After doing `OpenAIAgent.from

zzensin

After doing OpenAIAgent.from_tools how to get a list of tools that have already been added, and add more without redefining the agent entirely?

32 comments

LLogan M

defining the agent is a no-op, so adding more tools that way should be fine?

zzensin

what do you mean? I've got a whole chat history going. It seems terribly inefficient to OpenAIAgent.from_tools constantly in a chat

LLogan M

Chat history is already in memory and can also he passed in? It really is a no-op

I suppose a method could be added for adding additional tools, but it really is not that much of a difference to re-init the agent

This is a pretty common pattern especially for stateless server apis, etc. All the state is in the chat history

zzensin

The reason is I have a tool for creating new tools. I have done this in autogen and coded it myself from nothing, but would like to port it all to llama-index. This is one of the more important tools to me. Here's the autogen version: https://github.com/microsoft/autogen/blob/main/notebook/agentchat_inception_function.ipynb

zzensin

I could get it to work if I could see an example of how to define a llama-index tool without using the function inspection method llama-index is using.

zzensin

Plain Text

_function_config = {
            'type': 'function',
            'function': {
                "name": name,
                "description": description,
                "parameters": {"type": "object", "properties": json_args},
                "required": ["url"],
            }}

llama_tool = FunctionTool.from_defaults(
            lambda **args: self.execute_func(name, packages, code, **args),
            name,
            description,
            _function_config
        )

that didn't work. AttributeError: 'dict' object has no attribute 'schema'

LLogan M

Yea that probably wont work. You could skip the from_defaults and just init the tool
https://github.com/run-llama/llama_index/blob/c01beee1fab7c0de22869ce74f34ebd1f1d54722/llama-index-core/llama_index/core/tools/function_tool.py#L31

Or making something new that extends the base class
https://github.com/run-llama/llama_index/blob/c01beee1fab7c0de22869ce74f34ebd1f1d54722/llama-index-core/llama_index/core/tools/types.py#L97

zzensin

from_defaults rly can't do it? It seemed like it might since it also accepts name, description, fn_schema, tool_metadata

zzensin

FunctionTool looks relatively promising

zzensin

depending on what ToolMetadata wants. checking

LLogan M

Since you are defining a lambda, from_defaults wont work (like you mentioned, because of the function introspection seems to be incompatible with lambdas)

I would just do

Plain Text

llama_tool = FunctionTool.from_defaults(
    lambda **args: self.execute_func(name, packages, code, **args),
    ToolMetadata(description=description, name=name, fn_schema={...}
)

LLogan M

hard part would be manually getting the schema

LLogan M

but its doable, more annoying than hard lol

zzensin

what is the schema supposed to be though? the openai json spec? I'm already generating that

zzensin

it looks like it. The default is like:

Plain Text

parameters = {
    "type": "object",
    "properties": {
        "input": {"title": "input query string", "type": "string"},
    },
    "required": ["input"],
}

though i've never seen "object" before

zzensin

it's always "type": "function"

zzensin

everything else looks the same

zzensin

Plain Text

{
  'type': 'function',
  'function': {
      "name": name,
      "description": description,
      "parameters": {"type": "object", "properties": json_args},
      "required": ["url"],
}

oh it's obviously parameters lol. I've been awake too long

zzensin

no way this is gonna work haha

Attachment