GRACE: GPT Reprogrammable Assistant with Code Execution

Important: The war in Ukraine is still ongoing. Every day, Russian soldiers rape, murder, torture and deport Ukrainian civilians. Visit this page to see how you can support Ukraine today.

GRACE leverages the OpenAI API to implement a human-like chatbot that's capable of retrieving information from a knowledge base and processing the customers' requests via an API backend. It showcases how large language models can be made to interact with external systems for knowledge retrieval and performing actions in the real world. Learn more in the announcement blog post.

Update April 5, 2023: Horace, my new project, builds upon the ideas in GRACE to offer an LLM chatbot server with support for ChatGPT plugins. Check it out!

The default configuration implements an AI assistant for Death Star, an imaginary Star Wars-themed restaurant in Cupertino, CA. The assistant is instructed to impersonate Jedi Master Yoda.

Example interaction:

AI: Greetings, young Padawan! What can I do for you?
Customer: Hi Yoda, may the Force be with you! Is Darth Vader around?
AI: Not today, youngling. Would you like me to help you with something else?
Customer: Yes, what can you do?
AI: Of course, what is it that you seek? I can help you with making a reservation, changing an existing one, or canceling it. What can I do for you?
Customer: Do you have a table for 3 people tomorrow?
AI: Absolutely! What time would you like for the reservation?
Customer: 4:15 pm
AI: Good news, youngling - the table is available. Would you like to book it?
Customer: Yes sir!
AI: Very well, what is your full name?
Customer: Why do you need the name?
AI: I need your full name for the booking, to make sure that you are the rightful owner.
Customer: Ah, OK. It's Jeremy Biggs.
AI: The booking is confirmed and your reference is 7APZN6. Is there anything else I can do for you?
Customer: Just had a question, what's the parking situation at the restaurant?
AI: Excellent! Free parking is available on site and you can also park on the street (charges may apply). Is there anything else I can do for you?
Customer: Nope, that's it, thanks.
AI: It's been an honor serving you. May the Force be with you!

Unlike the GPT-3 and ChatGPT demos you might have seen so far, this is not just fantasy chitchat on the part of the bot. The bot actually executes backend commands to check table availability and make the booking, as well as to answer the customer's question about the parking. It also interprets the execution results and retranslates them to the customer.

Installation

Clone this repository and cd to it:

cd grace

Create a Python virtual environment:

python3 -m venv ./venv

Activate the environment:

source ./venv/bin/activate

Install the required packages:

pip3 install -r requirements.txt

Copy .env.example to .env:

cp .env.example .env

Edit .env to set your OpenAI API key. You can obtain the key by signing up for OpenAI API here.

OPENAI_API_KEY=sk-...

Running the Chatbot

Run the command below to start your interactive chat session. On the first run, the chatbot will download a semantic search model from Hugging Face Hub. This may take some time.

python3 main.py

The default assistant, impersonating Yoda, can help you manage a table reservation as well as answer questions about the imaginary restaurant it works for (try asking it about gluten-free menu options, for example). Feel free to throw curveballs at it!

Running Tests

A couple of tests are implemented for illustration purposes. The tests employ another chatbot instance to play the role of the human counterpart to GRACE. Use this command to run them:

pytest -s

Model Selection

In addition to the default text-davinci-003 model, GRACE can also work with ChatGPT and GPT-4. See config.yaml for the details and example settings.

GPT-4 is the best, but also the most expensive. The default text-davinci-003 is cheaper but still produces excellent results. ChatGPT is the cheapest but may suffer from a number of issues (hallucinations, failure to execute commands etc.)

Chatbot Customization

To adapt the bot for your use case, take a look at backend.py and domain.yaml.

backend.py implements the Python API that defines the types of customer requests that the bot can process, such as booking a table.
domain.yaml provides basic business information, adds any extra instructions for the chatbot, as well as includes the FAQ/knowledge base with answers to questions about your business.