agent.py

import os
from dotenv import load_dotenv

# Import models from SmolaAgents
from smolagents import CodeAgent, LiteLLMModel, OpenAIServerModel

# Import SmolaAgents tools
from smolagents.default_tools import FinalAnswerTool, PythonInterpreterTool

# Import custom tools
from tools import (
    AddDocumentToVectorStoreTool,
    ArxivSearchTool,
    DownloadFileFromLinkTool,
    DuckDuckGoSearchTool,
    QueryVectorStoreTool,
    ReadFileContentTool,
    TranscibeVideoFileTool,
    TranscribeAudioTool,
    VisitWebpageTool,
    WikipediaSearchTool,
    image_question_answering,
)

# Import utility functions
from utils import extract_final_answer, replace_tool_mentions


class BoomBot:
    def __init__(self, provider="anthropic"):
        """
        Initialize the BoomBot with the specified provider.

        Args:
            provider (str): The model provider to use (e.g., "groq", "qwen", "gemma", "anthropic", "deepinfra", "meta")
        """
        load_dotenv()
        self.provider = provider
        self.model = self._initialize_model()
        self.agent = self._create_agent()

    def _initialize_model(self):
        """
        Initialize the appropriate model based on the provider.

        Returns:
            The initialized model object
        """
        if self.provider == "qwen":
            qwen_model = "ollama_chat/qwen3:8b"
            return LiteLLMModel(
                model_id=qwen_model,
                device="cuda",
                num_ctx=32768,
                temperature=0.6,
                top_p=0.95,
            )
        elif self.provider == "gemma":
            gemma_model = "ollama_chat/gemma3:12b-it-qat"
            return LiteLLMModel(
                model_id=gemma_model,
                num_ctx=65536,
                temperature=1.0,
                device="cuda",
                top_k=64,
                top_p=0.95,
                min_p=0.0,
            )
        elif self.provider == "anthropic":
            model_id = "anthropic/claude-3-5-haiku-latest"
            return LiteLLMModel(
                model_id=model_id,
                temperature=0.6,
                max_tokens=8192,
                api_key=os.getenv("ANTHROPIC_API_KEY"),
            )

        elif self.provider == "deepinfra":
            deepinfra_model = "Qwen/Qwen3-235B-A22B"
            # return OpenAIServerModel(
            #     model_id=deepinfra_model,
            #     api_base="https://api.deepinfra.com/v1/openai",
            #     api_key=os.getenv("ANTHROPIC_API_KEY"),
            #     flatten_messages_as_text=True,
            #     max_tokens=8192,
            #     temperature=0.1,
            # )
            return LiteLLMModel(
                 model_id="deepinfra/"+ deepinfra_model,
                api_base="https://api.deepinfra.com/v1/openai",
                api_key=os.getenv("DEEPINFRA_API_KEY"),
                flatten_messages_as_text=True,
                max_tokens=8192,
                temperature=0.7,
            )
        elif self.provider == "meta":
            meta_model = "meta-llama/Llama-3.3-70B-Instruct-Turbo"
            # return OpenAIServerModel(
            #     model_id=meta_model,
            #     api_base="https://api.deepinfra.com/v1/openai",
            #     api_key=os.getenv("DEEPINFRA_API_KEY"),
            #     flatten_messages_as_text=True,
            #     max_tokens=8192,
            #     temperature=0.7,
            # )
            return LiteLLMModel(
                 model_id="deepinfra/"+ meta_model,
                api_base="https://api.deepinfra.com/v1/openai",
                api_key=os.getenv("DEEPINFRA_API_KEY"),
                flatten_messages_as_text=True,
                max_tokens=8192,
                temperature=0.7,
            )
        elif self.provider == "groq":
            # Default to use groq's claude-3-opus or llama-3
            model_id = "claude-3-opus-20240229"
            return LiteLLMModel(model_id=model_id, temperature=0.7, max_tokens=8192)
        else:
            raise ValueError(f"Unsupported provider: {self.provider}")

    def _create_agent(self):
        """
        Create and configure the agent with all necessary tools.

        Returns:
            The configured CodeAgent
        """
        # Initialize tools
        download_file = DownloadFileFromLinkTool()
        read_file_content = ReadFileContentTool()
        visit_webpage = VisitWebpageTool()
        transcribe_video = TranscibeVideoFileTool()
        transcribe_audio = TranscribeAudioTool()
        get_wikipedia_info = WikipediaSearchTool()
        web_searcher = DuckDuckGoSearchTool()
        arxiv_search = ArxivSearchTool()
        add_doc_vectorstore = AddDocumentToVectorStoreTool()
        retrieve_doc_vectorstore = QueryVectorStoreTool()

        # SmolaAgents default tools
        python_interpreter = PythonInterpreterTool()
        final_answer = FinalAnswerTool()

        # Combine all tools
        agent_tools = [
            web_searcher,
            download_file,
            read_file_content,
            visit_webpage,
            transcribe_video,
            transcribe_audio,
            get_wikipedia_info,
            arxiv_search,
            add_doc_vectorstore,
            retrieve_doc_vectorstore,
            image_question_answering,
            python_interpreter,
            final_answer,
        ]

        # Additional imports for the Python interpreter
        additional_imports = [
            "json",
            "os",
            "glob",
            "pathlib",
            "pandas",
            "numpy",
            "matplotlib",
            "seaborn",
            "sklearn",
            "tqdm",
            "argparse",
            "pickle",
            "io",
            "re",
            "datetime",
            "collections",
            "math",
            "random",
            "csv",
            "zipfile",
            "itertools",
            "functools",
            "open",
            "requests"
        ]

        # Create the agent
        agent = CodeAgent(
            tools=agent_tools,
            max_steps=12,
            model=self.model,
            add_base_tools=False,
            stream_outputs=True,
            additional_authorized_imports=additional_imports,
        )

        # Modify the system prompt
        modified_prompt = replace_tool_mentions(agent.system_prompt)
        agent.system_prompt = modified_prompt

        return agent

    def _get_system_prompt(self):
        """
        Return the system prompt for the agent.

        Returns:
            str: The system prompt
        """
        return """
        YOUR BEHAVIOR GUIDELINES:
          • Do NOT make unfounded assumptions—always ground answers in reliable sources or search results.
          • For math or puzzles: break the problem into code/math, then solve programmatically.
        
        RESEARCH WORKFLOW:
          1. SEARCH  
             - Try web_search, wikipedia_search, or arxiv_search first.  
             - Refine your query rather than repeating the exact same terms.  
             - If one search tool yields insufficient info, switch to another before downloading.
          2. VISIT  
             - Use visit_webpage to extract and read page content when a promising link appears after one of the SEARCH tools.
             - For each visited link, also download the file and add to the vector store, you might need to query this later, especially if you have a lot of search results.
          3. EVALUATE  
             - ✅ If the page or search snippet fully answers the question, respond immediately.  
             - ❌ If not, move on to deeper investigation.
          4. DOWNLOAD  
             - Use download_file_from_link tool on relevant links found (yes you can download webpages as html).  
             - For arXiv papers, target the /pdf/ or DOI link (e.g https://arxiv.org/pdf/2011.10672).
          5. INDEX & QUERY  
             - Add downloaded documents to the vector store with add_document_to_vector_store.  
             - Use query_downloaded_documents for detailed answers.
          6. READ
             - You have access to a read_file_content tool to read most types of files (html, pdf, text). 
             - You can also directly interact with downloaded files (csv, excel) in your python code.
             - Use query_downloaded_documents if you have added docs to vector store.
             
        FALLBACK & ADAPTATION:
          • If a tool fails, reformulate your query or try a different search method before dropping to download.
          • If a tool fails multiple times, try a different tool.
          • For arXiv: you might discover a paper link via web_search tool and then directly use download_file_from_link tool
        
        COMMON TOOL CHAINS (conceptual outlines):
        These are just guidelines, each task might require a unique workflow.
        A tool can provide useful information for the task, it will not always contain the answer. You need to work to get to a final_answer that makes sense.
        
          • FACTUAL Qs:  
              web_search → final_answer  
          • CURRENT EVENTS:  
              To have some summary information use web_search, that might output a promising website to visit and read content from using (visit_webpage or download_file_from_link and read_file_content)
              web_search → visit_webpage → final_answer  
          • DOCUMENT-BASED Qs:  
              web_search → download_file_from_link → add_document_to_vector_store → query_downloaded_documents → final_answer  
          • ARXIV PAPERS:  
              The arxiv search tool provides a list of results with summary content, to inspect the whole paper you need to download it with download_file_from_link tool.
              arxiv_search → download_file_from_link → read_file_content
              If that fails
              arxiv_search → download_file_from_link → add_document_to_vector_store → query_downloaded_documents  
          • MEDIA ANALYSIS:  
              download_file_from_link → transcribe_video/transcribe_audio/describe_image → final_answer
        
        FINAL ANSWER FORMAT:
        ** Do not name any python variables final_answer, this causes problems with tools.
          - Begin with "FINAL ANSWER: "  
          - Number → digits only (e.g., 42)  no units unless specified
          - String → exact text (e.g., Pope Francis) without quotation marks
          - List → comma-separated, one space, no brackets unless specified(e.g., 2, 3, 4)  
          - Conclude with: FINAL ANSWER: <your_answer>
        """

    def run(self, question: str, task_id: str, to_download) -> str:
        """
        Run the agent with the given question, task_id, and download flag.

        Args:
            question (str): The question or task for the agent to process
            task_id (str): A unique identifier for the task
            to_download (Bool): Flag indicating whether to download resources

        Returns:
            str: The agent's response
        """
        prompt = self._get_system_prompt()
        # Task introduction
        prompt += "\nHere is the Task you need to solve:\n\n"
        prompt += f"Task: {question}\n\n"

        # Include download instructions if applicable
        if to_download:
            link = f"https://agents-course-unit4-scoring.hf.space/files/{task_id}"
            prompt += (
                "IMPORTANT: Before solving the task, you must download a required file.\n"
                f"Use the `download_file_from_link` tool with this link: {link}\n"
                "After downloading, use the appropriate tool to read or process the file "
                "before attempting to solve the task.\n\n"
            )

        # Run the agent with the given question
        result = self.agent.run(prompt)

        # Extract the final answer from the result
        final_answer = extract_final_answer(result)

        return final_answer



if __name__ == "__main__":
    import time
    from utils import load_online_qas, extract_final_answer
    import requests
    import json

    agent = BoomBot(provider="gemma")
    file_online = load_online_qas(file_path = r"../../Final_Assignment_Template/allqas.jsonl", has_file=True)
    results = []

    excluded_keywords = ["youtube", "video", "chess"]

    for entry in file_online:
        task_id = entry["task_id"]
        question = entry["Question"]
        real_answer = entry["Final answer"]
        file_name = entry.get("file_name", "")
        to_download = file_name != ""
        link = f"https://agents-course-unit4-scoring.hf.space/files/{task_id}"

        # Check exclusion and file availability
        if any(kw in question.lower() for kw in excluded_keywords):
            llm_answer = "NOT ATTEMPTED"
            processed_answer = llm_answer
        else:
            try:
                response = requests.get(link)
                if response.status_code != 200:
                    llm_answer = "NOT ATTEMPTED"
                    processed_answer = llm_answer
                else:
                 
                    llm_answer = agent.run(question, task_id, to_download)
                    processed_answer = str(extract_final_answer(llm_answer))
                    # time.sleep(10)
            except Exception as e:
                llm_answer = processed_answer = f"[Error] {e}"
                # time.sleep(6)

        results.append({
            "question": question,
            "llm_answer": llm_answer,
            "processed_answer": processed_answer.strip(),
            "real_answer": real_answer
        })

        print("REAL ANSWER:", real_answer)

    # Save all results to file
    with open("llm_eval.json", "w", encoding="utf-8") as f:
        json.dump(results, f, indent=2, ensure_ascii=False)