Default to exclude_unset to avoid Pydantic serializer warnings for OpenAIObject-derived types.

Redacts or excludes fields from StandardLoggingPayload before callbacks receive it. This method handles two features: 1. tur

Setup guardrail with all UofT patterns (mimics the policy template sub-guardrail)

Transform a Responses API tools into a Chat Completion tools

Transform generate_content request to litellm completion format Args: model: The model name contents: Genera

Convert Anthropic thinking param to Responses API reasoning param. ``thinking.budget_tokens`` is bucketed via the shared ``r

Update an existing API key's parameters. Parameters: - key: str - The key to update - key_alias: Optional[str] - User-friendly key a

Verify a password against a stored hash. Supports scrypt and SHA256.

Encode model_id into an output item ID for encrypted-content items. Format: ``encitem_{base64("litellm:model_id:{model_id};item_id:{original_

Shared response-building logic (also used by the handler).

Build a synthetic FastAPI Request for sampling sub-calls. Converts the original MCP connection's HTTP headers into ASGI scope format so that

Call a regular MCP tool using the MCP client. Args: mcp_server: The MCP server configuration original_tool_n

Creates an AiohttpTransport with RequestNotRead error handling Note: aiohttp TCPConnector ssl parameter accepts: - SSLContex

Return (start_date, end_date) for the common 7-day range used in UI spend tests.

Extract citation tokens and search queries from Perplexity API response and add them to the usage object using standard LiteLLM field

If `exc` is an HTTPException with a dict `detail`, mutate it in place to add `guardrail_name` and `guardrail_mode` taken from the callback in

Walk a Bedrock Converse request body and collect text content. Returns (texts, holders) where each holder is the (container, key) pair t

Check if database is connected and raise HTTPException if not.

Helper method to get tools from a single MCP server with prefixed names. Args: server (MCPServer): The server to query t

Check if a model has zero cost (no configured pricing). Uses the router's get_model_group_info method to get pricing information. Args:

Build a MagicMock prisma_client whose user-credentials table returns ``row`` for find_unique and behaves async-correctly for upsert/find_many.

Enforce the authorized model and apply Presidio PII masking to a ``response.create`` message before it is forwarded to the upstream

Validate a managed GCS file_id and return (bucket, url-encoded-object-path).

Subset of the call kwargs limited to `keys`, for exact comparison.

Convert a list of database records to domain models.

Apply guardrails to inputs from policy names and/or a direct list of guardrail names. Runs all guardrails in order; if one fails, the error

For masking logged request/response. Return a modified version of the request/result.

Actual callback function that Boto3 will call.

Check if background responses are complete and track their cost. - Get all status="queued" or "in_progress" and file_purpose="respons

Parse a Responses API streaming chunk and convert to OpenAI format. Args: chunk: Dict containing the Responses API event

Public entry point — returns the model cost map dict. 1. If ``LITELLM_LOCAL_MODEL_COST_MAP`` is set, uses the local backup only. 2. Othe

Perform a health check on a specific MCP server. Args: server_id: The ID of the server to health check mcp_a

Query the DB for managed IDs the caller owns and return an OpenAI-style paginated list response. Returns ``None`` when ``prisma_client`` is u

Get a paginated list of teams with filtering and sorting options. Parameters: user_id: Optional[str] Only return teams w

Performs the actual redaction on the logging object and result.

Main entry point - find and rotate keys that are due for rotation. Uses PodLockManager to ensure only one pod runs rotation in multi-

Process SSO JWT access token and extract team IDs and user role if available. This function decodes the JWT access token and extracts team I

Render a prompt template with the given variables. Args: prompt_id: The ID of the prompt template to render

Resolve the final guardrails for a single policy, including inheritance. This method: 1. Resolves the inheritance chain

Save the active rules, let the test install its own, then restore.

Recursively walk a request body dict/list and resolve any passthrough managed IDs. Skips litellm internal keys (``litellm_*``). Returns

Safely sets an attribute on the span, ensuring the value is a primitive type.

Sanitize messages for tool calling to handle common issues when modify_params=True: Case A: Missing tool_result for tool_use (orphaned tool

Receive message from backend and store it

Converts the ProxyException instance to a dictionary.

Args: request: TokenCountRequest call_endpoint: bool - When set to "True" it will call the token counting endpoint - e.g Anthropi

Transform a Chat Completion response into a Responses API response

Transform OpenAI embedding input to Gemini batchEmbedContents format. Each input element becomes a separate EmbedContentRequest, supporting

Transform OpenAI embedding input to Gemini embedContent format (multimodal). Args: input: GeminiEmbeddingInput with text, data URIs,

Transform request to provider-specific format. Returns: TextToSpeechRequestData: A structured dict containing:

Transform the Veo video creation response. Veo returns: { "name": "operations/generate_1234567890",

Update the polling state in Redis Uses OpenAI Response object format with native status types: https://platform.openai.com/d

Add proxy hooks to litellm.callbacks