Fast, cached check for whether a callback is an async function. Falls back gracefully if the object cannot be weak-referenced or cached.

Create a tool call dict matching the ChatCompletionMessageToolCall schema.

Return the union of the ``model_info`` of every capability rule matching ``model``. Later rules override earlier ones on key conflicts. Returns `

Normalize a reasoning effort value based on model capabilities. Degradation chains: - "max" → max / xhigh / high - "xhigh" → x

Update data dictionary with model credentials (in-place). Args: data: Data dictionary to update credentials: Credentials fro

Process Gemini embedContent response (single embedding for multimodal input). Args: input: Original input model_response: Em

Return the json str of the request This is currently in BETA, and tested for `/chat/completions` -> `litellm.completion` calls.

Safely get Vertex AI credentials without mutating the litellm_params dict. Unlike get_vertex_ai_credentials(), this does NOT pop val

Recursively remove OCI-incompatible fields from a JSON schema. Strips ``title`` keys, removes ``None``-valued ``default`` entries, normalises

Check if the given model + provider supports 'response_schema' as a param. Parameters: model (str): The model name to be checked. cu

Transform user_api_key_dict to a metadata dict with prefixed keys. Converts keys like 'user_id' to 'user_api_key_user_id' to clearly

Short-circuit web-search-only requests by executing the search directly. Claude Code sends web search as a separate, standalone /v1/

Update the router settings.

Return minimal kwargs for set_attributes with Responses API defaults.

Fingerprint API keys, scrub them, and normalize multipart boundaries. Order matters in two ways: 1. vcrpy's ``filter_headers`` config option

Build OAuth protected resource response with the appropriate URL pattern. For pass-through MCP servers (``MCPServer.is_oauth_passthrough``),

Return every byteItem payload (file scans) submitted to Model Armor.

Check if the user has access to the vector store. Delegates to :func:`can_user_access_vector_store`, which honors: - PROXY_ADMIN bypass

Helper: return dict[attr_name] = value of all set_attribute calls.

Construct Azure realtime WebSocket URL. Args: api_base: Azure API base URL (will be converted from https:// to wss://)

Instantiate ContentFilterGuardrail with a given category.

Internal method to process the cached result Checks the call type and converts the cached result to the appropriate model response o

Decrypts a dictionary of variables and returns them.

Reject the batch if the caller is not authorized for the upload target. For managed files, ``target_model_names`` (from the unified file id)

Legacy path: execute search + build patch + run follow-up call.

Get allowed MCP servers for a team. Unions two sources: - Legacy team.object_permission (mcp_servers, mcp_access_groups,

Get Azure config based on model type. When *base_model* is provided (e.g. ``"azure/gpt-5.2"``), it is used for model-type detection i

Checks the request 'source' if a cli state token was passed in This is used to authenticate through the CLI login flow. The

Resolve the key's unified `access_group_ids` (LiteLLM_AccessGroupTable) to MCP server IDs as additive grants: a group attached to the

Internal function. Returns the token and url for the call. Handles logic if it's google ai studio vs. vertex ai. Returns

Helper to get a valid ClientSession for the current event loop. This handles the case where the session was created in a different

Tolerantly read ``web_search_requests`` from a ``server_tool_use`` value that may be ``None``, a ``dict``, a ``ServerToolUse`` pydantic insta

Handles Connection Errors when reading a Virtual Key from LiteLLM DB Use this if you don't want failed DB queries to block LLM API re

Optional continuous profiling via Grafana Pyroscope. Off by default. Enable with LITELLM_ENABLE_PYROSCOPE=true. Requires: pi

Run invoke_agent_a2a with standard patches applied.

Check if the current environment is already running as the target IAM role. This handles multiple AWS environments: - IRSA (

Check if the input contains multimodal data (data URIs, file references, GCS URLs, or nested lists for combined embeddings). Args:

Check if the error message indicates a database permission error. Permission errors should NOT be marked as applied, as the migratio

Find a deployment whose declared keywords appear in `user_message`. Returns (model_name, matched_keyword) or None when no keyword ma

Check if a token exists in the deprecated keys table and is still within its grace period. Returns the active_token_id if found and valid, o

Build a mock FastAPI Request with configurable headers.

Build a model_call_details dict that mimics real proxy/SDK flows.

Map OpenAI size format to OpenRouter aspect_ratio format. Uses the same mapping as image generation since OpenRouter handles

Stub out the DB helpers resolve_and_validate_end_user_id delegates to.

Reject URL-valued ``model``/``file_id`` unless admin-allowlisted. Some providers (HuggingFace, Oobabooga, Gemini files) accept a URL in the i

Walk the precedence chain and return the first matching credential name. Checks (in order): 1. project_model_config[model_name][provider

Run non-streaming post-call guardrail hooks on an assembled streaming response, then fire success logging via ``dispatch_success_hand

Engine/app-mode search request (vertex_engine_id set).

Only raise exception for "BLOCKED" actions, not for "ANONYMIZED" actions. If `self.mask_request_content` or `self.mask_response_cont

Removes or redacts base64-encoded file data (e.g., PDFs, images, audio) from messages and responses before sending to SQS. B

Strips the 'data: ' prefix from Server-Sent Events (SSE) chunks. Some providers like sagemaker send it as `data:`, need to handle bo