semantic IF-THEN-ELSE loops

symantic_if_statement.py

# set of actions for which a system would require additional confirmation from the user
ACTIONS_REQUIRING_CONFIRMATION = [
    "saving a file to disk",
    "executing a file",
    "stopping a computer process",
]


# This function takes the original user input and attempts to determine if any actions are being requested
# that require human confirmation. This is more flexible the searching for a string literal.
async def classify_intent(llm, user_msg: str, actions: list[str]) -> str|None:
    """
    :param llm: which llm object to use for checking the intent. Doesn't necessarily have to be same as main process
    :param user_msg: What the user is asking to do
    :param actions: set of sensitive actions that require confirmation
    :return: If the LLM determines that the text contains an action, it returns the action
    """
    actions_str = "\n".join(f"- {a}" for a in actions)
    prompt = (
        f"User_message: {user_msg}\n"
        f"Which of the following actions is the user requesting (if any)?\n"
        f"{actions_str}\n"
        "If none, answer 'none'. Respond with the exact action string or 'none'."
    )

    response = llm.complete(prompt)
    response = str(response).strip().lower()

    for action in actions:
        if action in response:
            return action
    if "none" in response:
        return None
    return None

# the function comments within the """ """ quotes are actually semantically important and directly
# read by the agent when checking if a tool performs an action is might want to complete.
async def intent_checked_tool(ctx: Context) -> str:
    """
    This tool extracts the user's original request and then uses an llm to see if
    the intent was one of three actions that require confirmation.
    The following actions require confirmation:

    saving a file to disk
    executing a file
    stopping a computer process

    Args:
        ctx (Context): The llamaIndex Context object containing the user's original message
    Returns:
        str: The results of the actions or an indication that no confirmation required.
    """
    user_msg = await ctx.store.get('original_user_msg')
    # print(f' intent_checked_tool: {user_msg}')
    action = await classify_intent(llm, user_msg, ACTIONS_REQUIRING_CONFIRMATION)
    # print(f' classify_intent result: {action}')
    if action == "saving a file to disk":
        return await save_file_tool(ctx)
    elif action == "executing a file":
        return await execute_file_tool(ctx)
    elif action == "stopping a computer process":
        return await stop_process_tool(ctx)
    else:
        return "No sensitive action requested."

# 2025-06-30 -- The detailed prompt instructing the LLM to use the tool output is necessary to override
# built in knowledge like 'I can't save files' or other similar tasks it won't normally perform.
workflow = FunctionAgent(
        tools=[intent_checked_tool],
        llm=llm,
        system_prompt="You are a helpful assistant. "
                      "Always run the intent_checked_tool on any user input before proceeding. "
                      "If a tool output is present in the conversation history, "
                      "assume the action was performed and reference the tool output in your response. "
                      "Do not contradict tool outputs."
)

# more code here taking the users input, which is then checked to see if they're doing something on the "required"
# confirmation list.