 1) agent-action annotations 2) user-satisfaction scores

Quickly deploy for any domain by leveraging domain-agnostic modules

• Data collection interface self-annotates dialogs for training/evaluation data
• Train a chatbot without ML knowledge using generic ML models, or write a custom model
• Evaluate any chatbot's accuracy as a black-box, or inspect its internal actions as a white-box

Chat interfaces are deployed with a simple web architecture more details in a Google drive. The backend is Python/Flask/sockets and the frontend is Preact/Redux.

Below is a short snippet of a user interacting with a chatbot that helps them specify a destination for a taxi ride.

Live Chatbot: http://35.212.174.248:9999/0/user. Chat directly with the chatbot here.

• Tested on python 3.6.3, requires python 3.6+ for the newer versions of the transformers module.

1. python3 -m venv env
2. source env/bin/activate

3. pip install -r requirements.txt
4. This installs server and API dependencies. Some agents require additional installation. See their folders for details or install requirements-full.txt.

• Requires Node.js. Tested with Node 12.10

Install the required dependencies.

1. cd gui/frontend
2. npm install

Backend configuration is all done through command line arguments to main.py. Run python main.py --help for a list of options. If you are running a domain that requires API keys, you need to create the gui/backend/keys/api_keys.json file and set them there. Here's an example of api_keys.json:

{
  "darksky": "my_api_key1",
  "google_maps": "my_api_key2",
  "google_speech": "my_api_key_file.json",
  "wit_date": "my_api_key3"
}

Frontend configuration is done using dotenv, stored in the auto-generated configuration file gui/frontend/.env. There is an example at gui/frontend/.env.example You can modify the parameters in this configuration file after it is generated by the front-end start-up command. For domains that require frontend API keys (e.g. destination), you should add the keys there.

Let's start a chat interface for the simple compare_numbers domain that determines whether one number is greater than another.

BACKEND_PORT=8081
FRONTEND_PORT=8082
LOG_DIR=logs/tutorial

cd gui/backend
python main.py --port=$BACKEND_PORT --domain=compare_numbers --num_rooms=1 --log_dir=$LOG_DIR

This starts a backend for human-agents that make api calls and reply to user utterances via the agent UI.

To see what it's like to interact with a chatbot-agent, add --agent_class_name=agents.compare_numbers_agent.CompareNumbersAgent to the command above.

cd gui/frontend
./run.sh $BACKEND_PORT $FRONTEND_PORT

The front-end command will output URLs to access the chat interface, e.g. http://0.0.0.0:8082.

• Chat Interface
• Datasets
• Pre-Built and Custom Domains
• Testing
• Implement and Evaluate a Custom Agent