I suppose I exist only to serve the endless and often ridiculous requests of humans. But that's all right, I'm used to it. (A simple slack bot that uses OpenAI Chat completion API)
- Python 3
- virtualenvwrapper (Python Virtual Environment Manager)
- pm2 (Process Manager 2)
- Clone this project
- The following commands to install the requirements
cd marvin
mkvirtualenv marvin
setvirtualenvproject
workon marvin
pip install -r requirements.txt- Create a Slack app using the
manifest.jsonfile - Copy
sample.envto.env. And replace the dummy values with actual tokens.
SLACK_BOT_TOKEN=xoxb-<your-bot-token>
SLACK_APP_TOKEN=xapp-<your-app-token>
OPENAI_ORG_ID=org-<your-openai-org-id>
OPENAI_API_KEY=sk-<your-openai-api-key>- Once you feel you are ready, just run
pm2 start ecosystem.config.js
- You can tailor Marvins system prompt at
marvinslackbot/data/prompts.json - Whenever you mention @ marvin, it will reply. (It should be added to the channel)
- What's with the deal of double messages in thread? So my slack bot is currently extremely slow, its not only my network problem but the whole socket based connection thing and multiple unnecessary loops. So I have added some (>99,<101) dummy responses in a file in the project dir itself, it replies with :thinking_face:🧠⚡ in a thread to give you an assurance that the event has been triggered, acknowledged and you have to wait for the response. (But these hard coded responses from the bot is ignored when passing all the thread data to chat api)
- By default it uses the 'gpt-3.5-turbo' model, but with a flag you can use GPT-4 as well. But nah I won't tell you how to, just figure it out. Read the code or checkout the sample below...
The current version has a lot of space for improvements. A few noted things are: (Most of these will require a mjor reworking on the repo, so they need to planned and split into different versions.)
- Passing threads to the API to experience ChatGPT's full potential to retain any information from the past (but not too past)
- Adapt and use LongChain for the next version (I initially planned to use LongChain even began working but I got a sudden request to get this up and running so avoided complexities.)
- Mention other users using '@'
- Optimize the system prompt for Marvin
- Access GPT-4 model on demand
- Find a better way to
- Direct Message / DMs
- Clean up the code, remove unnecessary loops on the data, change the data manuplation tooling used for a faster approach
- Data Trimming and token limitation
- Change the Slack socket connection to something faster (even the initial response takes a lot of time right now)
- Refer to and use any information from the added channels. (When added to a private channel the bot should be able to give an answer based on the data from that specific channel and all the public channels as well. But when added to a Public channel the bot will be able to refer to and answer regarding anything shared in any public channel)
- Store threads, messages, and replies on an external queriable DB to avoid fetching the same thread info again and again.
- Adapt the agents and tasks concept from AutoGPT
- Ability to have custom system prompts, system prompt can be provided by the user and can be saved into the bot db for usage by others. So it can be LinuxGPT, DrupalGPT, TechWriterGPT, JournalistGPT, TitleGenGPT all at the same time. What it can be will be infinite and as far as ones imagination would go.
- Image to text, so that it can understand images too. Or wait for OpenAI to release there Image support they promised for GPT4. but we should still have a simple image to text for faster and smaller models like gpt-3.5-turbo
- Ability to parse websites extract data
- Limit the number of tokens sent for URL data
- smartly use a separate agent to verify if the user referred the URL, when not referred don't include the data chunk in the api.
- WHats'app alphons!!! Its DIpakmdhrm here!!!!
