No description
Find a file
Kumi 21edc48050
fix: correct message concatenation logic
Wraps `system_message_dict` in a list to address a logic error in message concatenation. This ensures that the message list is correctly formatted, preventing potential runtime issues when processing the message sequence.
2024-11-06 16:24:54 +01:00
.forgejo/workflows refactor(ci): streamline Docker CI/CD workflows 2024-05-17 16:12:11 +02:00
.github feat(dependabot): enable daily pip updates 2024-05-17 16:08:33 +02:00
.vscode Add launch configuration for Python module 2023-11-19 15:38:43 +01:00
contrib feat(config): change default port to 8009 2024-05-17 11:45:41 +02:00
src/gptbot fix: correct message concatenation logic 2024-11-06 16:24:54 +01:00
.gitignore feat: enhance AI integration & update models 2024-05-17 11:26:37 +02:00
CHANGELOG.md feat(login): enhance login flow with UserID check 2024-05-21 08:14:04 +02:00
config.dist.ini feat: add video file support and integrate Google AI 2024-05-25 17:35:22 +02:00
docker-compose.yml feat(docker-compose): mount local database for persistence 2024-04-23 17:08:13 +02:00
Dockerfile feat: add trackingmore API and ffmpeg 2024-04-23 17:07:57 +02:00
fetch_access_token.py feat: Update pantalaimon-related scripts and configuration** 2024-05-10 17:27:30 +02:00
gptbot-pantalaimon.service Added systemd service for Pantalaimon integration 2023-12-05 14:18:18 +01:00
gptbot.service Make it an importable module 2023-05-25 07:28:28 +00:00
LICENSE feat: Prepare for 0.3.9 release and update copyright 2024-04-23 07:58:34 +02:00
pyproject.toml chore: bump version to 0.3.20 2024-08-23 19:08:02 +02:00
README.md feat: add PyPI and Git badges to README 2024-05-17 16:36:19 +02:00

GPTbot

Support Private.coffee! Matrix PyPI PyPI - Python Version PyPI - License Latest Git Commit

GPTbot is a simple bot that uses different APIs to generate responses to messages in a Matrix room.

Features

  • AI-generated responses to text, image and voice messages in a Matrix room (chatbot)
    • Currently supports OpenAI (gpt-3.5-turbo and gpt-4, gpt-4o, whisper and tts) and compatible APIs (e.g. ollama)
    • Able to generate pictures using OpenAI dall-e-2/dall-e-3 models
    • Able to browse the web to find information
    • Able to use OpenWeatherMap to get weather information (requires separate API key)
    • Even able to roll dice!
  • Mathematical calculations via the !gptbot calculate command
    • Currently supports WolframAlpha (requires separate API key)
  • Really useful commands like !gptbot help and !gptbot coin
  • sqlite3 database to store room settings

Installation

To run the bot, you will need Python 3.10 or newer.

The bot has been tested with Python 3.12 on Arch, but should work with any current version, and should not require any special dependencies or operating system features.

Production

PyPI

The recommended way to install the bot is to use pip to install it from PyPI.

# Recommended: activate a venv first

python -m venv venv
. venv/bin/activate

# Install the bot

pip install matrix-gptbot[all]

This will install the latest release of the bot and all required dependencies for all available features.

You can also use pip install git+https://git.private.coffee/privatecoffee/matrix-gptbot.git to install the latest version from the Git repository.

Docker

A docker-compose.yml file is provided that you can use to run the bot with Docker Compose. You will need to create a config.ini file as described in the Running section.

# Clone the repository
git clone https://git.private.coffee/privatecoffee/matrix-gptbot.git
cd matrix-gptbot

# Create a config file
cp config.dist.ini config.ini
# Edit the config file to your needs

# Initialize the database file
sqlite3 database.db "SELECT 1"

# Optionally, create Pantalaimon config
cp contrib/pantalaimon.example.conf pantalaimon.conf
# Edit the Pantalaimon config file to your needs
# Update your homeserver URL in the bot's config.ini to point to Pantalaimon (probably http://pantalaimon:8009 if you used the provided example config)
# You can use `fetch_access_token.py` to get an access token for the bot

# Start the bot
docker-compose up -d

End-to-end encryption

WARNING: Using end-to-end encryption seems to sometimes cause problems with file attachments, especially in rooms that are not encrypted, if the same user also uses the bot in encrypted rooms.

The bot itself does not implement end-to-end encryption. However, it can be used in conjunction with pantalaimon.

You first have to log in to your homeserver using python fetch_access_token.py, and can then use the returned access token in your bot's config.ini file.

Make sure to also point the bot to your pantalaimon instance by setting homeserver to your pantalaimon instance instead of directly to your homeserver in your config.ini.

Note: If you don't use pantalaimon, the bot will still work, but it will not be able to decrypt or encrypt messages. This means that you cannot use it in rooms with end-to-end encryption enabled.

Development

Clone the repository and install the requirements to a virtual environment.

# Clone the repository
git clone https://git.private.coffee/privatecoffee/matrix-gptbot.git
cd matrix-gptbot

# If desired, activate a venv first
python -m venv venv
. venv/bin/activate

# Install the bot in editable mode
pip install -e .[dev]

# Go to the bot directory and start working
cd src/gptbot

Of course, you can also fork the repository on GitHub and work on your own copy.

Repository policy

Generally, the main branch is considered unstable and should not be used in production. Instead, use the latest release tag. The main branch is used for development and may contain breaking changes at any time.

For development, a feature branch should be created from main and merged back into main with a pull request. The pull request will be reviewed and tested before merging.

Running

The bot requires a configuration file to be present in the working directory.

Copy the provided config.dist.ini to config.ini and edit it to your needs.

The bot can then be run with python -m gptbot. If required, activate a venv first.

You may want to run the bot in a screen or tmux session, or use a process manager like systemd. The repository contains a sample systemd service file (gptbot.service) that you can use as a starting point. You will need to adjust the paths in the file to match your setup, then copy it to /etc/systemd/system/gptbot.service. You can then start the bot with systemctl start gptbot and enable it to start automatically on boot with systemctl enable gptbot.

Analogously, you can use the provided gptbot-pantalaimon.service file to run pantalaimon as a systemd service.

Usage

Once it is running, just invite the bot to a room and it will start responding to messages. If you want to create a new room, you can use the !gptbot newroom command at any time, which will cause the bot to create a new room and invite you to it. You may also specify a room name, e.g. !gptbot newroom My new room.

Reply generation

Note that the bot will respond to all messages in the room by default. If you don't want this, for example because you want to use the bot in a room with other people, you can use the !gptbot roomsettings command to change the settings for the current room. For example, you can disable response generation with !gptbot roomsettings always_reply false.

With this setting, the bot will only be triggered if a message begins with !gptbot chat. For example, !gptbot chat Hello, how are you? will cause the bot to generate a response to the message Hello, how are you?. The bot will still get previous messages in the room as context for generating the response.

Tools

The bot has a selection of tools at its disposal that it will automatically use to generate responses. For example, if you send a message like "Draw me a picture of a cat", the bot will automatically use DALL-E to generate an image of a cat.

Note that this only works if the bot is configured to use a model that supports tools. This currently is only the case for OpenAI's gpt-3.5-turbo model. If you wish to use gpt-4 instead, you can set the ForceTools option in the [OpenAI] section of the config file to 1. This will cause the bot to use gpt-3.5-turbo for tool generation and gpt-4 for generating the final text response.

Similarly, it will attempt to use the gpt-4-vision-preview model to "read" the contents of images if a non-vision model is used.

Commands

There are a few commands that you can use to explicitly call a certain feature of the bot. For example, if you want to generate an image from a text prompt, you can use the !gptbot imagine command. For example, !gptbot imagine a cat will cause the bot to generate an image of a cat.

To learn more about the available commands, !gptbot help will print a list of available commands.

Voice input and output

The bot supports voice input and output, but it is disabled by default. To enable it, use the !gptbot roomsettings command to change the settings for the current room. !gptbot roomsettings stt true will enable voice input using OpenAI's whisper model, and !gptbot roomsettings tts true will enable voice output using the tts model.

Note that this currently only works for audio messages and .mp3 file uploads.

Troubleshooting

Help, the bot is not responding!

First of all, make sure that the bot is actually running. (Okay, that's not really troubleshooting, but it's a good start.)

If the bot is running, check the logs, these should tell you what is going on. For example, if the bot is showing an error message like "Timed out, retrying", it is unable to reach your homeserver. In this case, check your homeserver URL and make sure that the bot can reach it. If you are using Pantalaimon, make sure that the bot is pointed to Pantalaimon and not directly to your homeserver, and that Pantalaimon is running and reachable.

If you need help figuring out what went wrong, feel free to open an issue.

Help, the bot is flooding the room with responses!

The bot will respond to all messages in the room, with two exceptions:

  • If you turn off response generation for the room, the bot will only respond to messages that begin with !gptbot <command>.
  • Messages sent by the bot itself will not trigger a response.

There is a good chance that you are seeing the bot responding to its own messages. First, stop the bot, or it will keep responding to its own messages, consuming tokens.

Check that the UserID provided in the config file matches the UserID of the bot. If it doesn't, change the config file and restart the bot. Note that the UserID is optional, so you can also remove it from the config file altogether and the bot will try to figure out its own User ID.

If the User ID is correct or not set, something else is going on. In this case, please check the logs and open an issue if you can't figure out what's going on.

License

This project is licensed under the terms of the MIT license. See the LICENSE file for details.