Implementation:BerriAI Litellm LLM Request Utils

Attribute	Value
Sources	litellm/litellm_core_utils/llm_request_utils.py
Domains	Request Processing, Model Selection, Safety
Last Updated	2026-02-15 16:00 GMT

Overview

Provides utilities for sanitizing LLM request parameters, selecting cheapest models by provider, and extracting proxy request headers.

Description

This module contains three utility functions for LLM request handling:

• _ensure_extra_body_is_safe -- Ensures that the extra_body dictionary attached to a request is JSON-serializable. Specifically handles Langfuse TextPromptClient objects in extra_body["metadata"]["prompt"] by converting them to plain dictionaries via __dict__.
• pick_cheapest_chat_models_from_llm_provider -- Queries the global model cost map to find the n cheapest chat-mode models for a given provider. Models are sorted by combined input + output cost per token. Returns an empty list if the provider is not found.
• get_proxy_server_request_headers -- Extracts the original HTTP request headers from litellm_params["proxy_server_request"]["headers"], useful for custom logging or header forwarding in proxy deployments.

Usage

Import _ensure_extra_body_is_safe when processing user-supplied extra_body parameters before sending to provider APIs. Import pick_cheapest_chat_models_from_llm_provider for model selection in health checks or fallback logic. Import get_proxy_server_request_headers in custom callbacks to access the original proxy request headers.

Code Reference

Source Location

litellm/litellm_core_utils/llm_request_utils.py (83 lines)

Signature

def _ensure_extra_body_is_safe(extra_body: Optional[Dict]) -> Optional[Dict]

def pick_cheapest_chat_models_from_llm_provider(
    custom_llm_provider: str, n: int = 1
) -> list[str]

def get_proxy_server_request_headers(litellm_params: Optional[dict]) -> dict

Import

from litellm.litellm_core_utils.llm_request_utils import (
    _ensure_extra_body_is_safe,
    pick_cheapest_chat_models_from_llm_provider,
    get_proxy_server_request_headers,
)

I/O Contract

_ensure_extra_body_is_safe

Direction	Name	Type	Description
Input	extra_body	Optional[Dict]	The extra_body dictionary from a request
Output	return	Optional[Dict]	Sanitized extra_body with non-serializable objects converted

pick_cheapest_chat_models_from_llm_provider

Direction	Name	Type	Description
Input	custom_llm_provider	str	Provider name (e.g., "openai", "anthropic")
Input	n	int	Number of cheapest models to return (default: 1)
Output	return	list[str]	List of model name strings sorted by ascending cost

get_proxy_server_request_headers

Direction	Name	Type	Description
Input	litellm_params	Optional[dict]	The litellm_params dictionary from a call
Output	return	dict	Headers dictionary from the proxy server request, or empty dict

Usage Examples

from litellm.litellm_core_utils.llm_request_utils import (
    _ensure_extra_body_is_safe,
    pick_cheapest_chat_models_from_llm_provider,
    get_proxy_server_request_headers,
)

# Sanitize extra_body before sending to provider
extra_body = {"metadata": {"prompt": some_langfuse_prompt_client}}
safe_body = _ensure_extra_body_is_safe(extra_body)

# Pick the 3 cheapest OpenAI chat models
cheapest = pick_cheapest_chat_models_from_llm_provider("openai", n=3)
# e.g., ["gpt-4o-mini", "gpt-3.5-turbo", "gpt-4o"]

# Extract proxy request headers in a custom callback
def my_callback(kwargs, response, start_time, end_time):
    headers = get_proxy_server_request_headers(kwargs.get("litellm_params"))
    user_agent = headers.get("user-agent", "unknown")

Related Pages

• Implementation:BerriAI_Litellm_Health_Check_Helpers -- uses pick_cheapest_chat_models_from_llm_provider for wildcard health checks
• Implementation:BerriAI_Litellm_Get_Model_Cost_Map -- provides the model cost data used for cheapest-model selection