# Configuration and constants for GAIA Agent Evaluator
import os

# --- API Configuration ---
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
LLAMA_API_URL = "https://api-inference.huggingface.co/models/mistralai/Mixtral-8x7B-Instruct-v0.1"
HF_API_TOKEN = os.getenv("HF_API_TOKEN")
HEADERS = {"Authorization": f"Bearer {HF_API_TOKEN}"} if HF_API_TOKEN else {}

# --- Model Configuration ---
USE_LOCAL_MODEL = True  # Set to False to use remote API model instead
USE_LLAMACPP = True  # Set to True to use llama-cpp-python instead of transformers

# Configuration for llama-cpp-python model
# LLAMACPP_CONFIG = {
#     "model_path": None,  # Will use a default small model if None
#     # Using a smaller GGUF model to avoid download issues
#     "model_url": "https://huggingface.co/eachadea/ggml-gridlocked-alpha-3b/resolve/main/ggml-gridlocked-3b-q4_0.bin",
#     "n_ctx": 2048,
#     "n_gpu_layers": 0,  # Use 0 for CPU-only
#     "max_tokens": 1024,
#     "temperature": 0.7
# }

LLAMACPP_CONFIG = {
    "model_path": None,  # Use a default or provide a local path manually
    "model_url": "https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf",
    "n_ctx": 2048,
    "n_gpu_layers": 0,  # CPU-only, adjust as needed for GPU
    "max_tokens": 1024,
    "temperature": 0.7
}

# Backup configuration for transformers model
LOCAL_MODEL_CONFIG = {
    "model_name": "TinyLlama/TinyLlama-1.1B-Chat-v0.6",
    "device": "cpu",
    "max_tokens": 1024,
    "temperature": 0.5
}

# --- Request Configuration ---
MAX_RETRIES = 3
RETRY_DELAY = 2  # seconds

# --- Knowledge Base Configuration ---
KNOWLEDGE_BASE_PATH = os.path.join(os.path.dirname(__file__), 'data', 'knowledge_base.txt')

def load_knowledge_base():
    """Load knowledge base from file"""
    try:
        with open(KNOWLEDGE_BASE_PATH, 'r', encoding='utf-8') as f:
            return f.read()
    except FileNotFoundError:
        os.makedirs(os.path.dirname(KNOWLEDGE_BASE_PATH), exist_ok=True)
        return ""

# --- Tool Pattern Matching ---
YOUTUBE_PATTERNS = ["youtube.com", "youtu.be", "video", "watch?v=", "channel"]

REVERSE_TEXT_PATTERNS = ["rewsna", "ecnetnes", "sdrow"]

WIKIPEDIA_PATTERNS = [
    "wikipedia", "article", "published", "featured article", 
    "promoted", "nominated", "discography", "studio albums",
    "encyclopedia", "wiki", "featured content"
]

WEB_SEARCH_PATTERNS = [
    # Time indicators
    "current", "latest", "recent", "2023", "2022", "2021", "2020", "today",
    # Question words
    "how many", "where", "when", "who", "which", "what", "whose",
    # Sports and competitions
    "yankee", "walks", "athletes", "olympics", "competition", "pitcher", "baseball",
    # Specific entities that need web lookup
    "malko", "taishō tamai", "universe today", "nedoshivina",
    "specimens", "polish-language", "actor", "played",
    # Geographic and demographic
    "country", "nationality", "first name", "award number", "city",
    # Publications and research
    "published", "paper", "study", "research", "journal", "author",
    # Statistics and data
    "statistics", "data", "facts", "information about", "number of"
]

AI_PATTERNS = ["agent", "gaia", "llm", "ai", "evaluation", "tool", "artificial intelligence"]

FILE_PATTERNS = ["excel", "xlsx", "csv", "python", ".py", "mp3", "wav", "audio", "voice memo"]

# --- Answer Cleaning Patterns ---
ANSWER_PREFIXES_TO_REMOVE = [
    "final answer:", "answer:", "the answer is:", "result:", 
    "solution:", "conclusion:", "final answer is:", "direct answer:",
    "based on the context:", "according to:", "the result is:"
]

LLM_RESPONSE_MARKERS = ["<answer>", "<response>", "Answer:", "Response:", "Assistant:"]
LLM_END_MARKERS = ["</answer>", "</response>", "Human:", "User:"]

# Ensure knowledge base is loaded correctly
GAIA_KNOWLEDGE = load_knowledge_base()