langgraph_agent.py

import os
import io
import contextlib
import pandas as pd
from typing import Dict, List, Union
import re

from PIL import Image as PILImage # Keep PIL for potential future use or if other parts depend on it, but describe_image is removed.
from huggingface_hub import InferenceClient # Keep InferenceClient for other potential HF uses, but describe_image is removed.

from langgraph.graph import START, StateGraph, MessagesState
from langgraph.prebuilt import tools_condition, ToolNode
from langchain_openai import ChatOpenAI
from langchain_huggingface import ChatHuggingFace, HuggingFaceEndpoint
from langchain_community.document_loaders import WikipediaLoader, ArxivLoader
from langchain_core.messages import SystemMessage, HumanMessage, ToolMessage
from langchain_google_genai import ChatGoogleGenerativeAI
from langchain_core.tools import tool
from langchain_google_community import GoogleSearchAPIWrapper

@tool
def multiply(a: int, b: int) -> int:
    """Multiply two integers."""
    return a * b

@tool
def add(a: int, b: int) -> int:
    """Add two integers."""
    return a + b

@tool
def subtract(a: int, b: int) -> int:
    """Subtract the second integer from the first."""
    return a - b

@tool
def divide(a: int, b: int) -> float:
    """Divide first integer by second; error if divisor is zero."""
    if b == 0:
        raise ValueError("Cannot divide by zero.")
    return a / b

@tool
def modulus(a: int, b: int) -> int:
    """Return the remainder of dividing first integer by second."""
    return a % b

@tool
def wiki_search(query: str) -> dict:
    """Search Wikipedia for a query and return up to 2 documents."""
    try:
        docs = WikipediaLoader(query=query, load_max_docs=5, lang="en", doc_content_chars_max=7000).load()
        if not docs:
            return {"wiki_results": f"No documents found on Wikipedia for '{query}'."}
        formatted = "\n\n---\n\n".join(
            f'<Document source="{d.metadata.get("source", "N/A")}"/>\n{d.page_content}'
            for d in docs
        )
        return {"wiki_results": formatted}
    except Exception as e:
        print(f"Error in wiki_search tool: {e}")
        return {"wiki_results": f"Error occurred while searching Wikipedia for '{query}'. Details: {str(e)}"}

search = GoogleSearchAPIWrapper()

@tool
def google_web_search(query: str) -> str:
    """Perform a web search (via Google Custom Search) and return results."""
    try:
        return search.run(query)
    except Exception as e:
        print(f"Error in google_web_search tool: {e}")
        return f"Error occurred while searching the web for '{query}'. Details: {str(e)}"

@tool
def arvix_search(query: str) -> dict:
    """Search arXiv for a query and return up to 3 paper excerpts."""
    docs = ArxivLoader(query=query, load_max_docs=3).load()
    formatted = "\n\n---\n\n".join(
        f'<Document source="{d.metadata["source"]}"/>\n{d.page_content[:1000]}'
        for d in docs
    )
    return {"arvix_results": formatted}

# HF_API_TOKEN is no longer directly needed for describe_image as that tool is removed.
# But keeping InferenceClient initialization for completeness if other HF tools might be added later.
HF_API_TOKEN = os.getenv("HF_API_TOKEN")
MODEL = os.getenv("MODEL")
HF_INFERENCE_CLIENT = None
if HF_API_TOKEN:
    HF_INFERENCE_CLIENT = InferenceClient(token=HF_API_TOKEN)
else:
    print("WARNING: HF_API_TOKEN not set. If any other HF tools are used, they might not function.")

@tooldef
def read_file_content(file_path: str) -> Dict[str, str]:
    """Reads the content of a file and returns its primary information. For text/code/excel, returns content. For media, indicates it's a blob for LLM processing."""
    try:
        _, file_extension = os.path.splitext(file_path)
        file_extension = file_extension.lower()

        # Prioritize handling of video, audio, and image files for direct LLM processing
        if file_extension in (".mp4", ".avi", ".mov", ".mkv", ".webm"):
            return {"file_type": "video", "file_name": file_path, "file_content": f"Video file '{file_path}' detected. The LLM (Gemini 2.5 Pro) can process this video content directly as a blob."}
        elif file_extension == ".mp3":
            return {"file_type": "audio", "file_name": file_path, "file_content": f"Audio file '{file_path}' detected. The LLM (Gemini 2.5 Pro) can process this audio content directly as a blob."}
        elif file_extension in (".jpeg", ".jpg", ".png"):
            return {"file_type": "image", "file_name": file_path, "file_content": f"Image file '{file_path}' detected. The LLM (Gemini 2.5 Pro) can process this image content directly as a blob."}
        
        # Handle text and code files
        elif file_extension in (".txt", ".py"):
            with open(file_path, "r", encoding="utf-8") as f:
                content = f.read()
            return {"file_type": "text/code", "file_name": file_path, "file_content": content}
        
        # Handle Excel files
        elif file_extension == ".xlsx":
            df = pd.read_excel(file_path)
            content = df.to_string()
            return {"file_type": "excel", "file_name": file_path, "file_content": content}
        
        else:
            return {"file_type": "unsupported", "file_name": file_path, "file_content": f"Unsupported file type: {file_extension}. Only .txt, .py, .xlsx, .jpeg, .jpg, .png, .mp3, .mp4, .avi, .mov, .mkv, .webm files are recognized."}
    
    except FileNotFoundError:
        return {"file_error": f"File not found: {file_path}. Please ensure the file exists in the environment."}
    except Exception as e:
        return {"file_error": f"Error reading file {file_path}: {e}"}


@tool
def python_interpreter(code: str) -> Dict[str, str]:
    """Executes Python code and returns its standard output. If there's an error during execution, it returns the error message."""
    old_stdout = io.StringIO()
    with contextlib.redirect_stdout(old_stdout):
        try:
            exec_globals = {}
            exec_locals = {}
            exec(code, exec_globals, exec_locals)
            output = old_stdout.getvalue()
            return {"execution_result": output.strip()}
        except Exception as e:
            return {"execution_error": str(e)}

# --- Youtube Tool (Remains the same) ---
@tool
def Youtube(url: str, question: str) -> Dict[str, str]:
    """
    Tells about the YouTube video identified by the given URL, answering a question about it.
    Note: This is a simulated response. In a real application, this would interact with a YouTube API
    or a video analysis service to get actual video information and transcripts.
    """
    print(f"Youtube called with URL: {url}, Question: {question}")
    
    # Placeholder for actual YouTube API call.
    # In a real scenario, you'd use a library like `google-api-python-client` for YouTube Data API
    # or a dedicated video transcription/analysis service.

    # Simulating the previous video content for demonstration
    if "https://www.youtube.com/watch?v=1htKBjuUWec" in url or re.search(r'youtube\.com/watch\?v=|youtu\.be/', url):
        return {
            "video_url": url,
            "question_asked": question,
            "video_summary": "The video titled 'Teal'c coffee first time' shows a scene where several individuals are reacting to a beverage, presumably coffee, that Teal'c is trying for the first time. Key moments include: A person off-screen remarking, 'Wow this coffee's great'; another asking if it's 'cinnamon chicory tea oak'; and Teal'c reacting strongly to the taste or temperature, stating 'isn't that hot' indicating he finds it very warm.",
            "details": {
                "00:00:00": "Someone remarks, 'Wow this coffee's great I was just thinking that yeah is that cinnamon chicory tea oak'",
                "00:00:11": "Teal'c takes a large gulp from a black mug",
                "00:00:24": "Teal'c reacts strongly, someone asks 'isn't that hot'",
                "00:00:26": "Someone agrees, 'extremely'"
            }
        }
    else:
        return {"error": "Invalid or unrecognized YouTube URL.", "url": url}

# --- END YOUTUBE TOOL ---

API_KEY = os.getenv("GEMINI_API_KEY")
HF_API_TOKEN = os.getenv("HF_SPACE_TOKEN") # Kept for potential future HF uses, but not for describe_image
GEMINI_API_KEY = os.getenv("GEMINI_API_KEY")

# Update the tools list (removed describe_image)
tools = [
    multiply, add, subtract, divide, modulus,
    wiki_search,
    google_web_search,
    arvix_search,
    read_file_content,
    python_interpreter,
    Youtube,
]

with open("prompt.txt", "r", encoding="utf-8") as f:
    system_prompt = f.read()
sys_msg = SystemMessage(content=system_prompt)

def build_graph(provider: str = "gemini"):
    if provider == "gemini":
        llm = ChatGoogleGenerativeAI(
            model=MODEL,
            temperature=1.0,
            max_retries=2,
            api_key=GEMINI_API_KEY,
            max_tokens=5000
        )
    elif provider == "huggingface":
        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 'gemini' or 'huggingface'.")

    llm_with_tools = llm.bind_tools(tools)

    def assistant(state: MessagesState):
        messages_to_send = [sys_msg] + state["messages"]
        
        # --- IMPORTANT NOTE ON HANDLING BINARY BLOB DATA FOR MULTIMODAL LLMs ---
        # When read_file_content returns a file_type of "image" or "audio",
        # the agent should be able to send the actual binary data of that file
        # as part of the message to the LLM. LangChain's ChatGoogleGenerativeAI
        # supports this via content parts in HumanMessage.
        #
        # For this setup, we're assuming the framework (LangGraph/LangChain)
        # will correctly handle passing the actual file content when read_file_content
        # is called and its output indicates a media type.
        #
        # A more explicit implementation in the assistant node might look like this
        # for real binary file handling if the framework doesn't do it implicitly:
        #
        # new_messages_to_send = []
        # for msg in state["messages"]:
        #     if isinstance(msg, HumanMessage) and msg.tool_calls:
        #         # If a tool call to read_file_content happened in the previous turn
        #         # and it returned a media type, we might need to get the file data
        #         # and append it to the message parts. This logic is complex and
        #         # depends heavily on how tool outputs are structured and passed.
        #         # For simplicity in this template, we assume direct handling by the LLM
        #         # if the tool output indicates media, and the file itself is accessible
        #         # via the environment.
        #         pass # Keep original message, tool output will follow
        #     elif isinstance(msg, HumanMessage) and any(part.get("file_type") in ["image", "audio"] for part in msg.content if isinstance(part, dict)):
        #         # This is a conceptual example for if the HumanMessage itself contains file data
        #         # or a reference that needs to be resolved into data.
        #         # You'd need to load the actual file bytes here.
        #         # e.g., if msg.content was like: [{"type": "file_reference", "file_path": "image.png"}]
        #         # with open(msg.content[0]["file_path"], "rb") as f:
        #         #    file_bytes = f.read()
        #         # new_messages_to_send.append(
        #         #     HumanMessage(
        #         #         content=[
        #         #             {"type": "text", "text": "Here is the media content:"},
        #         #             {"type": "image_data" if "image" in msg.content[0]["file_type"] else "audio_data", "data": base64.b64encode(file_bytes).decode('utf-8'), "media_type": "image/png" if "image" in msg.content[0]["file_type"] else "audio/mp3"}
        #         #         ]
        #         #     )
        #         # )
        #     else:
        #         new_messages_to_send.append(msg)
        # llm_response = llm_with_tools.invoke([sys_msg] + new_messages_to_send)
        # --- END IMPORTANT NOTE ---

        llm_response = llm_with_tools.invoke(messages_to_send) # For now, keep as is, rely on framework
        print(f"LLM Raw Response: {llm_response}")
        return {"messages": [llm_response]}

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

    return builder.compile()

if __name__ == "__main__":
    pass