deer-flow/src/llms/llm.py

# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
# SPDX-License-Identifier: MIT

import logging
import os
from pathlib import Path
from typing import Any, Dict, get_args

import httpx
from langchain_core.language_models import BaseChatModel
from langchain_deepseek import ChatDeepSeek
from langchain_google_genai import ChatGoogleGenerativeAI
from langchain_openai import AzureChatOpenAI, ChatOpenAI

from src.config import load_yaml_config
from src.config.agents import LLMType
from src.llms.providers.dashscope import ChatDashscope

logger = logging.getLogger(__name__)

# Cache for LLM instances
_llm_cache: dict[LLMType, BaseChatModel] = {}

# Allowed LLM configuration keys to prevent unexpected parameters from being passed
# to LLM constructors (Issue #411 - SEARCH_ENGINE warning fix)
ALLOWED_LLM_CONFIG_KEYS = {
    # Common LLM configuration keys
    "model",
    "api_key",
    "base_url",
    "api_base",
    "max_retries",
    "timeout",
    "max_tokens",
    "temperature",
    "top_p",
    "frequency_penalty",
    "presence_penalty",
    "stop",
    "n",
    "stream",
    "logprobs",
    "echo",
    "best_of",
    "logit_bias",
    "user",
    "seed",
    # SSL and HTTP client settings
    "verify_ssl",
    "http_client",
    "http_async_client",
    # Platform-specific keys
    "platform",
    "google_api_key",
    # Azure-specific keys
    "azure_endpoint",
    "azure_deployment",
    "api_version",
    "azure_ad_token",
    "azure_ad_token_provider",
    # Dashscope/Doubao specific keys
    "extra_body",
    # Token limit for context compression (removed before passing to LLM)
    "token_limit",
    # Default headers
    "default_headers",
    "default_query",
}


def _get_config_file_path() -> str:
    """Get the path to the configuration file."""
    return str((Path(__file__).parent.parent.parent / "conf.yaml").resolve())


def _get_llm_type_config_keys() -> dict[str, str]:
    """Get mapping of LLM types to their configuration keys."""
    return {
        "reasoning": "REASONING_MODEL",
        "basic": "BASIC_MODEL",
        "vision": "VISION_MODEL",
        "code": "CODE_MODEL",
    }


def _get_env_llm_conf(llm_type: str) -> Dict[str, Any]:
    """
    Get LLM configuration from environment variables.
    Environment variables should follow the format: {LLM_TYPE}__{KEY}
    e.g., BASIC_MODEL__api_key, BASIC_MODEL__base_url
    """
    prefix = f"{llm_type.upper()}_MODEL__"
    conf = {}
    for key, value in os.environ.items():
        if key.startswith(prefix):
            conf_key = key[len(prefix) :].lower()
            conf[conf_key] = value
    return conf


def _create_llm_use_conf(llm_type: LLMType, conf: Dict[str, Any]) -> BaseChatModel:
    """Create LLM instance using configuration."""
    llm_type_config_keys = _get_llm_type_config_keys()
    config_key = llm_type_config_keys.get(llm_type)

    if not config_key:
        raise ValueError(f"Unknown LLM type: {llm_type}")

    llm_conf = conf.get(config_key, {})
    if not isinstance(llm_conf, dict):
        raise ValueError(f"Invalid LLM configuration for {llm_type}: {llm_conf}")

    # Get configuration from environment variables
    env_conf = _get_env_llm_conf(llm_type)

    # Merge configurations, with environment variables taking precedence
    merged_conf = {**llm_conf, **env_conf}

    # Filter out unexpected parameters to prevent LangChain warnings (Issue #411)
    # This prevents configuration keys like SEARCH_ENGINE from being passed to LLM constructors
    allowed_keys_lower = {k.lower() for k in ALLOWED_LLM_CONFIG_KEYS}
    unexpected_keys = [key for key in merged_conf.keys() if key.lower() not in allowed_keys_lower]
    for key in unexpected_keys:
        removed_value = merged_conf.pop(key)
        logger.warning(
            f"Removed unexpected LLM configuration key '{key}'. "
            f"This key is not a valid LLM parameter and may have been placed in the wrong section of conf.yaml. "
            f"Valid LLM config keys include: model, api_key, base_url, max_retries, temperature, etc."
        )

    # Remove unnecessary parameters when initializing the client
    if "token_limit" in merged_conf:
        merged_conf.pop("token_limit")

    if not merged_conf:
        raise ValueError(f"No configuration found for LLM type: {llm_type}")

    # Add max_retries to handle rate limit errors
    if "max_retries" not in merged_conf:
        merged_conf["max_retries"] = 3

    # Handle SSL verification settings
    verify_ssl = merged_conf.pop("verify_ssl", True)

    # Create custom HTTP client if SSL verification is disabled
    if not verify_ssl:
        http_client = httpx.Client(verify=False)
        http_async_client = httpx.AsyncClient(verify=False)
        merged_conf["http_client"] = http_client
        merged_conf["http_async_client"] = http_async_client

    # Check if it's Google AI Studio platform based on configuration
    platform = merged_conf.get("platform", "").lower()
    is_google_aistudio = platform == "google_aistudio" or platform == "google-aistudio"

    if is_google_aistudio:
        # Handle Google AI Studio specific configuration
        gemini_conf = merged_conf.copy()

        # Map common keys to Google AI Studio specific keys
        if "api_key" in gemini_conf:
            gemini_conf["google_api_key"] = gemini_conf.pop("api_key")

        # Remove base_url and platform since Google AI Studio doesn't use them
        gemini_conf.pop("base_url", None)
        gemini_conf.pop("platform", None)

        # Remove unsupported parameters for Google AI Studio
        gemini_conf.pop("http_client", None)
        gemini_conf.pop("http_async_client", None)

        return ChatGoogleGenerativeAI(**gemini_conf)

    if "azure_endpoint" in merged_conf or os.getenv("AZURE_OPENAI_ENDPOINT"):
        return AzureChatOpenAI(**merged_conf)

    # Check if base_url is dashscope endpoint
    if "base_url" in merged_conf and "dashscope." in merged_conf["base_url"]:
        if llm_type == "reasoning":
            merged_conf["extra_body"] = {"enable_thinking": True}
        else:
            merged_conf["extra_body"] = {"enable_thinking": False}
        return ChatDashscope(**merged_conf)

    if llm_type == "reasoning":
        merged_conf["api_base"] = merged_conf.pop("base_url", None)
        return ChatDeepSeek(**merged_conf)
    else:
        return ChatOpenAI(**merged_conf)


def get_llm_by_type(llm_type: LLMType) -> BaseChatModel:
    """
    Get LLM instance by type. Returns cached instance if available.
    """
    if llm_type in _llm_cache:
        return _llm_cache[llm_type]

    conf = load_yaml_config(_get_config_file_path())
    llm = _create_llm_use_conf(llm_type, conf)
    _llm_cache[llm_type] = llm
    return llm


def get_configured_llm_models() -> dict[str, list[str]]:
    """
    Get all configured LLM models grouped by type.

    Returns:
        Dictionary mapping LLM type to list of configured model names.
    """
    try:
        conf = load_yaml_config(_get_config_file_path())
        llm_type_config_keys = _get_llm_type_config_keys()

        configured_models: dict[str, list[str]] = {}

        for llm_type in get_args(LLMType):
            # Get configuration from YAML file
            config_key = llm_type_config_keys.get(llm_type, "")
            yaml_conf = conf.get(config_key, {}) if config_key else {}

            # Get configuration from environment variables
            env_conf = _get_env_llm_conf(llm_type)

            # Merge configurations, with environment variables taking precedence
            merged_conf = {**yaml_conf, **env_conf}

            # Check if model is configured
            model_name = merged_conf.get("model")
            if model_name:
                configured_models.setdefault(llm_type, []).append(model_name)

        return configured_models

    except Exception as e:
        # Log error and return empty dict to avoid breaking the application
        print(f"Warning: Failed to load LLM configuration: {e}")
        return {}


def _get_model_token_limit_defaults() -> dict[str, int]:
    """
    Get default token limits for common LLM models.
    These are conservative limits to prevent token overflow errors (Issue #721).
    Users can override by setting token_limit in their config.
    """
    return {
        # OpenAI models
        "gpt-4o": 120000,
        "gpt-4-turbo": 120000,
        "gpt-4": 8000,
        "gpt-3.5-turbo": 4000,
        # Anthropic Claude
        "claude-3": 180000,
        "claude-2": 100000,
        # Google Gemini
        "gemini-2": 180000,
        "gemini-1.5-pro": 180000,
        "gemini-1.5-flash": 180000,
        "gemini-pro": 30000,
        # Bytedance Doubao
        "doubao": 200000,
        # DeepSeek
        "deepseek": 100000,
        # Ollama/local
        "qwen": 30000,
        "llama": 4000,
        # Default fallback for unknown models
        "default": 100000,
    }


def _infer_token_limit_from_model(model_name: str) -> int:
    """
    Infer a reasonable token limit from the model name.
    This helps protect against token overflow errors when token_limit is not explicitly configured.
    
    Args:
        model_name: The model name from configuration
        
    Returns:
        A conservative token limit based on known model capabilities
    """
    if not model_name:
        return 100000  # Safe default
    
    model_name_lower = model_name.lower()
    defaults = _get_model_token_limit_defaults()
    
    # Try exact or prefix matches
    for key, limit in defaults.items():
        if key in model_name_lower:
            return limit
    
    # Return safe default if no match found
    return defaults["default"]


def get_llm_token_limit_by_type(llm_type: str) -> int:
    """
    Get the maximum token limit for a given LLM type.
    
    Priority order:
    1. Explicitly configured token_limit in conf.yaml
    2. Inferred from model name based on known model capabilities
    3. Safe default (100,000 tokens)
    
    This helps prevent token overflow errors (Issue #721) even when token_limit is not configured.

    Args:
        llm_type (str): The type of LLM (e.g., 'basic', 'reasoning', 'vision', 'code').

    Returns:
        int: The maximum token limit for the specified LLM type (conservative estimate).
    """
    llm_type_config_keys = _get_llm_type_config_keys()
    config_key = llm_type_config_keys.get(llm_type)

    conf = load_yaml_config(_get_config_file_path())
    model_config = conf.get(config_key, {})
    
    # First priority: explicitly configured token_limit
    if "token_limit" in model_config:
        configured_limit = model_config["token_limit"]
        if configured_limit is not None:
            return configured_limit
    
    # Second priority: infer from model name
    model_name = model_config.get("model")
    if model_name:
        inferred_limit = _infer_token_limit_from_model(model_name)
        return inferred_limit
    
    # Fallback: safe default
    return _get_model_token_limit_defaults()["default"]


# In the future, we will use reasoning_llm and vl_llm for different purposes
# reasoning_llm = get_llm_by_type("reasoning")
# vl_llm = get_llm_by_type("vision")