agent.py

import os
import json
import re
from typing import Tuple, Dict, Any
from transformers import pipeline, AutoTokenizer, AutoModelForSeq2SeqLM # Import AutoTokenizer and AutoModelForSeq2SeqLM

from tools.asr_tool import transcribe_audio
from tools.excel_tool import analyze_excel
from tools.search_tool import search_duckduckgo
from tools.math_tool import calculate_math # Make sure to import your math tool

class GaiaAgent:
    def __init__(self):
        token = os.getenv("HUGGINGFACEHUB_API_TOKEN")
        if not token:
            raise ValueError("Missing HUGGINGFACEHUB_API_TOKEN environment variable.")

        # Specify the model and load tokenizer and model separately for better control
        model_name = "google/flan-t5-large"
        self.tokenizer = AutoTokenizer.from_pretrained(model_name, token=token)
        self.model = AutoModelForSeq2SeqLM.from_pretrained(model_name, token=token)

        # Use the pipeline with the loaded model and tokenizer
        self.llm = pipeline(
            "text2text-generation",
            model=self.model,
            tokenizer=self.tokenizer,
            device="cpu", # Consider "cuda" if you have a GPU
            max_new_tokens=256,
            do_sample=False, # Set to True if you want to use temperature and top_p/k
            # temperature=0.1, # Removed, as it's not a valid pipeline initialization flag here
        )

        self.system_prompt = """You are a general AI assistant. 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."""

    def extract_final_answer(self, text: str) -> str:
        """Extrahera det slutliga svaret från modellens output"""
        final_answer_match = re.search(r'FINAL ANSWER:\s*(.+?)(?:\n|$)', text, re.IGNORECASE)
        if final_answer_match:
            return final_answer_match.group(1).strip()
        sentences = text.strip().split('\n')
        return sentences[-1].strip() if sentences else text.strip()

    def needs_tool(self, question: str) -> Tuple[str, bool]:
        """Bestäm vilket verktyg som behövs baserat på frågan"""
        question_lower = question.lower()

        if any(ext in question_lower for ext in ['.mp3', '.wav', '.m4a', '.flac']):
            return 'audio', True
        if any(ext in question_lower for ext in ['.xlsx', '.xls', '.csv']):
            return 'excel', True
        if any(keyword in question_lower for keyword in ['search', 'find', 'lookup', 'http', 'www.', 'wikipedia', 'albums', 'discography', 'published', 'website']):
            return 'search', True
        if any(keyword in question_lower for keyword in ['calculate', 'compute', 'sum', 'average', 'count', 'what is', 'solve']):
            return 'math', True
        return 'llm', False

    def process_with_tools(self, question: str, tool_type: str) -> Tuple[str, str]:
        """Bearbeta frågan med specifika verktyg"""
        trace_log = f"Detected {tool_type} task. Processing...\n"

        try:
            if tool_type == 'audio':
                audio_files = re.findall(r'\b[\w\-_]+\.(mp3|wav|m4a|flac)\b', question, re.IGNORECASE)
                if audio_files:
                    result = transcribe_audio(audio_files[0])
                    trace_log += f"Audio transcription: {result}\n"
                    return result, trace_log
                else:
                    return "No audio file mentioned in the question.", trace_log

            elif tool_type == 'excel':
                excel_files = re.findall(r'\b[\w\-_]+\.(xlsx|xls|csv)\b', question, re.IGNORECASE)
                if excel_files:
                    result = analyze_excel(excel_files[0])
                    trace_log += f"Excel analysis: {result}\n"
                    return result, trace_log
                else:
                    return "No Excel file mentioned in the question.", trace_log

            elif tool_type == 'search':
                search_query = question # This might need refinement to extract just the search query
                result = search_duckduckgo(search_query)
                trace_log += f"Search results: {result}\n"
                return result, trace_log

            elif tool_type == 'math':
                math_expression_match = re.search(r'calculate (.+)', question, re.IGNORECASE)
                if math_expression_match:
                    expression = math_expression_match.group(1).strip()
                    result = calculate_math(expression)
                    trace_log += f"Math calculation: {result}\n"
                    return result, trace_log
                else:
                    return "No clear mathematical expression found in the question.", trace_log

        except Exception as e:
            trace_log += f"Error using {tool_type} tool: {str(e)}\n"
            return f"Error: {str(e)}", trace_log

        return "No valid input found for tool", trace_log

    def reason_with_llm(self, question: str, context: str = "") -> Tuple[str, str]:
        """Använd LLM för reasoning med kontext"""
        trace_log = "Using LLM for reasoning...\n"

        # Combine system prompt, context, and question, ensuring it fits token limit
        if context:
            prompt = f"{self.system_prompt}\n\nContext: {context}\n\nQuestion: {question}\n\nPlease analyze this step by step and provide your final answer."
        else:
            prompt = f"{self.system_prompt}\n\nQuestion: {question}\n\nPlease analyze this step by step and provide your final answer."

        # Tokenize and truncate if necessary
        inputs = self.tokenizer(prompt, return_tensors="pt", truncation=True, max_length=self.tokenizer.model_max_length)

        try:
            # Generate response using the model's generate method for more control
            # You can add generation arguments here, e.g., temperature, top_k, etc.
            outputs = self.model.generate(
                inputs.input_ids,
                max_new_tokens=256,
                do_sample=False, # Set to True to enable temperature and other sampling parameters
                # temperature=0.1, # Example: Only if do_sample is True
            )
            response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)

            trace_log += f"LLM response: {response}\n"
            return response, trace_log
        except Exception as e:
            trace_log += f"Error with LLM: {str(e)}\n"
            return f"Error: {str(e)}", trace_log

    def __call__(self, question: str) -> Tuple[str, str]:
        """Huvudfunktion som bearbetar frågan"""
        total_trace = f"Processing question: {question}\n"

        tool_type, needs_tool = self.needs_tool(question)
        total_trace += f"Tool needed: {tool_type}\n"

        context = ""
        if needs_tool and tool_type != 'llm':
            tool_result, tool_trace = self.process_with_tools(question, tool_type)
            total_trace += tool_trace
            context = tool_result

        llm_response, llm_trace = self.reason_with_llm(question, context)
        total_trace += llm_trace

        final_answer = self.extract_final_answer(llm_response)
        total_trace += f"Final answer extracted: {final_answer}\n"

        return final_answer, total_trace