agent.py

import os, time

from langchain_core.messages import SystemMessage, HumanMessage, ToolMessage, AIMessage
from langchain_google_genai import ChatGoogleGenerativeAI
from langgraph.graph import StateGraph, MessagesState, START, END
from langgraph.prebuilt import ToolNode, tools_condition

from tools import my_tools

try:
    if os.path.exists("system_prompt.txt"):
        with open("system_prompt.txt", "r") as f:
            system_prompt = f.read()
    else:
        system_prompt = ""
except Exception as e:
    print(f"Error reading system prompt file: {e}")
    system_prompt = ""
sys_msg = SystemMessage(system_prompt)

class MyMessagesState(MessagesState):
    tool_call_cnt: int=0

class MyAgent:
    def __init__(
            self,
            model_name: str="gemini-2.0-flash",
            tools: list | None=None,
            max_tool_call: int | None=5
        ):

        self.graph = self._build_graph(model_name, tools, max_tool_call)

    def _build_graph(
            self,
            model_name: str,
            tools: list | None,
            max_tool_call: int | None
        ):
        self.model_name = model_name
        # Delays to avoid exceeding rate limits
        if "gemini-2.5" in model_name:
            self.delay = 6
        elif "gemini-2.0" in model_name:
            self.delay = 4
        else:
            self.delay = 0
        self.start_time = None

        llm = ChatGoogleGenerativeAI(model=model_name)
        if not tools:
            tools = my_tools
        llm_with_tools = llm.bind_tools(tools)

        def _delay_wait():
            if self.start_time is not None:
                elapsed = time.time() - self.start_time
                if elapsed < self.delay:
                    time.sleep(self.delay - elapsed)
            self.start_time = time.time()

        def assistant_with_tools(state: MyMessagesState) -> MyMessagesState:
            print("assistant_with_tools")
            _delay_wait()
            return {
                "messages": [llm_with_tools.invoke(state["messages"])],
                "tool_call_cnt": state["tool_call_cnt"]
            }

        def assistant_without_tools(state: MyMessagesState) -> MyMessagesState:
            print("assistant_without_tools")
            _delay_wait()
            return {
                "messages": [llm.invoke(state["messages"][:2])],
                "tool_call_cnt": state["tool_call_cnt"]
            }

        def increment_tool_call_cnt(state: MyMessagesState) -> MyMessagesState:
            print("increment_tool_call_cnt")
            return {
                "messages": state["messages"],
                "tool_call_cnt": state["tool_call_cnt"] + 1
            }

        def check_tool_call_status(state: MyMessagesState):
            # Initialize tool calling counter
            if "tool_call_cnt" not in state:
                state["tool_call_cnt"] = 0

            # If tool call error, fall back to LLM without tools instead of return failure message
            if "messages" in state:
                last_msg = state["messages"][-1]
                if isinstance(last_msg, AIMessage) and "sorry" in last_msg.content.lower():
                    return "assistant_without_tools"
                tool_msg = state["messages"][-2]
                if isinstance(tool_msg, ToolMessage) and tool_msg.status == "error":
                    return "assistant_without_tools"

            target_node = tools_condition(state)
            if (target_node == "tools"):
                # If tool calling count exceeds limit, fallback to LLM without tools
                if (max_tool_call is not None) and state["tool_call_cnt"] >= max_tool_call:
                    return "assistant_without_tools"
            return target_node

        builder = StateGraph(MyMessagesState)
        # Add nodes
        builder.add_node("assistant_with_tools", assistant_with_tools)
        builder.add_node("assistant_without_tools", assistant_without_tools)
        builder.add_node("tools", ToolNode(tools))
        builder.add_node("increment_tool_call_cnt", increment_tool_call_cnt)
        # Add edges
        builder.add_edge(START, "assistant_with_tools")
        builder.add_conditional_edges("assistant_with_tools", check_tool_call_status)
        builder.add_edge("tools", "increment_tool_call_cnt")
        builder.add_edge("increment_tool_call_cnt", "assistant_with_tools")
        builder.add_edge("assistant_without_tools", END)
        return builder.compile()

    def ask(self, question: str, final_answer_only: bool=True) -> str:
        if "gemma" in self.model_name:
            messages = [
                HumanMessage(f"{system_prompt} {question}")
           ]
        else:
            messages = [
                sys_msg,
                HumanMessage(question)
            ]
        messages = self.graph.invoke({"messages": messages, "tool_call_cnt": 0})
        response = messages["messages"][-1].content.strip()
        return self.format_output(response) if final_answer_only else response

    def format_output(self, output: str) -> str:
        # Find the text behind "FINAL ANSWER: "
        if "FINAL ANSWER: " in output:
            return output.split("FINAL ANSWER: ")[-1].strip()
        # Find the first text between ** and ** (bold markdown)
        elif "**" in output:
            return output.split("**")[1].strip()
        else:
            return output

    def __call__(self, question: str) -> str:
        return self.ask(question)

if __name__ == "__main__":
    from main import main as main_fn
    main_fn()