agent.py

import os
from dotenv import load_dotenv

# Load environment variables
load_dotenv()

# Set protobuf implementation to avoid C++ extension issues
os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"

# Load keys from environment
groq_api_key = os.getenv("GROQ_API_KEY")
serper_api_key = os.getenv("SERPER_API_KEY")
hf_token = os.getenv("HUGGINGFACE_INFERENCE_TOKEN")

# ---- Imports ----
from langgraph.graph import START, StateGraph, MessagesState
from langgraph.prebuilt import tools_condition, ToolNode
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 Chroma
from langchain_core.documents import Document
from langchain_core.messages import SystemMessage, HumanMessage
from langchain_core.tools import tool
from langchain.tools.retriever import create_retriever_tool
from langchain.vectorstores import Chroma
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.schema import Document
import json

# ---- Tools ----

@tool
def multiply(a: int, b: int) -> int:
    return a * b

@tool
def add(a: int, b: int) -> int:
    return a + b

@tool
def subtract(a: int, b: int) -> int:
    return a - b

@tool
def divide(a: int, b: int) -> float:
    if b == 0:
        raise ValueError("Cannot divide by zero.")
    return a / b

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

@tool
def wiki_search(query: str) -> str:
    search_docs = WikipediaLoader(query=query, load_max_docs=2).load()
    formatted = "\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}

@tool
def web_search(query: str) -> str:
    search_docs = TavilySearchResults(max_results=3).invoke(query=query)
    formatted = "\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}

@tool
def arvix_search(query: str) -> str:
    search_docs = ArxivLoader(query=query, load_max_docs=3).load()
    formatted = "\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}

# ---- Embedding & Vector Store Setup ----

embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/all-mpnet-base-v2")

json_QA = []
with open('metadata.jsonl', 'r') as jsonl_file:
    for line in jsonl_file:
        json_QA.append(json.loads(line))

documents = [
    Document(
        page_content=f"Question : {sample['Question']}\n\nFinal answer : {sample['Final answer']}",
        metadata={"source": sample["task_id"]}
    )
    for sample in json_QA
]

vector_store = Chroma.from_documents(
    documents=documents,
    embedding=embeddings,
    persist_directory="./chroma_db",
    collection_name="my_collection"
)
vector_store.persist()
print("Documents inserted:", vector_store._collection.count())

@tool
def similar_question_search(query: str) -> str:
    matched_docs = vector_store.similarity_search(query, 3)
    formatted = "\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 matched_docs
        ]
    )
    return {"similar_questions": formatted}

# ---- System Prompt ----

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...
"""

sys_msg = SystemMessage(content=system_prompt)

# ---- Tool List ----

tools = [
    multiply, add, subtract, divide, modulus,
    wiki_search, web_search, arvix_search, similar_question_search
]

# ---- Graph Definition ----

def build_graph(provider: str = "groq"):
    if provider == "groq":
        llm = ChatGroq(model="qwen-qwq-32b", temperature=0, api_key=groq_api_key)
    elif provider == "google":
        llm = ChatGoogleGenerativeAI(model="gemini-2.0-flash", temperature=0)
    elif provider == "huggingface":
        llm = ChatHuggingFace(
            llm=HuggingFaceEndpoint(repo_id="mosaicml/mpt-30b", temperature=0)
        )
    else:
        raise ValueError("Invalid provider: choose 'groq', 'google', or 'huggingface'.")

    llm_with_tools = llm.bind_tools(tools)

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

    def retriever(state: MessagesState):
        similar = vector_store.similarity_search(state["messages"][0].content)
        if similar:
            example_msg = HumanMessage(content=f"Here is a similar question:\n\n{similar[0].page_content}")
            return {"messages": [sys_msg] + state["messages"] + [example_msg]}
        return {"messages": [sys_msg] + state["messages"]}

    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")

    return builder.compile()