110antisocial011 / autogptcommands

AutoGPT with commands, read/write/del access. Edit .env file for working changes

Auto-GPT: An Autonomous GPT-4 Experiment

---

THIS IS A FORK.

I have updated the project to WORK. This is enough for baseline functionality- but it isn't perfect! The original project is getting bigger every day, and unfortunately, I'm not impressed. I've begun development on a langchain from scratch which is currently private but working much, much faster than this.

When the kinks are ironed out, it'll be here :)

HOW TO USE

IMPORTANT: Please rename your .env.template file to just .env, this will make your environment variables viewable. If it isn't recognized and you are getting missing API key errors, please hardcode your .env filepath to whatever it may be in config.py (line 5)

ALSO IMPORTANT If you don't have a SearX search server, then remove those lines (all 3 - server, user, pass included) from the .env file. You will need some type of search, so please setup a google search API if you want to have search functionality.

If you don't want to use any type of search, please use the --no-search flag.

(so run python main.py --no-search or ./run.sh --no-search from the right directories)

• download dependencies please!
• setup your .env file according to the template with at LEAST your keys, desired LLM + filepathing
• run python main.py from /Auto-GPT/scripts/ or ./run.sh from /Auto-GPT/ Enjoy! If you have issues, I'm hoping it's not too difficult to patch them up, so please let me know if there are. Part of the reason I forked this was to keep a snapshot of it before it got complexified and all messy.

Remember, this script uses SearX. Remove the functionality from commands.py on lines 104/105 if you don't want to use it - but it's free and decent.

It's also worth it to mention that this script requires very, very precise directions. Order matters, and so does what tools you tell it to use. Unless you're using GPT-4, then that model does a bit better with vagueness, but not really.

TODO:

• Fix the horrible, horrible json prompting
  • Possibly remove json altogether. There are better formats - even XML would be less intensive on the LLM than json. We shouldn't be using LLMs illustrious processing power for json formatting!
• Switch to pinecone or ChromaDB over this crappy array-based short-term memory

Changelog:

• Made SearX optional
• Added run.sh and run.bat with flags
• Made search optional altogether using --no-search flag
• Made the .env file parsing a little better (MAKE SURE YOU RENAME .env.template to .env!)

Example Gallery

Click on the images to view them in full size.

This is a crappy example meant to showcase some basic functionality. The original goal prompt was: Name: ResearcherGPT Role: You will delegate by using "start_agent" and tell them to research and return all the information about the "shadow government" key figures. Then, retrieve the information using "memory_list" and summarize it into an article. Goals: ['Delegate to a sub-agent by using "start_agent"', 'Specify to the agent that they must return a summary of key players in the world', 'Make sure to delegate multiple agents', "Then, when they're done, retrieve the information from memory", 'Summarize and write an article']

While my memory.py file's permanent_memory attribute accidentally had something about Elon musk in it - I feel like this is a funny and simple showcase of some of the abilities.

There will always be slight issues sometimes when it comes to parsing Json. This is normal. If it does, please simplify your directions. It can almost always be broken down into subtasks.

Removing JSON feature to come soon - langchain project is still in development. At the moment it's about 3x faster than this project.