agent.py

import os
from dotenv import load_dotenv
from langgraph.graph import START, StateGraph, MessagesState
from langgraph.prebuilt import ToolNode, tools_condition
from langchain_core.tools import tool
from langchain_core.messages import SystemMessage, HumanMessage, AIMessage
from langchain_google_genai import ChatGoogleGenerativeAI
from langchain_groq import ChatGroq
from langchain_huggingface import ChatHuggingFace, HuggingFaceEndpoint, HuggingFaceEmbeddings
from langchain_community.tools.tavily_search import TavilySearchResults
from langchain_community.document_loaders import WikipediaLoader, ArxivLoader
from langchain_community.vectorstores import SupabaseVectorStore
from langchain.tools.retriever import create_retriever_tool
from supabase.client import create_client

load_dotenv()

# --- System Prompt Loader ---
def load_system_prompt(path="system_prompt.txt") -> SystemMessage:
    try:
        with open(path, encoding="utf-8") as f:
            return SystemMessage(content=f.read())
    except FileNotFoundError:
        return SystemMessage(content="You are a helpful assistant.")

sys_msg = load_system_prompt()

# --- Math Tools Factory ---
def math_tool(fn):
    return tool(fn)

@math_tool
def add(a: int, b: int) -> int:        return a + b
@math_tool
def subtract(a: int, b: int) -> int:   return a - b
@math_tool
def multiply(a: int, b: int) -> int:   return a * b
@math_tool
def divide(a: int, b: int) -> float:
    if b == 0: raise ValueError("Cannot divide by zero.")
    return a / b

@math_tool
def modulus(a: int, b: int) -> int:    return a % b

# --- Document Formatting Helper ---
def format_docs(docs, key: str, max_chars: int = None) -> dict:
    content = "\n\n---\n\n".join(
        f'<Document source="{d.metadata.get("source","")}" page="{d.metadata.get("page","")}" />\n'
        f'{d.page_content[:max_chars] if max_chars else d.page_content}\n</Document>'
        for d in docs
    )
    return {key: content}

# --- Info Tools ---
@tool
def wiki_search(query: str) -> dict:
    docs = WikipediaLoader(query=query, load_max_docs=2).load()
    return format_docs(docs, "wiki_results")

@tool
def web_search(query: str) -> dict:
    docs = TavilySearchResults(max_results=3).invoke(query=query)
    return format_docs(docs, "web_results")

@tool
def arvix_search(query: str) -> dict:
    docs = ArxivLoader(query=query, load_max_docs=3).load()
    return format_docs(docs, "arvix_results", max_chars=1000)

# --- Vector Retriever Setup ---
def build_vector_retriever():
    embed_model = HuggingFaceEmbeddings(model_name="sentence-transformers/all-mpnet-base-v2")
    supa = create_client(os.getenv("SUPABASE_URL"), os.getenv("SUPABASE_SERVICE_KEY"))
    vs = SupabaseVectorStore(
        client=supa,
        embedding=embed_model,
        table_name="documents",
        query_name="match_documents_langchain"
    )
    return vs.as_retriever()

# --- LLM Factory ---
def get_llm(provider: str):
    if provider == "google":
        return ChatGoogleGenerativeAI(model="gemini-2.0-flash", temperature=0)
    if provider == "groq":
        return ChatGroq(model="qwen-qwq-32b", temperature=0)
    if provider == "huggingface":
        return ChatHuggingFace(llm=HuggingFaceEndpoint(
            url="https://api-inference.huggingface.co/models/Meta-DeepLearning/llama-2-7b-chat-hf",
            temperature=0))
    raise ValueError(f"Unsupported provider: {provider}")

# --- Build Graph ---
def build_graph(provider: str = "google"):
    # tools list
    retriever = build_vector_retriever()
    question_tool = create_retriever_tool(
        retriever=retriever,
        name="Question Search",
        description="Retrieve similar Q&A from vector store"
    )
    tools = [
        add, subtract, multiply, divide, modulus,
        wiki_search, web_search, arvix_search,
        question_tool
    ]

    # LLM w/ tools
    llm = get_llm(provider).bind_tools(tools)

    # Nodes
    def assistant(state: MessagesState):
        msgs = [sys_msg] + state["messages"]
        resp = llm.invoke({"messages": msgs})
        return {"messages": [resp]}

    def retriever_node(state: MessagesState):
        query = state["messages"][-1].content
        doc = retriever.similarity_search(query, k=1)[0]
        text = doc.page_content
        answer = text.split("Final answer :")[-1].strip() if "Final answer :" in text else text
        return {"messages": [AIMessage(content=answer)]}

    # Graph assembly
    graph = StateGraph(MessagesState)
    graph.add_node("retriever", retriever_node)
    graph.add_node("assistant", assistant)
    graph.add_node("tools", ToolNode(tools))
    graph.add_edge(START, "retriever")
    graph.add_edge("retriever", "assistant")
    graph.add_conditional_edges("assistant", tools_condition)
    graph.add_edge("tools", "assistant")
    graph.set_entry_point("retriever")
    graph.set_finish_point("assistant")

    return graph.compile()