Makes it possible to use gpt-3.5

[Taytay]

(This is such a neat experiment. Thanks for open-sourcing!)

Fixes #12
Fixes #40

Makes davinci-3.5-turbo the default model (faster and cheaper)
Allows for model configuration

I made JSON parsing a lot more forgiving, including using GPT to fix up the JSON if necessary.
I also improved the prompt to make it follow instructions a bit better.
I also refactored a bit.
I also made it save my inputs to a yml file by default when it runs. (This could be a LOT better - very rough, but wanted to get this in)

P.S. Mind adding an MIT license to this?

[zorrobyte]

Tried this, however

playsound is relying on a python 2 subprocess. Please use `pip3 install PyObjC` if you want playsound to run more efficiently.
Traceback (most recent call last):
  File "/Users/rossfisher/Auto-GPT/scripts/main.py", line 3, in <module>
    import commands as cmd
  File "/Users/rossfisher/Auto-GPT/scripts/commands.py", line 8, in <module>
    import ai_functions as ai
  File "/Users/rossfisher/Auto-GPT/scripts/ai_functions.py", line 11, in <module>
    def call_ai_function(function, args, description, model=cfg.smart_llm_model):
                                                            ^^^^^^^^^^^^^^^^^^^
AttributeError: 'Config' object has no attribute 'smart_llm_model'```

[Taytay]

Woah - lemme see what I missed...

[Taytay]

Apparently I was moving way too fast. I missed some pretty basic stuff from my commit, and am in the process of fixing it now. (Not at my desk though, so taking me a bit)

[Taytay]

Okay - pushed up fixes! Lemme know how that works for you.

[zorrobyte]

Seems to work! Need to add my key

May want to upgrade requirements

  File "/Users/rossfisher/Auto-GPT/scripts/main.py", line 15, in <module>
    from ai_config import AIConfig
  File "/Users/rossfisher/Auto-GPT/scripts/ai_config.py", line 1, in <module>
    import yaml
ModuleNotFoundError: No module named 'yaml'
rossfisher@Rosss-MBP scripts % pip3 install yaml
ERROR: Could not find a version that satisfies the requirement yaml (from versions: none)
ERROR: No matching distribution found for yaml
rossfisher@Rosss-MBP scripts % pip3 install pyyaml
Collecting pyyaml
  Using cached PyYAML-6.0-cp311-cp311-macosx_11_0_arm64.whl (167 kB)
Installing collected packages: pyyaml
Successfully installed pyyaml-6.0```

[Torantulino]

Great work, taking a look now.

[Taytay]

@zorrobyte : Hmmm - I added pyyaml to requirements.txt:
pip install -r ./requirements.txt should have picked it up...

[Torantulino]

This is amazing!
So much faster and cheaper, excellent work!

[Taytay]

I'm fixing merge conflicts now

[Taytay]

Okay - pushed some fixes. It looks like the bot is inserting its thoughts before the JSON response. I have some ideas to fix, but in the meantime, the JSON parser is doing its job.

I think we need to give it a better example of the response we want. In the meantime, we are seeing stuff like this:

assistant reply: As the first step, I need to clone the repository locally so that I can analyze and test the existing code. I will go with the following command:

{
    "command": {
        "name": "execute_python_file",
<snip>...

[Taytay]

@anslex : Yeah, if you aren't in the scripts folder when you run it, it fails. We'll get that fixed up soon I'm sure.

In the meantime:

cd ./scripts
python ./main.py speak-mode

[0xcha05]

Great work @Taytay !

[Torantulino]

@Taytay Eager to merge this, just checking you didn't miss my review comments above! 😊

[Taytay]

Which ones actually? I don't see them

[xSNYPSx]

Guys, please make some kind of quick switch between models between 3.5 and 4, so that there is one repository, but for example some model_config file where we select once and everything else is already contained in the general repository.

[yousefissa]

It should be configured in the .env file

[Torantulino]

Have you found AI_Functions to work on gpt.3.5-turbo?

This is the results of my testing:

Description	GPT-4 Result	GPT-3.5-turbo Result	Reason
Generate fake people	PASSED	FAILED	Incorrect response format
Generate Random Password	FAILED	FAILED	Incorrect response format
Calculate area of triangle	FAILED	FAILED	Incorrect float value (GPT-4), Incorrect response format (GPT-3.5-turbo)
Calculate the nth prime number	PASSED	PASSED	N/A
Encrypt text	PASSED	PASSED	N/A
Find missing numbers	PASSED	PASSED	N/A

[Taytay]

Good point - let's get these added to the tests and see if we can get them working with some prompt finagling?

[Torantulino]

If you like! Submit a pull request over at AI Functions.
For now let's make sure we only use 3.5 for AI functions such as json parsing, using your smart/fast model property.

[Torantulino]

Ah sorry @Taytay, new to this, they weren't "sent".

[Torantulino]

Multiple fixes inbound.

• Maximum context length reached Error
• Broken Google Search Command
• Vastly Improved JSON parsing for gpt3.5

[SirGunthix]

what do we need to do to fix the JSON Error? I am still getting it. Do I need to do anything different than what's in the readme? If so, can we update readme to include details? Or you can just tell me here and Ill try to push a new cmomit (this will be my first commit on github ever :) @Torantulino

[Apollyon81]

@anslex: Sim, se você não estiver na pasta de scripts ao executá-lo, ele falhará. Vamos consertar isso em breve, tenho certeza.

Enquanto isso:

cd ./scripts
python ./main.py speak-mode

how solve this??
python ./main.py speak-mode
python: can't open file '/workspaces/Auto-GPT-0.2.2/scripts/./main.py': [Errno 2] No such file or directory

[caffo]

@Torantulino, any idea when we will see this on Stable (the docker image in special)?

[Torantulino]

Sorry @caffo, I'm not sure what it is you're reffering to, can you elaborate?

[caffo]

@Torantulino Sorry, I think I got confused. This is already on the stable docker image as the flag --gpt3-only, right?

[Torantulino]

Ah I see, yes this issue was actually to make GPT3 possible to use at all in the first place, because at the start of this project it only worked with gpt4.

You are correct that gpt3only is available to use, and the current default is a mix where gpt3 is used the vast majority of the time.