crewAI/src/crewai/llm.py

import logging
import os
import re
import sys
import threading
import warnings
from contextlib import contextmanager
from typing import Any, Dict, List, Optional, Union

import litellm
from litellm import get_supported_openai_params

from crewai.utilities.exceptions.context_window_exceeding_exception import (
    LLMContextLengthExceededException,
)


class FilteredStream:
    def __init__(self, original_stream):
        self._original_stream = original_stream
        self._lock = threading.Lock()

    def write(self, s) -> int:
        with self._lock:
            if (
                "Give Feedback / Get Help: https://github.com/BerriAI/litellm/issues/new"
                in s
                or "LiteLLM.Info: If you need to debug this error, use `litellm.set_verbose=True`"
                in s
            ):
                return 0
            return self._original_stream.write(s)

    def flush(self):
        with self._lock:
            return self._original_stream.flush()


LLM_CONTEXT_WINDOW_SIZES = {
    # openai
    "gpt-4": 8192,
    "gpt-4o": 128000,
    "gpt-4o-mini": 128000,
    "gpt-4-turbo": 128000,
    "o1-preview": 128000,
    "o1-mini": 128000,
    # gemini
    "gemini-2.0-flash": 1048576,
    "gemini-1.5-pro": 2097152,
    "gemini-1.5-flash": 1048576,
    "gemini-1.5-flash-8b": 1048576,
    # deepseek
    "deepseek-chat": 128000,
    # groq
    "gemma2-9b-it": 8192,
    "gemma-7b-it": 8192,
    "llama3-groq-70b-8192-tool-use-preview": 8192,
    "llama3-groq-8b-8192-tool-use-preview": 8192,
    "llama-3.1-70b-versatile": 131072,
    "llama-3.1-8b-instant": 131072,
    "llama-3.2-1b-preview": 8192,
    "llama-3.2-3b-preview": 8192,
    "llama-3.2-11b-text-preview": 8192,
    "llama-3.2-90b-text-preview": 8192,
    "llama3-70b-8192": 8192,
    "llama3-8b-8192": 8192,
    "mixtral-8x7b-32768": 32768,
    "llama-3.3-70b-versatile": 128000,
    "llama-3.3-70b-instruct": 128000,
}

DEFAULT_CONTEXT_WINDOW_SIZE = 8192
CONTEXT_WINDOW_USAGE_RATIO = 0.75


@contextmanager
def suppress_warnings():
    with warnings.catch_warnings():
        warnings.filterwarnings("ignore")

        # Redirect stdout and stderr
        old_stdout = sys.stdout
        old_stderr = sys.stderr
        sys.stdout = FilteredStream(old_stdout)
        sys.stderr = FilteredStream(old_stderr)

        try:
            yield
        finally:
            # Restore stdout and stderr
            sys.stdout = old_stdout
            sys.stderr = old_stderr


class LLM:
    def __init__(
        self,
        model: str,
        timeout: Optional[Union[float, int]] = None,
        temperature: Optional[float] = None,
        top_p: Optional[float] = None,
        n: Optional[int] = None,
        stop: Optional[Union[str, List[str]]] = None,
        max_completion_tokens: Optional[int] = None,
        max_tokens: Optional[int] = None,
        presence_penalty: Optional[float] = None,
        frequency_penalty: Optional[float] = None,
        logit_bias: Optional[Dict[int, float]] = None,
        response_format: Optional[Dict[str, Any]] = None,
        seed: Optional[int] = None,
        logprobs: Optional[bool] = None,
        top_logprobs: Optional[int] = None,
        base_url: Optional[str] = None,
        api_version: Optional[str] = None,
        api_key: Optional[str] = None,
        callbacks: List[Any] = [],
        **kwargs,
    ):
        self.model = model
        self.timeout = timeout
        self.temperature = temperature
        self.top_p = top_p
        self.n = n
        self.stop = stop
        self.max_completion_tokens = max_completion_tokens
        self.max_tokens = max_tokens
        self.presence_penalty = presence_penalty
        self.frequency_penalty = frequency_penalty
        self.logit_bias = logit_bias
        self.response_format = response_format
        self.seed = seed
        self.logprobs = logprobs
        self.top_logprobs = top_logprobs
        self.base_url = base_url
        self.api_version = api_version
        self.api_key = api_key
        self.callbacks = callbacks
        self.context_window_size = 0
        self.kwargs = kwargs

        litellm.drop_params = True
        litellm.set_verbose = False
        self.set_callbacks(callbacks)
        self.set_env_callbacks()

    def call(self, messages: List[Dict[str, str]], callbacks: List[Any] = []) -> str:
        with suppress_warnings():
            if callbacks and len(callbacks) > 0:
                self.set_callbacks(callbacks)

            try:
                params = {
                    "model": self.model,
                    "messages": messages,
                    "timeout": self.timeout,
                    "temperature": self.temperature,
                    "top_p": self.top_p,
                    "n": self.n,
                    "stop": self.stop,
                    "max_tokens": self.max_tokens or self.max_completion_tokens,
                    "presence_penalty": self.presence_penalty,
                    "frequency_penalty": self.frequency_penalty,
                    "logit_bias": self.logit_bias,
                    "response_format": self.response_format,
                    "seed": self.seed,
                    "logprobs": self.logprobs,
                    "top_logprobs": self.top_logprobs,
                    "api_base": self.base_url,
                    "api_version": self.api_version,
                    "api_key": self.api_key,
                    "stream": False,
                    **self.kwargs,
                }

                # Remove None values to avoid passing unnecessary parameters
                params = {k: v for k, v in params.items() if v is not None}

                # Claude 4.6+ models reject the temperature parameter
                if self._is_anthropic_no_prefill_model():
                    params.pop("temperature", None)

                response = litellm.completion(**params)
                return response["choices"][0]["message"]["content"]
            except Exception as e:
                if not LLMContextLengthExceededException(
                    str(e)
                )._is_context_limit_error(str(e)):
                    logging.error(f"LiteLLM call failed: {str(e)}")

                raise  # Re-raise the exception after logging

    def supports_assistant_prefill(self) -> bool:
        """Check if the model supports assistant message prefill.

        Some Anthropic models (Claude 4.6+) reject requests where the
        last message has the assistant role.  Returns True for models
        that support prefill or where the capability cannot be determined.
        """
        try:
            info = litellm.get_model_info(self.model)
            provider = info.get("litellm_provider", "")
            prefill = info.get("supports_assistant_prefill")
            if "anthropic" in provider and prefill is False:
                return False
        except Exception:
            pass

        # Fallback heuristic for model names not in the litellm registry
        model_lower = self.model.lower()
        if "claude" in model_lower:
            match = re.search(r"claude.*?(\d+)[.-](\d+)", model_lower)
            if match:
                major, minor = int(match.group(1)), int(match.group(2))
                if (major == 4 and minor >= 6) or major >= 5:
                    return False
        return True

    def _is_anthropic_no_prefill_model(self) -> bool:
        """Return True when the model is an Anthropic model that does not
        support assistant prefill.  Used to also drop parameters that these
        models reject (e.g. temperature)."""
        return not self.supports_assistant_prefill()

    def supports_function_calling(self) -> bool:
        try:
            params = get_supported_openai_params(model=self.model)
            return "response_format" in params
        except Exception as e:
            logging.error(f"Failed to get supported params: {str(e)}")
            return False

    def supports_stop_words(self) -> bool:
        try:
            params = get_supported_openai_params(model=self.model)
            return "stop" in params
        except Exception as e:
            logging.error(f"Failed to get supported params: {str(e)}")
            return False

    def get_context_window_size(self) -> int:
        # Only using 75% of the context window size to avoid cutting the message in the middle
        if self.context_window_size != 0:
            return self.context_window_size

        self.context_window_size = int(
            DEFAULT_CONTEXT_WINDOW_SIZE * CONTEXT_WINDOW_USAGE_RATIO
        )
        for key, value in LLM_CONTEXT_WINDOW_SIZES.items():
            if self.model.startswith(key):
                self.context_window_size = int(value * CONTEXT_WINDOW_USAGE_RATIO)
        return self.context_window_size

    def set_callbacks(self, callbacks: List[Any]):
        callback_types = [type(callback) for callback in callbacks]
        for callback in litellm.success_callback[:]:
            if type(callback) in callback_types:
                litellm.success_callback.remove(callback)

        for callback in litellm._async_success_callback[:]:
            if type(callback) in callback_types:
                litellm._async_success_callback.remove(callback)

        litellm.callbacks = callbacks

    def set_env_callbacks(self):
        """
        Sets the success and failure callbacks for the LiteLLM library from environment variables.

        This method reads the `LITELLM_SUCCESS_CALLBACKS` and `LITELLM_FAILURE_CALLBACKS`
        environment variables, which should contain comma-separated lists of callback names.
        It then assigns these lists to `litellm.success_callback` and `litellm.failure_callback`,
        respectively.

        If the environment variables are not set or are empty, the corresponding callback lists
        will be set to empty lists.

        Example:
            LITELLM_SUCCESS_CALLBACKS="langfuse,langsmith"
            LITELLM_FAILURE_CALLBACKS="langfuse"

        This will set `litellm.success_callback` to ["langfuse", "langsmith"] and
        `litellm.failure_callback` to ["langfuse"].
        """
        success_callbacks_str = os.environ.get("LITELLM_SUCCESS_CALLBACKS", "")
        success_callbacks = []
        if success_callbacks_str:
            success_callbacks = [
                callback.strip() for callback in success_callbacks_str.split(",")
            ]

        failure_callbacks_str = os.environ.get("LITELLM_FAILURE_CALLBACKS", "")
        failure_callbacks = []
        if failure_callbacks_str:
            failure_callbacks = [
                callback.strip() for callback in failure_callbacks_str.split(",")
            ]

        litellm.success_callback = success_callbacks
        litellm.failure_callback = failure_callbacks