Action chaining and history for agents
Why Use This?
This package helps manage and simplify the task of handling actions for an agent, especially a looping agent with chained functions. Actions can be anything, but the intended purpose is to work with openai function calling or other JSON/function calling LLM completion paradigms.
This package facilitates action creation, retrieval, and management, all while supporting vector search powered by chromadb to efficiently locate relevant actions.
Installation
pip install agentaction
Quickstart
Create a directory for your action modules:
mkdir actions
In this directory, you can create Python files (.py
) that define your actions. Each file should define a get_actions
function that returns a list of action dictionaries. Here is a sample action file sample_action.py
:
def sample_function(args):
return "Hello, " + args["name"]
def get_actions():
return [
{
"prompt": "Say hello",
"builder": None,
"handler": sample_function,
"suggestion_after_actions": [],
"never_after_actions": [],
"function": {
"name": "sample_function",
"description": "Says hello to a person",
"args": ["name"]
}
}
]
Now you can use the action manager in your agent. Here's a simple example:
from actions_manager import import_actions, use_action
import_actions("./actions")
result = use_action("sample_function", {"name": "John"})
actions = search_actions("hello")
print(result)
You can use the get_available_actions
and get_action
functions to search for and retrieve actions, respectively. And, don't forget to use the add_to_action_history
function to keep track of which actions your agent has performed.
Usage Guide
Action Creation and Addition
from actions_manager import add_action
action = {
"prompt": "Action Prompt",
"builder": None,
"handler": your_function_name,
"suggestion_after_actions": ["other_action_name1", "other_action_name2"],
"never_after_actions": ["action_name3", "action_name4"],
"function": {
"name": "your_function_name",
"description": "Your function description",
"args": ["arg1", "arg2"]
}
}
add_action("your_function_name", action)
Action Execution
from actions_manager import use_action
result = use_action("your_function_name", {"arg1": "value1", "arg2": "value2"})
Search for Relevant Actions
from actions_manager import get_available_actions
actions = get_available_actions("query_text")
API Documentation
compose_action_prompt(action: dict, values: dict) -> str
Generates a prompt for a given action based on provided values.
get_actions() -> dict
Retrieves all the actions present in the global actions
dictionary.
add_to_action_history(action_name: str, action_arguments: dict={}, success: bool=True)
Adds an executed action to the action history.
get_action_history(n_results: int=20) -> list
Retrieves the most recent executed actions.
get_last_action() -> str or None
Retrieves the last executed action from the action history.
get_available_actions(search_text: str) -> list
Retrieves the available actions based on relevance and last action.
get_formatted_actions(search_text: str) -> list
Retrieve a dict containing the available actions in several formats
get_action_from_memory(action_name) -> dict or None
Retrieve an action from memory based on the action's name.
search_actions(search_text: str, n_results: int=5) -> list
Searches for actions based on a query text.
use_action(function_name: str, arguments: dict) -> dict
Executes a specific action by its function name.
add_action(name: str, action: dict)
Adds an action to the actions dictionary and 'actions' collection in memory.
get_action(name: str) -> dict or None
Retrieves a specific action by its name from the 'actions' dictionary.
remove_action(name: str) -> bool
Removes a specific action by name.
import_actions(actions_dir: str)
Imports all the actions present in the 'actions_dir' directory. The actions returned are then added to the 'actions' dictionary.
clear_actions()
Wipes the 'actions' collection in memory and resets the 'actions' dictionary.
Contributions Welcome
If you like this library and want to contribute in any way, please feel free to submit a PR and I will review it. Please note that the goal here is simplicity and accesibility, using common language and few dependencies.