agent.py

import os
from dotenv import load_dotenv
from supabase.client import create_client
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

load_dotenv()


def load_system_prompt(path: str = "system_prompt.txt") -> SystemMessage:
    """
    Load system prompt from a file, fallback to a default if missing.

    Args:
        path: File path to the system prompt.

    Returns:
        SystemMessage containing the loaded or default prompt.
    """
    try:
        with open(path, encoding="utf-8") as f:
            content = f.read()
    except FileNotFoundError:
        content = "You are a helpful assistant."
    return SystemMessage(content=content)


def math_tool(func):
    """
    Wrap a Python function as a LangChain tool.

    Args:
        func: Callable to wrap.

    Returns:
        A LangChain tool.
    """
    return tool(func)


@math_tool
def add(a: int, b: int) -> int:
    """Return a + b."""
    return a + b


@math_tool
def subtract(a: int, b: int) -> int:
    """Return a - b."""
    return a - b


@math_tool
def multiply(a: int, b: int) -> int:
    """Return a * b."""
    return a * b


@math_tool
def divide(a: int, b: int) -> float:
    """
    Return a / b.

    Raises:
        ValueError: If b is zero.
    """
    if b == 0:
        raise ValueError("Cannot divide by zero.")
    return a / b


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


def format_docs(docs, key: str, max_chars: int = None) -> dict:
    """
    Convert document list into labeled XML-style chunks.

    Args:
        docs: Iterable of Document objects.
        key: Dict key for formatted results.
        max_chars: Optionally truncate content.

    Returns:
        {key: formatted_string}
    """
    entries = []
    for d in docs:
        content = d.page_content if max_chars is None else d.page_content[:max_chars]
        entries.append(
            f'<Document source="{d.metadata.get("source","")}" page="{d.metadata.get("page","")}">\n'
            f"{content}\n</Document>"
        )
    return {key: "\n\n---\n\n".join(entries)}


@tool
def wiki_search(query: str) -> dict:
    """Search Wikipedia (2 docs) and format results."""
    docs = WikipediaLoader(query=query, load_max_docs=2).load()
    return format_docs(docs, "wiki_results")


@tool
def web_search(query: str) -> dict:
    """Search the web via Tavily (3 docs) and format results."""
    docs = TavilySearchResults(max_results=3).invoke(query=query)
    return format_docs(docs, "web_results")


@tool
def arxiv_search(query: str) -> dict:
    """Search ArXiv (3 docs) and format results (truncate to 1k chars)."""
    docs = ArxivLoader(query=query, load_max_docs=3).load()
    return format_docs(docs, "arxiv_results", max_chars=1000)


def build_vector_retriever():
    """
    Create and return a Supabase-based vector retriever.

    Returns:
        Retriever for semantic similarity queries.
    """
    embed = HuggingFaceEmbeddings("sentence-transformers/all-mpnet-base-v2")
    supa = create_client(
        os.getenv("SUPABASE_URL"), os.getenv("SUPABASE_SERVICE_KEY")
    )
    store = SupabaseVectorStore(
        client=supa,
        embedding=embed,
        table_name="documents",
        query_name="match_documents_langchain",
    )
    return store.as_retriever()


def get_llm(provider: str = "google"):
    """
    Factory to select and return an LLM client.

    Args:
        provider: One of "google", "groq", "huggingface".

    Returns:
        Configured LLM client.

    Raises:
        ValueError: On unsupported provider.
    """
    if provider == "google":
        return ChatGoogleGenerativeAI("gemini-2.0-flash", temperature=0)
    if provider == "groq":
        return ChatGroq("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}")


def build_graph(provider: str = "google"):
    """
    Build and compile a StateGraph for retrieval + LLM responses.

    Args:
        provider: LLM provider key.

    Returns:
        A compiled StateGraph.
    """
    sys_msg = load_system_prompt()
    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,
        arxiv_search,
        question_tool,
    ]
    llm = get_llm(provider).bind_tools(tools)

    def retriever_node(state: MessagesState) -> dict:
        """
        Node: retrieve most relevant doc and extract its answer.
        """
        query = state["messages"][-1].content
        doc = retriever.similarity_search(query, k=1)[0]
        text = doc.page_content
        ans = text.split("Final answer :")[-1].strip() if "Final answer :" in text else text
        return {"messages": [AIMessage(content=ans)]}

    def assistant_node(state: MessagesState) -> dict:
        """
        Node: call LLM with system prompt + history.
        """
        msgs = [sys_msg] + state["messages"]
        resp = llm.invoke({"messages": msgs})
        return {"messages": [resp]}

    graph = StateGraph(MessagesState)
    graph.add_node("retriever", retriever_node)
    graph.add_node("assistant", assistant_node)
    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()