[Feature Request] Function calling

[gilcu3]

Do you plan supporting the function calling feature? Would you be open to accept PRs on that regard?

Currently, some functions such as DuckDuckGo search or Wolfram Alpha could greatly extend the model's capabilities.

[sigoden]

function call is for developers, not end users.

Currently the only promising form of extending model functionality is gpts. Technically, it is possible to implement a command line version of gpt through function call, but at this stage, there are neither users nor suppliers, so there is no chance.

[gilcu3]

I am not you what do you mean by only for developers. Certainly I can use both I mentioned as a user, because the model directly decides which function to call during a standard conversation.

[cboecking]

It seems to me that the best use of function calling is to make the call via cli and pipe in the results to aichat. One use case that I hope evolves is the ability to use aichat as a remote control over an ollama-type system with enhanced tools capabilities.

Is this in line with your comment @gilcu3 ?

[gilcu3]

@cboecking That is one possibility, although the usage I was suggesting is much simpler. When interacting with the bot it might realize it needs to search something online, or solve some math formula, and in such cases it automatically uses function calling to reply to the user. In the background aichat would call (via cli or using a library) the corresponding functions specified by the AI.

[tkanarsky]

tools are something you pass to the base model in a structured form, e.g. for OpenAI https://platform.openai.com/docs/api-reference/chat/create#chat-create-tools, and the API handles whether to return a chat message or request a tool invocation. it would make the most sense to allow the user to define tools and their interfaces in config.yaml, specify if the current model supports tool usage (which most OpenAI compatible API wrappers support, e.g. Huggingface TGI: https://huggingface.co/docs/text-generation-inference/en/basic_tutorials/using_guidance#chat-completion-with-tools), and if the completion requests a tool call then invoke the tool after prompting the user.

example config.yaml, in pseudocode:

- tools:
  list_directory: 
    shell-command: `ls %input`
    interface: {"input": "Path of directory to list"}
  convert:
    shell-command: `convert %source %destination`
    interface: {"source": "Source image name", "destination": "Target image name"}
- models:
  openai:gpt-4o:
    features:vision,tool-usage

in a chat:

temp) Convert image in current dir to jpg

*Run tool `list_directory .`?* (y/N): y
Output of `list_directory .`: foo.tiff doc.txt

Thanks for providing the file names. foo.tiff appears to be the image you're interested in.

*Run tool `convert foo.tiff foo.jpg`?* (y/N): y
Output of `convert foo.tiff foo.jpg`: 

I have obtained the filename of the image in the current directory and converted it to jpg as per your request.

temp)

and re: users vs developers, in many senses users of aichat are sort of developers too!

[sigoden]

#514 has implemented this function, you can try it. Welcome feedback and suggestions.

[gilcu3]

#514 has implemented this function, you can try it. Welcome feedback and suggestions.

Feature working perfectly, thank you! Will you accept PRs to add more example tools to llm-functions? I tested specifically with Wolfram Alpha factoring a big number...

[sigoden]

I welcome you to submit a PR @gilcu3

[sigoden]

There are two function types:

1. Dispatch Functions: Execute a script.
2. Retrieve Functions: Generate JSON data for further LLM processing.

Now the question is how to distinguish them?

• Function Name Conventions: Functions starting with retrieve_ or ending with _data are retrieve functions. All others are dispatch functions.
• Custom Fields: Use # @meta retrieve to mark a retrieve function, adding an is_retrieve: true field to the generated JSON data.
• Autodetect: The AIChat captures the script's output. If it's in JSON format, it acts as a retrieve function; otherwise, it's a dispatch function. However, this approach limits the capabilities of the script. For example, it can't handle progress bars or TUIs. Moreover, output is only available after the script has finished executing.

What are your suggestions?

[gilcu3]

I would go for either of the first 2 options, leaning towards the second. For the moment I cannot imagine any other type of function. For retrieve functions then it would make sense to keep their inputs and outputs as part of the session conversation.

[sigoden]

AIChat now fully supports the function calling feature.

• Supports two function types: retrieve and execute
• Supports multi-calls and multi-steps
• Works within a role/session
• Have function filter
• Executes function calls in parallel