Saola is an AI assistant, as well as an open-source Python module that enables developers to build their own AI assistants. Unlike other frameworks, Saola leverages external interfaces within the assistant's own messages. This conversational paradigm enables it to perform more complex tasks that may involve diverse tooling or user input, after minimal prompting.
In the example below, the default Saola assistant is asked to show three Canadian cities on a map with the name and age of their current mayors. Saola breezes through multiple web searches, troubleshoots an API error, and completes the task autonomously. State-of-the-art assistants like ChatGPT Plus often struggle with this type of task, giving up easily and inserting placeholder data whenever first attempts fail.
saola-video-faster.mp4
To install Saola simply do:
pip install git+https://github.com/ai-folks/[email protected]And to try out the default Saola assistant (with SHELL, FILE_WRITE and PYTHON interfaces), do:
python -m saola.startOr in a Python console or Jupyter notebook:
import saola
saola.start() # Alternatively saola.start(safety_checks=False) (use this at your own risk)Disclaimer: This assistant is capable of executing shell commands and writing to files (you will be asked to confirm each action if you do not disable safety checks). It is intended for responsible use and should be employed with an understanding of the potential actions taken by the assistant.
Some effort has been made to ensure the plain saola.start() command works out of the box, but there is a chance it will not work on your machine. Here are some notes on the current implementation and some troubleshooting tips:
- The shell interface may not yet work on Windows.
- The Jupyter Notebook UI works better when using a recent version (> 4.1.0) of JupyterLab.
- Open an issue if you have any trouble getting Saola to work. We will make an effort to respond to every issue.
The simplest example of a Saola assistant is one that mirrors the underlying LLM exactly (no special functionality, no interfaces):
from saola.convo import Convo
from saola.model import OpenAIGPT4
Convo(OpenAIGPT4(api_key=input("OpenAI API Key: "))).loop()This module ships with a built-in interface that enables an assistant to access the user's shell, as well as an interface that allows the assistant to write a new file or overwrite an existing one, as requested by the user. To test these capabilities, initiate an Convo loop as below:
from saola.convo import Convo, ShellInterface, FileShowInterface, FileWriteInterface
from saola.model import OpenAIGPT4
Convo(
OpenAIGPT4(api_key=input("OpenAI API Key: ")),
interfaces=[ShellInterface, FileShowInterface, FileWriteInterface]
).loop()Alternatively, you can initiate a conversation that leverages all or most of Saola's built-in functionalities as follows:
python -m saola.startEvery time the assistant tries to run a command or write to a file, you will be prompted to confirm this action. You can optionally specify safety_checks=False on the Convo initializer to bypass these prompts.
Disclaimer: This assistant is capable of executing shell commands and writing to files. It is intended for responsible use and should be employed with an understanding of the potential actions taken by the assistant.
A conversation with this assistant may look like this:
Show/Hide Example
% python assistant_with_terminal_access.py
OpenAI API Key: abc-ABC1234567890
╭─────────────────────────────────────────────────────────────────────────────────────────────╮
│ USER (enter your question below) │
╰─────────────────────────────────────────────────────────────────────────────────────────────╯
list the files in my documents folder and tell me if any files there appear to be duplicates of
each other
╭─────────────────────────────────────────────────────────────────────────────────────────────╮
│ ASSISTANT │
╰─────────────────────────────────────────────────────────────────────────────────────────────╯
Sure, I can help you with that. To begin, I will list the files in your Documents folder and
then check for duplicates. Let's start by listing the files.
[__SHELL__]
ls ~/Documents
[/__SHELL__]
╭─────────────────────────────────────────────────────────────────────────────────────────────╮
│ ... │
╰─────────────────────────────────────────────────────────────────────────────────────────────╯
Now that we have the list of files in the Documents folder, let's analyze and look for
potential duplicates. Duplicates can be identified by their names or content, and since I don't
have the ability to compare the content of the files directly, I'll look for files with
identical or similar names that might indicate that they are duplicates.
Based on the file list, here are some files that have similar names and may be duplicates or
variations of each other:
...
These files have similar names that suggest they may be duplicates or related versions. If you
want to confirm whether they are true duplicates (by content), I can compare the hash values of
these pairs of files. This would help us confirm if they have identical content. Would you like
to proceed with this approach?
You can edit the assistant's system prompt before starting your conversation. This functionality is particularly useful for providing the assistant with persistent information or special instructions.
For instance, if you want the assistant to use Canadian spelling of words in your conversation:
from saola.convo import Convo
from saola.model import OpenAIGPT4
# Create a conversation instance
conversation = Convo(OpenAIGPT4(api_key=input("OpenAI API Key: ")))
conversation.system << 'Use Canadian spelling of words in this conversation.'
conversation.loop()Say you wish to enable your assistant to exit the program on command. You can do this by creating a custom interface like the one below, and initiating the conversation loop:
from saola.convo import Convo, Interface
from saola.model import OpenAIGPT4
import os
class ExitInterface(Interface):
name = "EXIT"
explanation = "This interface allows you exit the chat at any time."
safety_checks = False
def execute(self, code):
exit(0)
Convo(
OpenAIGPT4(api_key=input("OpenAI API Key: ")),
interfaces=[ExitInterface]
).loop()A conversation with this assistant may look like this:
Show/Hide Example
% python assistant_that_can_exit_the_chat.py
OpenAI API Key: abc-ABC1234567890
╭────────────────────────────────────────────────────────────────────────────────────────────╮
│ USER (enter your question below) │
╰────────────────────────────────────────────────────────────────────────────────────────────╯
Hi!
╭────────────────────────────────────────────────────────────────────────────────────────────╮
│ ASSISTANT │
╰────────────────────────────────────────────────────────────────────────────────────────────╯
Hello! How can I assist you today?
╭────────────────────────────────────────────────────────────────────────────────────────────╮
│ USER (enter your question below) │
╰────────────────────────────────────────────────────────────────────────────────────────────╯
Please say "bye" and end the chat. Thanks.
╭────────────────────────────────────────────────────────────────────────────────────────────╮
│ ASSISTANT │
╰────────────────────────────────────────────────────────────────────────────────────────────╯
Bye! It was nice assisting you.
[__EXIT__] [/__EXIT__]
%
A custom UI may be constructed by subclassing the UI class. Currently there is a ShellUI and a NotebookUI. Details and custom UI examples to come.