agent.py

import os

from datasets import load_dataset
from dotenv import load_dotenv
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.schema import Document
from langchain.tools.retriever import create_retriever_tool
from langchain.vectorstores import Chroma
from langchain_community.document_loaders import ArxivLoader, WikipediaLoader
from langchain_community.tools.tavily_search import TavilySearchResults
from langchain_core.messages import HumanMessage, SystemMessage
from langchain_core.tools import tool
from langchain_google_genai import ChatGoogleGenerativeAI
from langchain_groq import ChatGroq
from langchain_huggingface import (ChatHuggingFace, HuggingFaceEmbeddings,
                                   HuggingFaceEndpoint)
from langgraph.graph import START, MessagesState, StateGraph
from langgraph.prebuilt import ToolNode, tools_condition
from huggingface_hub import login

login(token=os.environ["HUGGINGFACE_TOKEN"])

load_dotenv()

@tool
def calculator(query: str) -> str:
    """Perform basic arithmetic operations based on the provided query.

    Args:
        query: A mathematical query as a string, e.g., '2 + 2' or '5 * 6'."""

    try:
        # Evaluate the mathematical expression
        result = eval(query)
        return {"calculator_result": str(result)}
    except Exception as e:
        return {"error": f"Error evaluating the expression: {str(e)}"}


@tool
def wiki_search(query: str) -> str:
    """Search Wikipedia for a query and return maximum 2 results.

    Args:
        query: The search query."""
    search_docs = WikipediaLoader(query=query, load_max_docs=2).load()
    formatted_search_docs = "\n\n---\n\n".join(
        [
            f'<Document source="{doc.metadata["source"]}" page="{doc.metadata.get("page", "")}"/>\n{doc.page_content}\n</Document>'
            for doc in search_docs
        ]
    )
    return {"wiki_results": formatted_search_docs}


@tool
def web_search(query: str) -> str:
    """Search Tavily for a query and return maximum 3 results.

    Args:
        query: The search query."""
    search_docs = TavilySearchResults(max_results=3).invoke(query=query)
    formatted_search_docs = "\n\n---\n\n".join(
        [
            f'<Document source="{doc.metadata["source"]}" page="{doc.metadata.get("page", "")}"/>\n{doc.page_content}\n</Document>'
            for doc in search_docs
        ]
    )
    return {"web_results": formatted_search_docs}


@tool
def arvix_search(query: str) -> str:
    """Search Arxiv for a query and return maximum 3 result.

    Args:
        query: The search query."""
    search_docs = ArxivLoader(query=query, load_max_docs=3).load()
    formatted_search_docs = "\n\n---\n\n".join(
        [
            f'<Document source="{doc.metadata["source"]}" page="{doc.metadata.get("page", "")}"/>\n{doc.page_content[:1000]}\n</Document>'
            for doc in search_docs
        ]
    )
    return {"arvix_results": formatted_search_docs}


system_prompt = """You are a helpful assistant tasked with answering questions using a set of tools. 
Now, I will ask you a question. Report your thoughts, and finish your answer with the following template: 
FINAL ANSWER: [YOUR FINAL ANSWER]. 
YOUR FINAL ANSWER should be a number OR as few words as possible OR a comma separated list of numbers and/or strings. If you are asked for a number, don't use comma to write your number neither use units such as $ or percent sign unless specified otherwise. If you are asked for a string, don't use articles, neither abbreviations (e.g. for cities), and write the digits in plain text unless specified otherwise. If you are asked for a comma separated list, apply the above rules depending of whether the element to be put in the list is a number or a string.
Your answer should only start with "FINAL ANSWER: ", then follows with the answer. """

# System message
sys_msg = SystemMessage(content=system_prompt)

# build a retriever
embeddings = HuggingFaceEmbeddings(
    model_name="sentence-transformers/all-mpnet-base-v2"
)  #  dim=768

# Load the GAIA validation dataset
dataset = load_dataset("gaia-benchmark/GAIA", name="2023_level1", split="validation")

# Prepare the embeddings model
embeddings = HuggingFaceEmbeddings(model_name="all-MiniLM-L6-v2")

# Extract questions and their answers
documents = []
for entry in dataset:
    question = entry["Question"]
    answer = entry["Final answer"]

    # Create a document with both the question and the answer as metadata
    metadata = {
        "task_id": entry["task_id"],
        "steps": entry["Annotator Metadata"]["Steps"],
        "tools": entry["Annotator Metadata"]["Tools"],
        "answer": answer,
    }

    # Add the question to the list of documents
    documents.append(Document(page_content=question, metadata=metadata))

# Insert the documents into Chroma
vectorstore = Chroma.from_documents(
    documents=documents,
    embedding=embeddings,
    collection_name="gaia_validation",
    persist_directory="./chroma_store",
)

create_retriever_tool = create_retriever_tool(
    retriever=vectorstore.as_retriever(),
    name="Question Search",
    description="A tool to retrieve similar questions from a vector store.",
)


tools = [
    calculator,
    wiki_search,
    web_search,
    arvix_search,
]


# Build graph function
def build_graph(provider: str = "groq"):
    """Build the graph"""
    # Load environment variables from .env file
    if provider == "google":
        # Google Gemini
        llm = ChatGoogleGenerativeAI(model="gemini-2.0-flash", temperature=0)
    elif provider == "groq":
        # Groq https://console.groq.com/docs/models
        llm = ChatGroq(
            model="qwen-qwq-32b", temperature=0
        )  # optional : qwen-qwq-32b gemma2-9b-it
    elif provider == "huggingface":
        # TODO: Add huggingface endpoint
        llm = ChatHuggingFace(
            llm=HuggingFaceEndpoint(
                url="https://api-inference.huggingface.co/models/Meta-DeepLearning/llama-2-7b-chat-hf",
                temperature=0,
            ),
        )
    else:
        raise ValueError("Invalid provider. Choose 'google', 'groq' or 'huggingface'.")
    # Bind tools to LLM
    llm_with_tools = llm.bind_tools(tools)

    # Node
    def assistant(state: MessagesState):
        """Assistant node"""
        return {"messages": [llm_with_tools.invoke(state["messages"])]}

    def retriever(state: MessagesState):
        """Retriever node"""
        similar_question = vectorstore.similarity_search(state["messages"][0].content)
        example_msg = HumanMessage(
            content=f"Here I provide a similar question and answer for reference: \n\n{similar_question[0].page_content}",
        )
        return {"messages": [sys_msg] + state["messages"] + [example_msg]}

    builder = StateGraph(MessagesState)
    builder.add_node("retriever", retriever)
    builder.add_node("assistant", assistant)
    builder.add_node("tools", ToolNode(tools))
    builder.add_edge(START, "retriever")
    builder.add_edge("retriever", "assistant")
    builder.add_conditional_edges(
        "assistant",
        tools_condition,
    )
    builder.add_edge("tools", "assistant")

    # Compile graph
    return builder.compile()


# test
if __name__ == "__main__":
    question = "When was a picture of St. Thomas Aquinas first added to the Wikipedia page on the Principle of double effect?"
    # Build the graph
    graph = build_graph(provider="groq")
    # Run the graph
    messages = [HumanMessage(content=question)]
    messages = graph.invoke({"messages": messages})
    for m in messages["messages"]:
        m.pretty_print()