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

from PIL import Image as PILImage
from huggingface_hub import InferenceClient

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=2, lang="en").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 = os.getenv("HF_API_TOKEN")
HF_INFERENCE_CLIENT = None
if HF_API_TOKEN:
    HF_INFERENCE_CLIENT = InferenceClient(token=HF_API_TOKEN)
else:
    print("WARNING: HF_API_TOKEN not set. Image and Audio tools will not function.")

@tool
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, returns a prompt to use specific tools."""
    try:
        _, file_extension = os.path.splitext(file_path)
        file_extension = file_extension.lower()

        if 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}
        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}
        elif file_extension in (".jpeg", ".jpg", ".png"):
            return {"file_type": "image", "file_name": file_path, "file_content": f"Image file '{file_path}' detected. Use 'describe_image' tool to get a textual description."}
        elif file_extension == ".mp3":
            # For MP3, we indicate it's an audio file and expect the LLM to handle the blob directly.
            # In a real Langchain setup, you might actually read the bytes here and pass them
            # as a part of the message content to the LLM if it supports direct binary upload.
            # For now, this tool simply confirms its type for the agent.
            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."}
        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 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)}

@tool
def describe_image(image_path: str) -> Dict[str, str]:
    """Generates a textual description for an image file (JPEG, JPG, PNG) using an image-to-text model from the Hugging Face Inference API. Requires HF_API_TOKEN environment variable to be set."""
    if not HF_INFERENCE_CLIENT:
        return {"error": "Hugging Face API token not configured for image description. Cannot use this tool."}
    try:
        with open(image_path, "rb") as f:
            image_bytes = f.read()
        description = HF_INFERENCE_CLIENT.image_to_text(image_bytes)
        return {"image_description": description, "image_path": image_path}
    except FileNotFoundError:
        return {"error": f"Image file not found: {image_path}. Please ensure the file exists."}
    except Exception as e:
        return {"error": f"Error describing image {image_path}: {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")
GEMINI_API_KEY = os.getenv("GEMINI_API_KEY")

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

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="gemini-2.5-pro-preview-05-06",
            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"]
        
        # When sending messages to Gemini, if read_file_content identified an audio file,
        # you'll need to ensure the actual binary content of the audio file is included
        # in the message parts for the LLM to process it natively.
        # This part requires a bit more advanced handling than just text.
        # Langchain often handles this when you use `tool_code.File(...)` or similar constructs.
        # For simplicity in this prompt and code example, we're assuming the framework
        # will correctly pass the file content if `read_file_content` returns an audio type.
        
        # A more robust implementation would involve modifying the `assistant` node
        # to explicitly read the file bytes and add them to the message parts
        # if a file is detected in the input state.
        
        # Example of how you might include binary content (conceptual, depends on LangChain/API):
        # new_messages_to_send = []
        # for msg in messages_to_send:
        #    if isinstance(msg, HumanMessage) and "audio file" in msg.content: # Simplified check
        #        # Assume you can get the actual file path from the context
        #        file_path_from_context = "Strawberry pie.mp3" # Or extract from msg.content
        #        if os.path.exists(file_path_from_context):
        #            with open(file_path_from_context, "rb") as f:
        #                audio_bytes = f.read()
        #            new_messages_to_send.append(
        #                HumanMessage(
        #                    content=[
        #                        {"type": "text", "text": "Here is the audio file:"},
        #                        {"type": "media", "media_type": "audio/mp3", "data": audio_bytes}
        #                    ]
        #                )
        #            )
        #    else:
        #        new_messages_to_send.append(msg)
        # llm_response = llm_with_tools.invoke(new_messages_to_send)

        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