suna/backend/services/llm.py

"""
LLM API interface for making calls to various language models.

This module provides a unified interface for making API calls to different LLM providers
(OpenAI, Anthropic, Groq, xAI, etc.) using LiteLLM. It includes support for:
- Streaming responses
- Tool calls and function calling
- Retry logic with exponential backoff
- Model-specific configurations
- Comprehensive error handling and logging
"""

from typing import Union, Dict, Any, Optional, AsyncGenerator, List
import os
import json
import asyncio
from openai import OpenAIError
import litellm
from litellm.files.main import ModelResponse
from utils.logger import logger
from utils.config import config

# litellm.set_verbose=True
litellm.modify_params=True

# Constants
MAX_RETRIES = 2
RATE_LIMIT_DELAY = 30
RETRY_DELAY = 0.1

class LLMError(Exception):
    """Base exception for LLM-related errors."""
    pass

class LLMRetryError(LLMError):
    """Exception raised when retries are exhausted."""
    pass

def setup_api_keys() -> None:
    """Set up API keys from environment variables."""
    providers = ['OPENAI', 'ANTHROPIC', 'GROQ', 'OPENROUTER', 'XAI']
    for provider in providers:
        key = getattr(config, f'{provider}_API_KEY')
        if key:
            logger.debug(f"API key set for provider: {provider}")
        else:
            logger.warning(f"No API key found for provider: {provider}")

    # Set up OpenRouter API base if not already set
    if config.OPENROUTER_API_KEY and config.OPENROUTER_API_BASE:
        os.environ['OPENROUTER_API_BASE'] = config.OPENROUTER_API_BASE
        logger.debug(f"Set OPENROUTER_API_BASE to {config.OPENROUTER_API_BASE}")

    # Set up AWS Bedrock credentials
    aws_access_key = config.AWS_ACCESS_KEY_ID
    aws_secret_key = config.AWS_SECRET_ACCESS_KEY
    aws_region = config.AWS_REGION_NAME

    if aws_access_key and aws_secret_key and aws_region:
        logger.debug(f"AWS credentials set for Bedrock in region: {aws_region}")
        # Configure LiteLLM to use AWS credentials
        os.environ['AWS_ACCESS_KEY_ID'] = aws_access_key
        os.environ['AWS_SECRET_ACCESS_KEY'] = aws_secret_key
        os.environ['AWS_REGION_NAME'] = aws_region
    else:
        logger.warning(f"Missing AWS credentials for Bedrock integration - access_key: {bool(aws_access_key)}, secret_key: {bool(aws_secret_key)}, region: {aws_region}")

def get_openrouter_fallback(model_name: str) -> Optional[str]:
    """Get OpenRouter fallback model for a given model name."""
    # Skip if already using OpenRouter
    if model_name.startswith("openrouter/"):
        return None
    
    # Map models to their OpenRouter equivalents
    fallback_mapping = {
        "anthropic/claude-3-7-sonnet-latest": "openrouter/anthropic/claude-3.7-sonnet",
        "anthropic/claude-sonnet-4-20250514": "openrouter/anthropic/claude-sonnet-4",
        "xai/grok-4": "openrouter/x-ai/grok-4",
    }
    
    # Check for exact match first
    if model_name in fallback_mapping:
        return fallback_mapping[model_name]
    
    # Check for partial matches (e.g., bedrock models)
    for key, value in fallback_mapping.items():
        if key in model_name:
            return value
    
    # Default fallbacks by provider
    if "claude" in model_name.lower() or "anthropic" in model_name.lower():
        return "openrouter/anthropic/claude-sonnet-4"
    elif "xai" in model_name.lower() or "grok" in model_name.lower():
        return "openrouter/x-ai/grok-4"
    
    return None

async def handle_error(error: Exception, attempt: int, max_attempts: int) -> None:
    """Handle API errors with appropriate delays and logging."""
    delay = RATE_LIMIT_DELAY if isinstance(error, litellm.exceptions.RateLimitError) else RETRY_DELAY
    logger.warning(f"Error on attempt {attempt + 1}/{max_attempts}: {str(error)}")
    logger.debug(f"Waiting {delay} seconds before retry...")
    await asyncio.sleep(delay)

def prepare_params(
    messages: List[Dict[str, Any]],
    model_name: str,
    temperature: float = 0,
    max_tokens: Optional[int] = None,
    response_format: Optional[Any] = None,
    tools: Optional[List[Dict[str, Any]]] = None,
    tool_choice: str = "auto",
    api_key: Optional[str] = None,
    api_base: Optional[str] = None,
    stream: bool = False,
    top_p: Optional[float] = None,
    model_id: Optional[str] = None,
    enable_thinking: Optional[bool] = False,
    reasoning_effort: Optional[str] = 'low'
) -> Dict[str, Any]:
    """Prepare parameters for the API call."""
    params = {
        "model": model_name,
        "messages": messages,
        "temperature": temperature,
        "response_format": response_format,
        "top_p": top_p,
        "stream": stream,
    }

    if api_key:
        params["api_key"] = api_key
    if api_base:
        params["api_base"] = api_base
    if model_id:
        params["model_id"] = model_id

    # Handle token limits
    if max_tokens is not None:
        # For Claude 3.7 in Bedrock, do not set max_tokens or max_tokens_to_sample
        # as it causes errors with inference profiles
        if model_name.startswith("bedrock/") and "claude-3-7" in model_name:
            logger.debug(f"Skipping max_tokens for Claude 3.7 model: {model_name}")
            # Do not add any max_tokens parameter for Claude 3.7
        else:
            param_name = "max_completion_tokens" if 'o1' in model_name else "max_tokens"
            params[param_name] = max_tokens

    # Add tools if provided
    if tools:
        params.update({
            "tools": tools,
            "tool_choice": tool_choice
        })
        logger.debug(f"Added {len(tools)} tools to API parameters")

    # # Add Claude-specific headers
    if "claude" in model_name.lower() or "anthropic" in model_name.lower():
        params["extra_headers"] = {
            # "anthropic-beta": "max-tokens-3-5-sonnet-2024-07-15"
            "anthropic-beta": "output-128k-2025-02-19"
        }
        # params["mock_testing_fallback"] = True
        logger.debug("Added Claude-specific headers")

    # Add OpenRouter-specific parameters
    if model_name.startswith("openrouter/"):
        logger.debug(f"Preparing OpenRouter parameters for model: {model_name}")

        # Add optional site URL and app name from config
        site_url = config.OR_SITE_URL
        app_name = config.OR_APP_NAME
        if site_url or app_name:
            extra_headers = params.get("extra_headers", {})
            if site_url:
                extra_headers["HTTP-Referer"] = site_url
            if app_name:
                extra_headers["X-Title"] = app_name
            params["extra_headers"] = extra_headers
            logger.debug(f"Added OpenRouter site URL and app name to headers")

    # Add Bedrock-specific parameters
    if model_name.startswith("bedrock/"):
        logger.debug(f"Preparing AWS Bedrock parameters for model: {model_name}")

        if not model_id and "anthropic.claude-3-7-sonnet" in model_name:
            params["model_id"] = "arn:aws:bedrock:us-west-2:935064898258:inference-profile/us.anthropic.claude-3-7-sonnet-20250219-v1:0"
            logger.debug(f"Auto-set model_id for Claude 3.7 Sonnet: {params['model_id']}")

    fallback_model = get_openrouter_fallback(model_name)
    if fallback_model:
        params["fallbacks"] = [{
            "model": fallback_model,
            "messages": messages,
        }]
        logger.debug(f"Added OpenRouter fallback for model: {model_name} to {fallback_model}")

    # Apply Anthropic prompt caching (minimal implementation)
    # Check model name *after* potential modifications (like adding bedrock/ prefix)
    effective_model_name = params.get("model", model_name) # Use model from params if set, else original
    if "claude" in effective_model_name.lower() or "anthropic" in effective_model_name.lower():
        messages = params["messages"] # Direct reference, modification affects params

        # Ensure messages is a list
        if not isinstance(messages, list):
            return params # Return early if messages format is unexpected

        # Apply cache control to the first 4 text blocks across all messages
        cache_control_count = 0
        max_cache_control_blocks = 3

        for message in messages:
            if cache_control_count >= max_cache_control_blocks:
                break
                
            content = message.get("content")
            
            if isinstance(content, str):
                message["content"] = [
                    {"type": "text", "text": content, "cache_control": {"type": "ephemeral"}}
                ]
                cache_control_count += 1
            elif isinstance(content, list):
                for item in content:
                    if cache_control_count >= max_cache_control_blocks:
                        break
                    if isinstance(item, dict) and item.get("type") == "text" and "cache_control" not in item:
                        item["cache_control"] = {"type": "ephemeral"}
                        cache_control_count += 1

    # Add reasoning_effort for Anthropic models if enabled
    use_thinking = enable_thinking if enable_thinking is not None else False
    is_anthropic = "anthropic" in effective_model_name.lower() or "claude" in effective_model_name.lower()
    is_xai = "xai" in effective_model_name.lower() or model_name.startswith("xai/")
    is_kimi_k2 = "kimi-k2" in effective_model_name.lower() or model_name.startswith("moonshotai/kimi-k2")

    if is_kimi_k2:
        params["provider"] = {
            "order": ["groq", "together/fp8"]
        }

    if is_anthropic and use_thinking:
        effort_level = reasoning_effort if reasoning_effort else 'low'
        params["reasoning_effort"] = effort_level
        params["temperature"] = 1.0 # Required by Anthropic when reasoning_effort is used
        logger.info(f"Anthropic thinking enabled with reasoning_effort='{effort_level}'")

    # Add reasoning_effort for xAI models if enabled
    if is_xai and use_thinking:
        effort_level = reasoning_effort if reasoning_effort else 'low'
        params["reasoning_effort"] = effort_level
        logger.info(f"xAI thinking enabled with reasoning_effort='{effort_level}'")

    # Add xAI-specific parameters
    if model_name.startswith("xai/"):
        logger.debug(f"Preparing xAI parameters for model: {model_name}")
        # xAI models support standard parameters, no special handling needed beyond reasoning_effort

    return params

async def make_llm_api_call(
    messages: List[Dict[str, Any]],
    model_name: str,
    response_format: Optional[Any] = None,
    temperature: float = 0,
    max_tokens: Optional[int] = None,
    tools: Optional[List[Dict[str, Any]]] = None,
    tool_choice: str = "auto",
    api_key: Optional[str] = None,
    api_base: Optional[str] = None,
    stream: bool = False,
    top_p: Optional[float] = None,
    model_id: Optional[str] = None,
    enable_thinking: Optional[bool] = False,
    reasoning_effort: Optional[str] = 'low'
) -> Union[Dict[str, Any], AsyncGenerator, ModelResponse]:
    """
    Make an API call to a language model using LiteLLM.

    Args:
        messages: List of message dictionaries for the conversation
        model_name: Name of the model to use (e.g., "gpt-4", "claude-3", "openrouter/openai/gpt-4", "bedrock/anthropic.claude-3-sonnet-20240229-v1:0")
        response_format: Desired format for the response
        temperature: Sampling temperature (0-1)
        max_tokens: Maximum tokens in the response
        tools: List of tool definitions for function calling
        tool_choice: How to select tools ("auto" or "none")
        api_key: Override default API key
        api_base: Override default API base URL
        stream: Whether to stream the response
        top_p: Top-p sampling parameter
        model_id: Optional ARN for Bedrock inference profiles
        enable_thinking: Whether to enable thinking
        reasoning_effort: Level of reasoning effort

    Returns:
        Union[Dict[str, Any], AsyncGenerator]: API response or stream

    Raises:
        LLMRetryError: If API call fails after retries
        LLMError: For other API-related errors
    """
    # debug <timestamp>.json messages
    logger.info(f"Making LLM API call to model: {model_name} (Thinking: {enable_thinking}, Effort: {reasoning_effort})")
    logger.info(f"📡 API Call: Using model {model_name}")
    params = prepare_params(
        messages=messages,
        model_name=model_name,
        temperature=temperature,
        max_tokens=max_tokens,
        response_format=response_format,
        tools=tools,
        tool_choice=tool_choice,
        api_key=api_key,
        api_base=api_base,
        stream=stream,
        top_p=top_p,
        model_id=model_id,
        enable_thinking=enable_thinking,
        reasoning_effort=reasoning_effort
    )
    last_error = None
    for attempt in range(MAX_RETRIES):
        try:
            logger.debug(f"Attempt {attempt + 1}/{MAX_RETRIES}")
            # logger.debug(f"API request parameters: {json.dumps(params, indent=2)}")

            response = await litellm.acompletion(**params)
            logger.debug(f"Successfully received API response from {model_name}")
            # logger.debug(f"Response: {response}")
            return response

        except (litellm.exceptions.RateLimitError, OpenAIError, json.JSONDecodeError) as e:
            last_error = e
            await handle_error(e, attempt, MAX_RETRIES)

        except Exception as e:
            logger.error(f"Unexpected error during API call: {str(e)}", exc_info=True)
            raise LLMError(f"API call failed: {str(e)}")

    error_msg = f"Failed to make API call after {MAX_RETRIES} attempts"
    if last_error:
        error_msg += f". Last error: {str(last_error)}"
    logger.error(error_msg, exc_info=True)
    raise LLMRetryError(error_msg)

# Initialize API keys on module import
setup_api_keys()