agent.py

import os
import pandas as pd
from langchain_core.messages import HumanMessage, AIMessage
from langgraph.graph import StateGraph, MessagesState
from langgraph.prebuilt import ToolNode
from langchain_huggingface import HuggingFaceEndpoint
from tools import TOOLS  # your dictionary of tool functions

# --- Load local QA metadata for retriever ---
QA_PATH = "metadata.jsonl"
qa_pairs = pd.read_json(QA_PATH, lines=True)
qa_dict = {
    row["Question"].strip(): row["Final answer"].strip()
    for _, row in qa_pairs.iterrows()
}

# === LangGraph builder ===
def build_graph():
    llm = HuggingFaceEndpoint(
        repo_id="mistralai/Mistral-7B-Instruct-v0.3",
        task="text-generation",
        huggingfacehub_api_token=os.environ["HF_TOKEN"]
    )

    # Node 1: Retriever
    def retriever_node(state: MessagesState):
        query = state["messages"][-1].content.strip()
        if query in qa_dict:
            print("✅ Exact match found in retriever.")
            return {"messages": [AIMessage(content=qa_dict[query])]}
        print("🔍 No match. Sending to LLM.")
        return {"messages": state["messages"]}

    # Node 2: Assistant (LLM response parsing)
    def assistant_node(state: MessagesState):
        query = state["messages"][-1].content.strip()

        prompt = (
            "You are a helpful assistant for GAIA benchmark.\n"
            "If you can answer directly, output ONLY the answer.\n"
            "If you need to use a tool, reply in this format:\n"
            "use_tool: <tool_name>: <tool_input>\n"
            "Never explain anything."
        )

        full_prompt = f"{prompt}\n\nQuestion: {query}\nAnswer:"
        response = llm.invoke(full_prompt).strip()
        print(f"🧠 LLM said: {response}")

        if response.startswith("use_tool:"):
            return {
                "messages": state["messages"] + [AIMessage(content=response)],
                "tool_call": response  # carry tool signal
            }
        else:
            return {"messages": [AIMessage(content=response)]}

    # Node 3: Tool execution
    def tool_node(state: MessagesState):
        try:
            tool_signal = state.get("tool_call", "")
            _, tool_name, tool_input = tool_signal.split(":", 2)
            tool_name = tool_name.strip()
            tool_input = tool_input.strip()
            tool_fn = TOOLS.get(tool_name)

            if not tool_fn:
                print(f"❌ Unknown tool: {tool_name}")
                return {"messages": [AIMessage(content="Unknown")]}
            
            print(f"🔧 Using tool: {tool_name} with input: {tool_input}")
            tool_result = tool_fn(tool_input)
            return {"messages": [AIMessage(content=str(tool_result))]}

        except Exception as e:
            print(f"⚠️ Tool error: {e}")
            return {"messages": [AIMessage(content="Unknown")]}  # fail-safe

    # Build LangGraph
    builder = StateGraph(MessagesState)
    builder.add_node("retriever", retriever_node)
    builder.add_node("assistant", assistant_node)
    builder.add_node("tool", tool_node)

    builder.set_entry_point("retriever")
    builder.add_edge("retriever", "assistant")
    builder.add_edge("assistant", "tool")
    builder.add_edge("tool", "assistant")
    builder.set_finish_point("assistant")

    return builder.compile()

# === BasicAgent wrapper ===
class BasicAgent:
    def __init__(self):
        print("BasicAgent initialized with retriever + LLM + manual tool logic.")
        self.graph = build_graph()

    def __call__(self, question: str) -> str:
        print(f"📥 Question: {question[:100]}")
        result = self.graph.invoke({"messages": [HumanMessage(content=question)]})
        answer = result["messages"][-1].content.strip()
        print(f"📤 Answer: {answer}")
        return answer