diff --git a/gateway/routes/allowlist.py b/gateway/routes/allowlist.py index c81082fa46b..f7557983b91 100644 --- a/gateway/routes/allowlist.py +++ b/gateway/routes/allowlist.py @@ -95,6 +95,7 @@ GATEWAY_PATH_PREFIXES: tuple[str, ...] = ( "/vertex-ai/", "/assemblyai/", "/eu.assemblyai/", + "/deepgram/", "/langfuse/", "/vllm/", "/mistral/", diff --git a/litellm/constants.py b/litellm/constants.py index fd21a05d00d..b76f24fc84b 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -320,6 +320,9 @@ REALTIME_CREDENTIAL_RESOLUTION_TIMEOUT_SECONDS: Final = float( # RFC 6455 caps the close frame payload at 125 bytes, 2 of which carry the status code WEBSOCKET_CLOSE_REASON_MAX_BYTES: Final = 123 +DEEPGRAM_DEFAULT_API_BASE: Final = "https://api.deepgram.com/v1" +DEEPGRAM_LISTEN_DEFAULT_MODEL: Final = "nova-3" + BEDROCK_REALTIME_PENDING_SESSION_UPDATE_SCOPE_KEY: Final = "litellm.bedrock_realtime.pending_session_update" BEDROCK_REALTIME_SESSION_COMMITTED_SCOPE_KEY: Final = "litellm.bedrock_realtime.session_committed" BEDROCK_REALTIME_COMMITTED_FAILURE_SCOPE_KEY: Final = "litellm.bedrock_realtime.committed_failure" diff --git a/litellm/llms/deepgram/common_utils.py b/litellm/llms/deepgram/common_utils.py index a741b092a36..9b071ac8321 100644 --- a/litellm/llms/deepgram/common_utils.py +++ b/litellm/llms/deepgram/common_utils.py @@ -1,5 +1,200 @@ +import math +from collections.abc import Mapping, Sequence +from types import MappingProxyType +from typing import Final +from urllib.parse import parse_qs, urlparse + +import httpx + +import litellm +from litellm.constants import DEEPGRAM_DEFAULT_API_BASE, DEEPGRAM_LISTEN_DEFAULT_MODEL from litellm.llms.base_llm.chat.transformation import BaseLLMException +from litellm.types.utils import LlmProviders + +_WEBSOCKET_SCHEMES: Final = MappingProxyType({"https": "wss", "http": "ws", "wss": "wss", "ws": "ws"}) +DEEPGRAM_LISTEN_CALLBACK_PARAMS: Final = frozenset({"callback", "callback_method"}) +DEEPGRAM_LISTEN_STREAMING_PRICING_PREFIX: Final = "streaming/" +DEEPGRAM_LISTEN_MULTILINGUAL_LANGUAGE: Final = "multi" +DEEPGRAM_LISTEN_MULTILINGUAL_PRICING_SUFFIX: Final = "-multilingual" +DEEPGRAM_LISTEN_ADDON_PRICING_PARAMS: Final = MappingProxyType( + { + "redact": "redact", + "keyterm": "keyterm", + "detect_entities": "detect_entities", + "diarize": "diarize", + "diarize_model": "diarize", + } +) +_DISABLED_PARAM_VALUES: Final = frozenset({"", "false"}) +_SINGLE_VALUED_PARAMS: Final = frozenset({"model", "language"}) class DeepgramException(BaseLLMException): pass + + +def deepgram_listen_requested_model(query_string: str) -> str: + return httpx.QueryParams(query_string).get("model") or DEEPGRAM_LISTEN_DEFAULT_MODEL + + +def _first_occurrences(query_string: str) -> httpx.QueryParams: + """Authorization and pricing read the first ``model`` and ``language`` value; Deepgram must not see a second one.""" + items: Final = httpx.QueryParams(query_string).multi_items() + return httpx.QueryParams( + tuple( + (key, value) + for index, (key, value) in enumerate(items) + if key not in _SINGLE_VALUED_PARAMS or all(earlier != key for earlier, _ in items[:index]) + ) + ) + + +def deepgram_listen_websocket_target(api_base: str | None, query_string: str) -> str: + listen_url: Final = httpx.URL(f"{(api_base or DEEPGRAM_DEFAULT_API_BASE).rstrip('/')}/listen") + websocket_url: Final = listen_url.copy_with(scheme=_WEBSOCKET_SCHEMES.get(listen_url.scheme, listen_url.scheme)) + params: Final = _first_occurrences(query_string) + query: Final = params if params.get("model") else params.remove("model").add("model", DEEPGRAM_LISTEN_DEFAULT_MODEL) + return f"{websocket_url}?{query}" + + +def deepgram_listen_callback_params(query_string: str) -> tuple[str, ...]: + return tuple(sorted(DEEPGRAM_LISTEN_CALLBACK_PARAMS.intersection(httpx.QueryParams(query_string).keys()))) + + +def deepgram_listen_model(upstream_url: str) -> str: + models: Final = parse_qs(urlparse(upstream_url).query).get("model") + return models[0] if models else DEEPGRAM_LISTEN_DEFAULT_MODEL + + +def _param_enabled(values: Sequence[str]) -> bool: + return any(value.strip().lower() not in _DISABLED_PARAM_VALUES for value in values) + + +def deepgram_listen_pricing_model(upstream_url: str) -> str: + """Registry key, without the provider prefix, for the per-second base rate Deepgram bills a streaming session at: + the multilingual streaming entry when ``language=multi``, otherwise the model's own streaming entry. Pre-recorded + entries are never a substitute: Deepgram prices the two products differently.""" + streaming: Final = f"{DEEPGRAM_LISTEN_STREAMING_PRICING_PREFIX}{deepgram_listen_model(upstream_url)}" + language: Final = parse_qs(urlparse(upstream_url).query).get("language", ("",))[0] + if language.strip().lower() == DEEPGRAM_LISTEN_MULTILINGUAL_LANGUAGE: + return f"{streaming}{DEEPGRAM_LISTEN_MULTILINGUAL_PRICING_SUFFIX}" + return streaming + + +def deepgram_listen_registry_key(upstream_url: str) -> str: + return f"{LlmProviders.DEEPGRAM.value}/{deepgram_listen_pricing_model(upstream_url)}" + + +def deepgram_listen_is_priced(upstream_url: str) -> bool: + """Only an exact registry hit counts: the cost calculator resolves a missing ``streaming/`` row to the + pre-recorded ```` row, which is not the rate Deepgram bills a WebSocket session at.""" + registry_key: Final = deepgram_listen_registry_key(upstream_url) + try: + model_info: Final = litellm.get_model_info(model=registry_key, custom_llm_provider=LlmProviders.DEEPGRAM.value) + except Exception: + return False + return model_info["key"] == registry_key + + +def deepgram_listen_addon_pricing_models(upstream_url: str) -> tuple[str, ...]: + params: Final = parse_qs(urlparse(upstream_url).query) + return tuple( + sorted( + frozenset( + f"{DEEPGRAM_LISTEN_STREAMING_PRICING_PREFIX}{addon}" + for param, addon in DEEPGRAM_LISTEN_ADDON_PRICING_PARAMS.items() + if _param_enabled(params.get(param, ())) + ) + ) + ) + + +def _channel_count(value: object) -> int | None: + if isinstance(value, bool) or not isinstance(value, int): + return None + return value if value >= 1 else None + + +def _results_channel_count(frame: Mapping[str, object]) -> int | None: + channel_index: Final = frame.get("channel_index") + if not isinstance(channel_index, list) or len(channel_index) != 2: + return None + return _channel_count(channel_index[1]) + + +def _declared_channel_count(upstream_url: str) -> int | None: + declared: Final = parse_qs(urlparse(upstream_url).query).get("channels") + if not declared or not declared[0].isdigit(): + return None + return _channel_count(int(declared[0])) + + +def deepgram_listen_channel_count(websocket_messages: Sequence[Mapping[str, object]], upstream_url: str) -> int: + metadata_channels: Final = tuple( + channels + for frame in websocket_messages + if frame.get("type") == "Metadata" + if (channels := _channel_count(frame.get("channels"))) is not None + ) + if metadata_channels: + return metadata_channels[-1] + results_channels: Final = tuple( + channels + for frame in websocket_messages + if frame.get("type") == "Results" + if (channels := _results_channel_count(frame)) is not None + ) + if results_channels: + return max(results_channels) + return _declared_channel_count(upstream_url) or 1 + + +def _seconds(value: object) -> float | None: + if isinstance(value, bool) or not isinstance(value, (int, float)): + return None + return float(value) if math.isfinite(value) and value >= 0 else None + + +def _results_frame_end(frame: Mapping[str, object]) -> float | None: + start: Final = _seconds(frame.get("start")) + duration: Final = _seconds(frame.get("duration")) + return None if start is None or duration is None else start + duration + + +def _final_transcript(frame: Mapping[str, object]) -> str | None: + if frame.get("is_final") is not True: + return None + channel: Final = frame.get("channel") + alternatives: Final = channel.get("alternatives") if isinstance(channel, Mapping) else None + first: Final = alternatives[0] if isinstance(alternatives, list) and alternatives else None + transcript: Final = first.get("transcript") if isinstance(first, Mapping) else None + return transcript if isinstance(transcript, str) and transcript else None + + +def deepgram_listen_audio_seconds(websocket_messages: Sequence[Mapping[str, object]]) -> float: + metadata_durations: Final = tuple( + duration + for frame in websocket_messages + if frame.get("type") == "Metadata" + if (duration := _seconds(frame.get("duration"))) is not None and duration > 0 + ) + if metadata_durations: + return metadata_durations[-1] + return max( + ( + end + for frame in websocket_messages + if frame.get("type") == "Results" + if (end := _results_frame_end(frame)) is not None + ), + default=0.0, + ) + + +def deepgram_listen_transcript(websocket_messages: Sequence[Mapping[str, object]]) -> str: + return " ".join( + transcript + for frame in websocket_messages + if frame.get("type") == "Results" + if (transcript := _final_transcript(frame)) is not None + ) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index e0c726b1e2c..02df6897b41 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -20853,6 +20853,96 @@ "/v1/audio/transcriptions" ] }, + "deepgram/streaming/nova-3": { + "input_cost_per_second": 8e-05, + "litellm_provider": "deepgram", + "metadata": { + "calculation": "$0.0048/60 seconds = $0.00008000 per second", + "note": "Nova-3 monolingual streaming, pay as you go", + "original_pricing_per_minute": 0.0048 + }, + "mode": "audio_transcription", + "output_cost_per_second": 0.0, + "source": "https://deepgram.com/pricing", + "supported_endpoints": [ + "/v1/listen" + ] + }, + "deepgram/streaming/nova-3-multilingual": { + "input_cost_per_second": 9.667e-05, + "litellm_provider": "deepgram", + "metadata": { + "calculation": "$0.0058/60 seconds = $0.00009667 per second", + "note": "Nova-3 multilingual (language=multi) streaming, pay as you go", + "original_pricing_per_minute": 0.0058 + }, + "mode": "audio_transcription", + "output_cost_per_second": 0.0, + "source": "https://deepgram.com/pricing", + "supported_endpoints": [ + "/v1/listen" + ] + }, + "deepgram/streaming/redact": { + "input_cost_per_second": 3.333e-05, + "litellm_provider": "deepgram", + "metadata": { + "calculation": "$0.0020/60 seconds = $0.00003333 per second", + "note": "Redaction add-on (redact query param), streaming, pay as you go", + "original_pricing_per_minute": 0.002 + }, + "mode": "audio_transcription", + "output_cost_per_second": 0.0, + "source": "https://deepgram.com/pricing", + "supported_endpoints": [ + "/v1/listen" + ] + }, + "deepgram/streaming/keyterm": { + "input_cost_per_second": 2.167e-05, + "litellm_provider": "deepgram", + "metadata": { + "calculation": "$0.0013/60 seconds = $0.00002167 per second", + "note": "Keyterm Prompting add-on (keyterm query param), streaming, pay as you go", + "original_pricing_per_minute": 0.0013 + }, + "mode": "audio_transcription", + "output_cost_per_second": 0.0, + "source": "https://deepgram.com/pricing", + "supported_endpoints": [ + "/v1/listen" + ] + }, + "deepgram/streaming/detect_entities": { + "input_cost_per_second": 2.833e-05, + "litellm_provider": "deepgram", + "metadata": { + "calculation": "$0.0017/60 seconds = $0.00002833 per second", + "note": "Entity Detection add-on (detect_entities query param), streaming, pay as you go", + "original_pricing_per_minute": 0.0017 + }, + "mode": "audio_transcription", + "output_cost_per_second": 0.0, + "source": "https://deepgram.com/pricing", + "supported_endpoints": [ + "/v1/listen" + ] + }, + "deepgram/streaming/diarize": { + "input_cost_per_second": 3.333e-05, + "litellm_provider": "deepgram", + "metadata": { + "calculation": "$0.0020/60 seconds = $0.00003333 per second", + "note": "Speaker Diarization add-on (diarize / diarize_model query params), streaming, pay as you go", + "original_pricing_per_minute": 0.002 + }, + "mode": "audio_transcription", + "output_cost_per_second": 0.0, + "source": "https://deepgram.com/pricing", + "supported_endpoints": [ + "/v1/listen" + ] + }, "deepgram/whisper": { "input_cost_per_second": 0.0001, "litellm_provider": "deepgram", diff --git a/litellm/proxy/_lazy_features.py b/litellm/proxy/_lazy_features.py index ccbf0104d81..79aa2d16d24 100644 --- a/litellm/proxy/_lazy_features.py +++ b/litellm/proxy/_lazy_features.py @@ -201,6 +201,7 @@ LAZY_FEATURES: Final[tuple[LazyFeature, ...]] = ( "/cohere/", "/comprehendmedical", "/cursor/", + "/deepgram/", "/eu.assemblyai/", "/gemini/", "/gigachat/", diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py index 262f4e40ef4..76a51627d0c 100644 --- a/litellm/proxy/_types.py +++ b/litellm/proxy/_types.py @@ -71,6 +71,7 @@ from litellm.types.utils import ( StandardLoggingVectorStoreRequest, StandardPassThroughResponseObject, TextCompletionResponse, + TranscriptionResponse, ) from litellm.types.videos.main import VideoObject @@ -491,6 +492,7 @@ class LiteLLMRoutes(enum.Enum): "/gigachat", "/watsonx", "/nvidia_nim", + "/deepgram", ] ######################################################### @@ -4772,6 +4774,7 @@ PassThroughEndpointLoggingResultValues = ( | VideoObject | StandardPassThroughResponseObject | ResponsesAPIResponse + | TranscriptionResponse ) diff --git a/litellm/proxy/auth/user_api_key_auth.py b/litellm/proxy/auth/user_api_key_auth.py index abfd6e94372..6f6db73f99e 100644 --- a/litellm/proxy/auth/user_api_key_auth.py +++ b/litellm/proxy/auth/user_api_key_auth.py @@ -632,9 +632,11 @@ def _apply_budget_limits_to_end_user_params( verbose_proxy_logger.debug("Applied budget limits to end user %s", end_user_id) -async def user_api_key_auth_websocket(websocket: WebSocket): - # Accept the WebSocket connection +async def user_api_key_auth_websocket(websocket: WebSocket) -> UserAPIKeyAuth: + return await user_api_key_auth_websocket_for_model(websocket, model=websocket.query_params.get("model")) + +async def user_api_key_auth_websocket_for_model(websocket: WebSocket, model: str | None) -> UserAPIKeyAuth: ws_scope: Final = websocket.scope or {} scope_headers: Final = list(ws_scope.get("headers") or []) # ``get_request_route`` falls back to ``request.url.path`` when @@ -654,10 +656,6 @@ async def user_api_key_auth_websocket(websocket: WebSocket): request._url = websocket.url - query_params: Final = websocket.query_params - - model: Final = query_params.get("model") - async def return_body(): return _realtime_request_body(model) diff --git a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py index ed70369506d..44d9f11360d 100644 --- a/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py +++ b/litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py @@ -45,6 +45,13 @@ from litellm.litellm_core_utils.aws_partition import get_aws_dns_suffix from litellm.llms.anthropic.common_utils import AnthropicModelInfo from litellm.llms.azure.passthrough.transformation import foreign_azure_deployment from litellm.llms.custom_httpx.http_handler import get_async_httpx_client +from litellm.llms.deepgram.common_utils import ( + deepgram_listen_callback_params, + deepgram_listen_is_priced, + deepgram_listen_registry_key, + deepgram_listen_requested_model, + deepgram_listen_websocket_target, +) from litellm.llms.nvidia_nim.passthrough.transformation import nvidia_nim_model_group_in_path from litellm.llms.vertex_ai.vertex_llm_base import VertexBase from litellm.passthrough.main import AsyncPassthroughStreamingResponse @@ -57,6 +64,7 @@ from litellm.proxy.auth.user_api_key_auth import ( is_no_auth_dev_mode, user_api_key_auth, user_api_key_auth_websocket, + user_api_key_auth_websocket_for_model, ) from litellm.proxy.common_request_processing import open_sse_before_first_byte from litellm.proxy.common_utils.http_parsing_utils import ( @@ -2925,7 +2933,7 @@ async def _openai_websocket_refusal( return None -class _OpenAIWebsocketRelay(Protocol): +class _WebsocketRelay(Protocol): async def __call__( self, *, @@ -2939,7 +2947,7 @@ class _OpenAIWebsocketRelay(Protocol): ) -> None: ... -def _openai_websocket_relay() -> _OpenAIWebsocketRelay: +def _websocket_relay() -> _WebsocketRelay: return websocket_passthrough_request @@ -2957,6 +2965,15 @@ def _proxy_model_allowlists() -> _OpenAIWebsocketModelAllowlists: return resolve +def _negotiated_websocket_subprotocol(websocket: WebSocket) -> str | None: + requested_subprotocols: Final = tuple( + protocol.strip() + for protocol in (websocket.headers.get("sec-websocket-protocol") or "").split(",") + if protocol.strip() + ) + return requested_subprotocols[0] if requested_subprotocols else None + + @router.websocket("/openai_passthrough/{endpoint:path}") @router.websocket("/openai/{endpoint:path}") async def openai_websocket_proxy_route( @@ -2964,16 +2981,11 @@ async def openai_websocket_proxy_route( endpoint: str, user_api_key_dict: Annotated[UserAPIKeyAuth, Depends(user_api_key_auth_websocket)], general_settings: Annotated[Mapping[str, object], Depends(_proxy_general_settings)], - relay: Annotated[_OpenAIWebsocketRelay, Depends(_openai_websocket_relay)], + relay: Annotated[_WebsocketRelay, Depends(_websocket_relay)], model_allowlists: Annotated[_OpenAIWebsocketModelAllowlists, Depends(_proxy_model_allowlists)], ) -> None: """WebSocket passthrough for OpenAI prefixes (realtime / responses.connect).""" - requested_subprotocols: Final = tuple( - protocol.strip() - for protocol in (websocket.headers.get("sec-websocket-protocol") or "").split(",") - if protocol.strip() - ) - negotiated_subprotocol: Final = requested_subprotocols[0] if requested_subprotocols else None + negotiated_subprotocol: Final = _negotiated_websocket_subprotocol(websocket) refusal: Final = await _openai_websocket_refusal(user_api_key_dict, general_settings, model_allowlists) if refusal is not None: @@ -3032,6 +3044,69 @@ async def openai_websocket_proxy_route( ) +_DEEPGRAM_WS_MISSING_KEY_REASON: Final = ( + "Required 'DEEPGRAM_API_KEY' in environment to make pass-through calls to Deepgram." +) +_DEEPGRAM_WS_CALLBACK_REASON: Final = "Deepgram callback delivery is not supported through the proxy: remove {params}" +_DEEPGRAM_WS_UNPRICED_REASON: Final = ( + "No streaming price for '{registry_key}': add it to the model cost map to enable it" +) + + +async def deepgram_listen_user_api_key_auth(websocket: WebSocket) -> UserAPIKeyAuth: + return await user_api_key_auth_websocket_for_model( + websocket, model=deepgram_listen_requested_model(websocket.url.query) + ) + + +@router.websocket("/deepgram/v1/listen") +@router.websocket("/deepgram/listen") +async def deepgram_listen_websocket_route( + websocket: WebSocket, + user_api_key_dict: Annotated[UserAPIKeyAuth, Depends(deepgram_listen_user_api_key_auth)], + relay: Annotated[_WebsocketRelay, Depends(_websocket_relay)], +) -> None: + deepgram_api_key: Final = passthrough_endpoint_router.get_credentials( + custom_llm_provider=litellm.LlmProviders.DEEPGRAM.value, + region_name=None, + ) + if deepgram_api_key is None: + await websocket.close(code=1011, reason=_DEEPGRAM_WS_MISSING_KEY_REASON) + return + + await websocket.accept(subprotocol=_negotiated_websocket_subprotocol(websocket)) + callback_params: Final = deepgram_listen_callback_params(websocket.url.query) + if callback_params: + await websocket.close( + code=1008, + reason=_DEEPGRAM_WS_CALLBACK_REASON.format(params=", ".join(callback_params)), + ) + return + + target: Final = deepgram_listen_websocket_target( + api_base=get_secret_str("DEEPGRAM_API_BASE"), + query_string=websocket.url.query, + ) + if not deepgram_listen_is_priced(target): + await websocket.close( + code=1008, + reason=_DEEPGRAM_WS_UNPRICED_REASON.format(registry_key=deepgram_listen_registry_key(target)), + ) + return + + await relay( + websocket=websocket, + target=target, + custom_headers={ # mutable-ok: websocket_passthrough_request requires a plain dict of upstream headers + "Authorization": f"Token {deepgram_api_key}" + }, + user_api_key_dict=user_api_key_dict, + forward_headers=False, + endpoint=websocket.url.path, + accept_websocket=False, + ) + + class BaseOpenAIPassThroughHandler: @staticmethod async def _base_openai_pass_through_handler( diff --git a/litellm/proxy/pass_through_endpoints/llm_provider_handlers/deepgram_listen_passthrough_logging_handler.py b/litellm/proxy/pass_through_endpoints/llm_provider_handlers/deepgram_listen_passthrough_logging_handler.py new file mode 100644 index 00000000000..8386c154600 --- /dev/null +++ b/litellm/proxy/pass_through_endpoints/llm_provider_handlers/deepgram_listen_passthrough_logging_handler.py @@ -0,0 +1,96 @@ +from collections.abc import Mapping, Sequence +from types import MappingProxyType +from typing import Final +from urllib.parse import urlparse + +import litellm +from litellm._logging import verbose_proxy_logger +from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj +from litellm.llms.deepgram.common_utils import ( + deepgram_listen_addon_pricing_models, + deepgram_listen_audio_seconds, + deepgram_listen_channel_count, + deepgram_listen_is_priced, + deepgram_listen_model, + deepgram_listen_pricing_model, + deepgram_listen_registry_key, + deepgram_listen_transcript, +) +from litellm.proxy._types import PassThroughEndpointLoggingTypedDict +from litellm.types.utils import TranscriptionResponse + +DEEPGRAM_LISTEN_ROUTE_SUFFIX: Final = "/listen" + + +def _registry_cost(response: TranscriptionResponse, pricing_model: str) -> float | None: + try: + return litellm.completion_cost( + completion_response=response, + model=pricing_model, + custom_llm_provider=litellm.LlmProviders.DEEPGRAM.value, + call_type="transcription", + ) + except Exception as e: # noqa: BLE001 # an unpriced entry must not lose the spend row, only its cost + verbose_proxy_logger.debug("Deepgram listen passthrough: no registry price for '%s': %s", pricing_model, e) + return None + + +def _audio_cost(response: TranscriptionResponse, upstream_url: str) -> float | None: + if not deepgram_listen_is_priced(upstream_url): + verbose_proxy_logger.warning( + "Deepgram listen passthrough: no registry entry '%s'", deepgram_listen_registry_key(upstream_url) + ) + return None + base_cost: Final = _registry_cost(response, deepgram_listen_pricing_model(upstream_url)) + if base_cost is None: + return None + addon_costs: Final = tuple( + _registry_cost(response, pricing_model) for pricing_model in deepgram_listen_addon_pricing_models(upstream_url) + ) + return base_cost + sum(cost for cost in addon_costs if cost is not None) + + +class DeepgramListenPassthroughLoggingHandler: + @staticmethod + def is_deepgram_listen_route(url_route: str) -> bool: + path: Final = urlparse(url_route).path + return "/deepgram/" in path and path.endswith(DEEPGRAM_LISTEN_ROUTE_SUFFIX) + + def deepgram_listen_passthrough_handler( + self, + websocket_messages: Sequence[Mapping[str, object]], + logging_obj: LiteLLMLoggingObj, + upstream_url: str, + kwargs: Mapping[str, object] = MappingProxyType({}), + ) -> PassThroughEndpointLoggingTypedDict: + model: Final = deepgram_listen_model(upstream_url) + audio_seconds: Final = deepgram_listen_audio_seconds(websocket_messages) + channels: Final = deepgram_listen_channel_count(websocket_messages, upstream_url) + billed_seconds: Final = audio_seconds * channels + response: Final = TranscriptionResponse(text=deepgram_listen_transcript(websocket_messages)) + response._hidden_params["audio_transcription_duration"] = billed_seconds # pyright: ignore[reportPrivateUsage] # the cost calculator reads the billed duration off the response's hidden params + response_cost: Final = _audio_cost(response, upstream_url) + response._hidden_params["response_cost"] = response_cost # pyright: ignore[reportPrivateUsage] # the logger reads a precomputed cost off the response's hidden params + + provider: Final = litellm.LlmProviders.DEEPGRAM.value + logging_obj.model = model # rebind-ok: the spend logger reads model and cost off the shared logging object + logging_obj.model_call_details["model"] = model # rebind-ok: same shared logging object + logging_obj.model_call_details["custom_llm_provider"] = provider # rebind-ok: same shared logging object + logging_obj.model_call_details["response_cost"] = response_cost # rebind-ok: same shared logging object + verbose_proxy_logger.debug( + "Deepgram listen passthrough cost tracking: model %s, audio seconds %s, channels %s, cost %s", + model, + audio_seconds, + channels, + response_cost, + ) + logging_result: Final[PassThroughEndpointLoggingTypedDict] = { + "result": response, + "kwargs": { + **kwargs, + "model": model, + "custom_llm_provider": provider, + "response_cost": response_cost, + }, + } + return logging_result diff --git a/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py b/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py index fddb0382f06..ae1c543de56 100644 --- a/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py +++ b/litellm/proxy/pass_through_endpoints/pass_through_endpoints.py @@ -8,7 +8,7 @@ from base64 import b64encode from collections.abc import AsyncGenerator, Callable, Iterable, Mapping, Sequence from dataclasses import dataclass from datetime import datetime -from itertools import groupby +from itertools import count, groupby from types import MappingProxyType from typing import TYPE_CHECKING, Any, Final, TypedDict, cast from urllib.parse import urlencode, urlparse @@ -2122,6 +2122,14 @@ def _resolved_vertex_live_setup( return {**setup_data, "model": setup_model_rewriter(setup_model)} +def _json_object_frame(frame: str | bytes) -> dict[str, object] | None: + try: + decoded: Final = json.loads(frame if isinstance(frame, str) else frame.decode("utf-8")) + except (json.JSONDecodeError, UnicodeDecodeError): + return None + return decoded if isinstance(decoded, dict) else None + + def _truncated_close_reason(reason: str) -> str: """ Fit a close reason inside the byte budget a WebSocket close frame allows, without splitting a character @@ -2403,70 +2411,41 @@ async def websocket_passthrough_request( ) await upstream_ws.close() + def _extract_vertex_live_model_from_setup_response(setup_response: Mapping[str, object]) -> None: + extracted_model: Final = _extract_model_from_vertex_ai_setup(setup_response) + if not extracted_model: + verbose_proxy_logger.warning( + "WebSocket passthrough (%s): Failed to extract model from server setup response: %s", + endpoint, + setup_response, + ) + return + kwargs["model"] = extracted_model + kwargs["custom_llm_provider"] = "vertex_ai_language_models" + logging_obj.model = extracted_model + logging_obj.model_call_details["model"] = extracted_model + logging_obj.model_call_details["custom_llm_provider"] = "vertex_ai_language_models" + + is_vertex_live: Final = bool(endpoint and "/vertex_ai/live" in endpoint) + json_frame_ordinal: Final = count() + + async def relay_upstream_frame(upstream_message: str | bytes) -> None: + if isinstance(upstream_message, bytes): + await websocket.send_bytes(upstream_message) + else: + await websocket.send_text(upstream_message) + message_data: Final = _json_object_frame(upstream_message) + if message_data is None: + return + if is_vertex_live and next(json_frame_ordinal) == 0: + _extract_vertex_live_model_from_setup_response(message_data) + return + websocket_messages.append(message_data) + async def forward_upstream_to_client() -> Close | None: - """Forward messages from upstream to client WebSocket, returning the upstream's close frame""" try: - # Wait for the first response from upstream - raw_response = await upstream_ws.recv(decode=False) - # Ensure raw_response is bytes before decoding - if isinstance(raw_response, str): - raw_response = raw_response.encode("utf-8") - setup_response: Final[Mapping[str, object]] = json.loads(raw_response.decode("utf-8")) - verbose_proxy_logger.debug("Setup response: %s", setup_response) - - # Extract model and provider from setup response for Vertex AI Live - if endpoint and "/vertex_ai/live" in endpoint: - verbose_proxy_logger.debug( - "WebSocket passthrough (%s): Processing server setup response for model extraction", - endpoint, - ) - extracted_model: Final = _extract_model_from_vertex_ai_setup(setup_response) - if extracted_model: - kwargs["model"] = extracted_model - kwargs["custom_llm_provider"] = "vertex_ai_language_models" - # Update logging object with correct model - logging_obj.model = extracted_model - logging_obj.model_call_details["model"] = extracted_model - logging_obj.model_call_details["custom_llm_provider"] = "vertex_ai_language_models" - verbose_proxy_logger.debug( - "WebSocket passthrough (%s): Successfully extracted model '%s' and set provider to 'vertex_ai' from server setup response", - endpoint, - extracted_model, - ) - else: - verbose_proxy_logger.warning( - "WebSocket passthrough (%s): Failed to extract model from server setup response: %s", - endpoint, - setup_response, - ) - else: - verbose_proxy_logger.debug( - "WebSocket passthrough (%s): Not a Vertex AI Live endpoint, skipping model extraction", - endpoint, - ) - - # Send the setup response to the client - await websocket.send_text(json.dumps(setup_response)) - - # Now continuously forward messages from upstream to client - async for upstream_message in upstream_ws: - if isinstance(upstream_message, bytes): - await websocket.send_bytes(upstream_message) - # Parse and collect for cost tracking - try: - message_data: dict[str, object] = json.loads(upstream_message.decode()) - websocket_messages.append(message_data) - except (json.JSONDecodeError, UnicodeDecodeError): - pass - else: - await websocket.send_text(upstream_message) - # Parse and collect for cost tracking - try: - message_data = json.loads(upstream_message) - websocket_messages.append(message_data) - except json.JSONDecodeError: - pass - + while True: + await relay_upstream_frame(await upstream_ws.recv()) except (ConnectionClosedOK, ConnectionClosedError) as e: verbose_proxy_logger.debug("Upstream WebSocket connection closed: %s", e) return e.rcvd diff --git a/litellm/proxy/pass_through_endpoints/success_handler.py b/litellm/proxy/pass_through_endpoints/success_handler.py index 96fd081924a..de1a8ae1d93 100644 --- a/litellm/proxy/pass_through_endpoints/success_handler.py +++ b/litellm/proxy/pass_through_endpoints/success_handler.py @@ -26,6 +26,9 @@ from .llm_provider_handlers.cohere_passthrough_logging_handler import ( from .llm_provider_handlers.cursor_passthrough_logging_handler import ( CursorPassthroughLoggingHandler, ) +from .llm_provider_handlers.deepgram_listen_passthrough_logging_handler import ( + DeepgramListenPassthroughLoggingHandler, +) from .llm_provider_handlers.gemini_passthrough_logging_handler import ( GeminiPassthroughLoggingHandler, ) @@ -349,6 +352,21 @@ class PassThroughEndpointLogging: standard_logging_response_object = vertex_ai_live_handler_result["result"] kwargs = vertex_ai_live_handler_result["kwargs"] + elif DeepgramListenPassthroughLoggingHandler.is_deepgram_listen_route(url_route): + deepgram_handler_result: Final = ( + DeepgramListenPassthroughLoggingHandler().deepgram_listen_passthrough_handler( + websocket_messages=tuple( + message + for message in (response_body if isinstance(response_body, list) else ()) + if isinstance(message, dict) + ), + logging_obj=logging_obj, + upstream_url=str(httpx_response.request.url), + kwargs=kwargs, + ) + ) + standard_logging_response_object = deepgram_handler_result["result"] # rebind-ok: elif-chain + kwargs = deepgram_handler_result["kwargs"] # rebind-ok: elif-chain contract return_dict["standard_logging_response_object"] = standard_logging_response_object return_dict["kwargs"] = kwargs diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index e0c726b1e2c..02df6897b41 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -20853,6 +20853,96 @@ "/v1/audio/transcriptions" ] }, + "deepgram/streaming/nova-3": { + "input_cost_per_second": 8e-05, + "litellm_provider": "deepgram", + "metadata": { + "calculation": "$0.0048/60 seconds = $0.00008000 per second", + "note": "Nova-3 monolingual streaming, pay as you go", + "original_pricing_per_minute": 0.0048 + }, + "mode": "audio_transcription", + "output_cost_per_second": 0.0, + "source": "https://deepgram.com/pricing", + "supported_endpoints": [ + "/v1/listen" + ] + }, + "deepgram/streaming/nova-3-multilingual": { + "input_cost_per_second": 9.667e-05, + "litellm_provider": "deepgram", + "metadata": { + "calculation": "$0.0058/60 seconds = $0.00009667 per second", + "note": "Nova-3 multilingual (language=multi) streaming, pay as you go", + "original_pricing_per_minute": 0.0058 + }, + "mode": "audio_transcription", + "output_cost_per_second": 0.0, + "source": "https://deepgram.com/pricing", + "supported_endpoints": [ + "/v1/listen" + ] + }, + "deepgram/streaming/redact": { + "input_cost_per_second": 3.333e-05, + "litellm_provider": "deepgram", + "metadata": { + "calculation": "$0.0020/60 seconds = $0.00003333 per second", + "note": "Redaction add-on (redact query param), streaming, pay as you go", + "original_pricing_per_minute": 0.002 + }, + "mode": "audio_transcription", + "output_cost_per_second": 0.0, + "source": "https://deepgram.com/pricing", + "supported_endpoints": [ + "/v1/listen" + ] + }, + "deepgram/streaming/keyterm": { + "input_cost_per_second": 2.167e-05, + "litellm_provider": "deepgram", + "metadata": { + "calculation": "$0.0013/60 seconds = $0.00002167 per second", + "note": "Keyterm Prompting add-on (keyterm query param), streaming, pay as you go", + "original_pricing_per_minute": 0.0013 + }, + "mode": "audio_transcription", + "output_cost_per_second": 0.0, + "source": "https://deepgram.com/pricing", + "supported_endpoints": [ + "/v1/listen" + ] + }, + "deepgram/streaming/detect_entities": { + "input_cost_per_second": 2.833e-05, + "litellm_provider": "deepgram", + "metadata": { + "calculation": "$0.0017/60 seconds = $0.00002833 per second", + "note": "Entity Detection add-on (detect_entities query param), streaming, pay as you go", + "original_pricing_per_minute": 0.0017 + }, + "mode": "audio_transcription", + "output_cost_per_second": 0.0, + "source": "https://deepgram.com/pricing", + "supported_endpoints": [ + "/v1/listen" + ] + }, + "deepgram/streaming/diarize": { + "input_cost_per_second": 3.333e-05, + "litellm_provider": "deepgram", + "metadata": { + "calculation": "$0.0020/60 seconds = $0.00003333 per second", + "note": "Speaker Diarization add-on (diarize / diarize_model query params), streaming, pay as you go", + "original_pricing_per_minute": 0.002 + }, + "mode": "audio_transcription", + "output_cost_per_second": 0.0, + "source": "https://deepgram.com/pricing", + "supported_endpoints": [ + "/v1/listen" + ] + }, "deepgram/whisper": { "input_cost_per_second": 0.0001, "litellm_provider": "deepgram", diff --git a/tests/test_litellm/llms/deepgram/test_deepgram_common_utils.py b/tests/test_litellm/llms/deepgram/test_deepgram_common_utils.py new file mode 100644 index 00000000000..530888b70c4 --- /dev/null +++ b/tests/test_litellm/llms/deepgram/test_deepgram_common_utils.py @@ -0,0 +1,326 @@ +import math +from collections.abc import Mapping, Sequence +from typing import Final +from urllib.parse import parse_qs, urlparse + +import pytest + +import litellm +from litellm.llms.deepgram.common_utils import ( + deepgram_listen_addon_pricing_models, + deepgram_listen_audio_seconds, + deepgram_listen_callback_params, + deepgram_listen_channel_count, + deepgram_listen_is_priced, + deepgram_listen_model, + deepgram_listen_pricing_model, + deepgram_listen_registry_key, + deepgram_listen_requested_model, + deepgram_listen_transcript, + deepgram_listen_websocket_target, +) + +NOVA_3_URL: Final = "wss://api.deepgram.com/v1/listen?model=nova-3&encoding=linear16&sample_rate=16000" + + +def _results( + start: object, + duration: object, + transcript: str = "", + is_final: object = True, + channel_index: object = (0, 1), +) -> dict[str, object]: + return { + "type": "Results", + "start": start, + "duration": duration, + "is_final": is_final, + "channel_index": list(channel_index) if isinstance(channel_index, tuple) else channel_index, + "channel": {"alternatives": [{"transcript": transcript, "confidence": 0.9}]}, + } + + +def _metadata(duration: object, channels: object = 1) -> dict[str, object]: + return {"type": "Metadata", "request_id": "req-1", "duration": duration, "channels": channels} + + +@pytest.mark.parametrize( + ("api_base", "query_string", "expected"), + [ + pytest.param( + None, + "model=nova-3&encoding=linear16", + "wss://api.deepgram.com/v1/listen?model=nova-3&encoding=linear16", + id="default", + ), + pytest.param( + None, + "encoding=linear16&sample_rate=16000", + "wss://api.deepgram.com/v1/listen?encoding=linear16&sample_rate=16000&model=nova-3", + id="model added when missing", + ), + pytest.param( + None, + "model=&encoding=linear16", + "wss://api.deepgram.com/v1/listen?encoding=linear16&model=nova-3", + id="empty model replaced", + ), + pytest.param( + "http://localhost:9000/v1/", + "model=nova-2", + "ws://localhost:9000/v1/listen?model=nova-2", + id="custom base becomes ws", + ), + pytest.param( + "wss://dg.internal/v1", + "model=nova-3&keywords=a&keywords=b", + "wss://dg.internal/v1/listen?model=nova-3&keywords=a&keywords=b", + id="repeated keys preserved", + ), + pytest.param( + None, + "model=nova-2&encoding=linear16&model=nova-3", + "wss://api.deepgram.com/v1/listen?model=nova-2&encoding=linear16", + id="only the authorized first model reaches deepgram", + ), + pytest.param( + None, + "language=en&model=nova-3&language=multi", + "wss://api.deepgram.com/v1/listen?language=en&model=nova-3", + id="only the priced first language reaches deepgram", + ), + pytest.param( + None, + "model=&model=nova-2", + "wss://api.deepgram.com/v1/listen?model=nova-3", + id="blank first model is the default, later models dropped", + ), + ], +) +def test_deepgram_listen_websocket_target(api_base: str | None, query_string: str, expected: str): + assert deepgram_listen_websocket_target(api_base=api_base, query_string=query_string) == expected + + +@pytest.mark.parametrize( + ("query_string", "expected"), + [ + pytest.param("model=nova-3&encoding=linear16", (), id="no callback"), + pytest.param("model=nova-3&callback=https%3A%2F%2Fevil.example%2Fsink", ("callback",), id="callback"), + pytest.param( + "callback_method=put&model=nova-3&callback=wss%3A%2F%2Fevil.example", + ("callback", "callback_method"), + id="callback and method", + ), + pytest.param("model=nova-3&callback_method=put", ("callback_method",), id="method alone"), + pytest.param("model=nova-3&callbacks=x&my_callback=y", (), id="only exact names match"), + ], +) +def test_deepgram_listen_callback_params(query_string: str, expected: tuple[str, ...]): + assert deepgram_listen_callback_params(query_string) == expected + + +@pytest.mark.parametrize( + ("frames", "expected_seconds"), + [ + pytest.param((_results(0.0, 2.0), _results(2.0, 3.5), _metadata(6.25)), 6.25, id="metadata wins"), + pytest.param((_metadata(4.0), _results(0.0, 9.0), _metadata(5.5)), 5.5, id="last metadata wins"), + pytest.param((_results(0.0, 2.0), _results(2.0, 3.5), _results(1.0, 1.0)), 5.5, id="furthest results end"), + pytest.param((_results(0.0, 0.0), _metadata(0.0)), 0.0, id="zero metadata is a real zero"), + pytest.param( + (_metadata(0.0), _results(0.0, 2.0), _results(2.0, 3.5)), + 5.5, + id="handshake metadata zero does not hide streamed results", + ), + pytest.param((_metadata(0.0), _results(0.0, 2.0), _metadata(0.0)), 2.0, id="only zero metadata frames"), + pytest.param((_results(0.0, 1.5), _metadata("6.25")), 1.5, id="string metadata is ignored"), + pytest.param((_results(0.0, 1.5), _metadata(True)), 1.5, id="boolean metadata is ignored"), + pytest.param((_results(0.0, 1.5), _metadata(-3.0)), 1.5, id="negative metadata is ignored"), + pytest.param((_results(0.0, 1.5), _metadata(math.nan), _metadata(math.inf)), 1.5, id="nan/inf ignored"), + pytest.param((_results("0", 2.0), _results(0.0, None), _results(0.0, 0.75)), 0.75, id="malformed results"), + pytest.param(({"type": "SpeechStarted", "timestamp": 3.0}, {"type": "UtteranceEnd"}), 0.0, id="no usage"), + pytest.param((), 0.0, id="no frames"), + ], +) +def test_deepgram_listen_audio_seconds(frames: Sequence[Mapping[str, object]], expected_seconds: float): + assert deepgram_listen_audio_seconds(frames) == expected_seconds + + +@pytest.mark.parametrize( + ("frames", "upstream_url", "expected_channels"), + [ + pytest.param((_results(0.0, 2.0), _metadata(6.25)), NOVA_3_URL, 1, id="mono"), + pytest.param((_results(0.0, 2.0, channel_index=(0, 2)), _metadata(6.25, 2)), NOVA_3_URL, 2, id="stereo"), + pytest.param((_metadata(1.0, 3), _metadata(1.0, 5)), NOVA_3_URL, 5, id="last metadata wins"), + pytest.param( + (_metadata(1.0, 20), _results(0.0, 1.0, channel_index=(1, 2))), + NOVA_3_URL, + 20, + id="metadata beats channel_index", + ), + pytest.param( + (_results(0.0, 1.0, channel_index=(0, 2)), _results(0.0, 1.0, channel_index=(3, 4))), + NOVA_3_URL, + 4, + id="widest channel_index without metadata", + ), + pytest.param( + (_results(0.0, 1.0, channel_index=(0, 2)),), + f"{NOVA_3_URL}&channels=7&multichannel=true", + 2, + id="frames beat the declared query", + ), + pytest.param((), f"{NOVA_3_URL}&channels=7&multichannel=true", 7, id="declared query when no frames"), + pytest.param((), f"{NOVA_3_URL}&channels=0", 1, id="zero declared channels"), + pytest.param((), f"{NOVA_3_URL}&channels=-2", 1, id="negative declared channels"), + pytest.param((), f"{NOVA_3_URL}&channels=two", 1, id="non numeric declared channels"), + pytest.param((), NOVA_3_URL, 1, id="nothing declared"), + pytest.param((_metadata(1.0, "2"), _metadata(1.0, True), _metadata(1.0, 0)), NOVA_3_URL, 1, id="bad metadata"), + pytest.param((_metadata(1.0, 3), _metadata(1.0, True)), NOVA_3_URL, 3, id="boolean does not shadow a count"), + pytest.param((_metadata(1.0, 2.0), _metadata(1.0, -1)), NOVA_3_URL, 1, id="float and negative metadata"), + pytest.param( + (_metadata(1.0, 2), {**_results(0.0, 1.0), "channels": 9}, {"type": "UtteranceEnd", "channels": 11}), + NOVA_3_URL, + 2, + id="channels on non metadata frames ignored", + ), + pytest.param( + ( + _results(0.0, 1.0, channel_index=[0]), + _results(0.0, 1.0, channel_index=(0, "2")), + _results(0.0, 1.0, channel_index=(0, 0)), + ), + NOVA_3_URL, + 1, + id="bad channel_index", + ), + ], +) +def test_deepgram_listen_channel_count( + frames: Sequence[Mapping[str, object]], upstream_url: str, expected_channels: int +): + assert deepgram_listen_channel_count(frames, upstream_url) == expected_channels + + +def test_deepgram_listen_transcript_joins_final_results_only(): + frames = ( + _results(0.0, 1.0, "hello wor", is_final=False), + _results(0.0, 1.5, "hello world"), + _results(1.5, 0.5, "", is_final=True), + _results(2.0, 1.0, "how are you", is_final="yes"), + {"type": "Results", "start": 3.0, "duration": 1.0, "is_final": True, "channel": {"alternatives": []}}, + _results(4.0, 1.0, "goodbye"), + _metadata(5.0), + ) + assert deepgram_listen_transcript(frames) == "hello world goodbye" + + +@pytest.mark.parametrize( + ("upstream_url", "expected_model"), + [ + (NOVA_3_URL, "nova-3"), + ("wss://api.deepgram.com/v1/listen?encoding=linear16&model=nova-2-medical", "nova-2-medical"), + ("wss://api.deepgram.com/v1/listen?model=nova-3&model=nova-2", "nova-3"), + ("wss://api.deepgram.com/v1/listen?encoding=linear16", litellm.constants.DEEPGRAM_LISTEN_DEFAULT_MODEL), + ], +) +def test_deepgram_listen_model_comes_from_the_upstream_query(upstream_url: str, expected_model: str): + assert deepgram_listen_model(upstream_url) == expected_model + + +@pytest.mark.parametrize( + "query_string", + [ + "model=nova-2&language=en", + "language=en", + "model=&language=en", + "", + "model=nova-3-medical", + "model=nova-2&model=nova-3", + "model=&model=nova-3-medical", + ], +) +def test_requested_model_is_the_only_model_the_upstream_target_carries(query_string: str): + """Authorization runs against ``deepgram_listen_requested_model``; the upstream URL is built separately, so the + two must always agree or a key could be authorized for one model and reach another. Deepgram reads the last + repeated ``model``, so the target must carry exactly one.""" + target: Final = deepgram_listen_websocket_target(None, query_string) + assert parse_qs(urlparse(target).query)["model"] == [deepgram_listen_requested_model(query_string)] + assert deepgram_listen_requested_model(query_string) == deepgram_listen_model(target) + + +@pytest.mark.parametrize( + ("upstream_url", "expected"), + [ + pytest.param(NOVA_3_URL, "streaming/nova-3", id="monolingual"), + pytest.param(f"{NOVA_3_URL}&language=en", "streaming/nova-3", id="explicit language"), + pytest.param(f"{NOVA_3_URL}&language=multi", "streaming/nova-3-multilingual", id="multilingual"), + pytest.param(f"{NOVA_3_URL}&language=MULTI", "streaming/nova-3-multilingual", id="multilingual any case"), + pytest.param( + "wss://api.deepgram.com/v1/listen?model=nova-2&language=multi", + "streaming/nova-2-multilingual", + id="other model", + ), + pytest.param("wss://api.deepgram.com/v1/listen?encoding=linear16", "streaming/nova-3", id="default model"), + ], +) +def test_deepgram_listen_pricing_model_is_the_streaming_entry_never_the_prerecorded_one( + upstream_url: str, expected: str +): + assert deepgram_listen_pricing_model(upstream_url) == expected + assert deepgram_listen_registry_key(upstream_url) == f"deepgram/{expected}" + + +NOVA_2_URL: Final = "wss://api.deepgram.com/v1/listen?model=nova-2" + + +@pytest.mark.usefixtures("local_model_cost_map") +@pytest.mark.parametrize( + ("upstream_url", "extra_rows", "expected"), + [ + pytest.param(NOVA_3_URL, (), True, id="streaming entry present"), + pytest.param(f"{NOVA_3_URL}&language=multi", (), True, id="multilingual entry present"), + pytest.param(NOVA_2_URL, (), False, id="only the pre-recorded entry"), + pytest.param(f"{NOVA_2_URL}&language=multi", ("deepgram/streaming/nova-2",), False, id="needs multilingual"), + pytest.param("wss://api.deepgram.com/v1/listen?model=nova-99-unmapped", (), False, id="nothing priced"), + pytest.param(NOVA_2_URL, ("deepgram/streaming/nova-2",), True, id="operator-supplied streaming entry"), + pytest.param(NOVA_2_URL, ("streaming/nova-2",), False, id="a row under another key is not the entry"), + ], +) +def test_deepgram_listen_is_priced( + monkeypatch: pytest.MonkeyPatch, upstream_url: str, extra_rows: tuple[str, ...], expected: bool +): + """The bundled map prices only nova-3 for streaming; nova-2 has a pre-recorded row, which must never count.""" + monkeypatch.delitem(litellm.model_cost, "deepgram/streaming/nova-2", raising=False) + assert "deepgram/nova-2" in litellm.model_cost + for row in extra_rows: + monkeypatch.setitem(litellm.model_cost, row, dict(litellm.model_cost["deepgram/streaming/nova-3"])) + + assert deepgram_listen_is_priced(upstream_url) is expected + + +@pytest.mark.parametrize( + ("upstream_url", "expected"), + [ + pytest.param(NOVA_3_URL, (), id="no add-ons"), + pytest.param(f"{NOVA_3_URL}&redact=pci", ("streaming/redact",), id="redact"), + pytest.param(f"{NOVA_3_URL}&redact=pci&redact=ssn", ("streaming/redact",), id="repeated redact once"), + pytest.param(f"{NOVA_3_URL}&keyterm=a&keyterm=b", ("streaming/keyterm",), id="keyterm"), + pytest.param(f"{NOVA_3_URL}&detect_entities=true", ("streaming/detect_entities",), id="detect_entities"), + pytest.param(f"{NOVA_3_URL}&diarize=true", ("streaming/diarize",), id="diarize"), + pytest.param(f"{NOVA_3_URL}&diarize_model=v1", ("streaming/diarize",), id="diarize_model"), + pytest.param(f"{NOVA_3_URL}&diarize=true&diarize_model=latest", ("streaming/diarize",), id="diarize both once"), + pytest.param(f"{NOVA_3_URL}&detect_entities=false&diarize=FALSE&redact=", (), id="disabled"), + pytest.param( + f"{NOVA_3_URL}&detect_entities=false&detect_entities=true", + ("streaming/detect_entities",), + id="any enabling value wins", + ), + pytest.param( + f"{NOVA_3_URL}&diarize=true&redact=pci&keyterm=x&detect_entities=true", + ("streaming/detect_entities", "streaming/diarize", "streaming/keyterm", "streaming/redact"), + id="all, sorted", + ), + ], +) +def test_deepgram_listen_addon_pricing_models(upstream_url: str, expected: tuple[str, ...]): + assert deepgram_listen_addon_pricing_models(upstream_url) == expected diff --git a/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_deepgram_listen_passthrough_logging_handler.py b/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_deepgram_listen_passthrough_logging_handler.py new file mode 100644 index 00000000000..ac742c0ab46 --- /dev/null +++ b/tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_deepgram_listen_passthrough_logging_handler.py @@ -0,0 +1,302 @@ +"""Deepgram ``/v1/listen`` WebSocket passthrough: duration extraction and duration based cost tracking.""" + +from datetime import datetime +from types import SimpleNamespace +from typing import Final + +import pytest + +import litellm +from litellm.integrations.custom_logger import CustomLogger +from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj +from litellm.proxy._types import UserAPIKeyAuth +from litellm.proxy.pass_through_endpoints.llm_provider_handlers.deepgram_listen_passthrough_logging_handler import ( + DeepgramListenPassthroughLoggingHandler, +) +from litellm.proxy.pass_through_endpoints.success_handler import PassThroughEndpointLogging +from litellm.types.passthrough_endpoints.pass_through_endpoints import PassthroughStandardLoggingPayload +from litellm.types.utils import StandardLoggingPayload, TranscriptionResponse + +NOVA_3_URL: Final = "wss://api.deepgram.com/v1/listen?model=nova-3&encoding=linear16&sample_rate=16000" + +pytestmark: Final = pytest.mark.usefixtures("local_model_cost_map") + + +def _results(start: object, duration: object, transcript: str = "", is_final: object = True) -> dict[str, object]: + return { + "type": "Results", + "start": start, + "duration": duration, + "is_final": is_final, + "channel": {"alternatives": [{"transcript": transcript, "confidence": 0.9}]}, + } + + +def _metadata(duration: object, channels: int = 1) -> dict[str, object]: + return {"type": "Metadata", "request_id": "req-1", "duration": duration, "channels": channels} + + +@pytest.mark.parametrize( + ("url_route", "expected"), + [ + ("/deepgram/v1/listen", True), + ("/deepgram/listen", True), + ("/deepgram/v1/listen?model=nova-3", True), + ("/litellm/deepgram/v1/listen", True), + ("/deepgram/v1/speak", False), + ("/deepgram/v1/listen/extra", False), + ("/openai/v1/realtime", False), + ("/vertex_ai/live", False), + ("", False), + ], +) +def test_is_deepgram_listen_route(url_route: str, expected: bool): + assert DeepgramListenPassthroughLoggingHandler.is_deepgram_listen_route(url_route) is expected + + +def _logging_obj(call_id: str = "call-dg") -> LiteLLMLoggingObj: + return LiteLLMLoggingObj( + model="unknown", + messages=[{"role": "user", "content": "WebSocket connection"}], + stream=True, + call_type="pass_through_endpoint", + start_time=datetime.now(), + litellm_call_id=call_id, + function_id="websocket_passthrough", + ) + + +def _registry_cost(pricing_model: str, seconds: float) -> float: + """Derives the expected charge from the live cost map rather than pinning a vendor price.""" + per_second: Final = litellm.model_cost[f"deepgram/{pricing_model}"]["input_cost_per_second"] + assert per_second > 0 + return per_second * seconds + + +def _cost(upstream_url: str, *frames: dict[str, object]) -> float: + handler_result = DeepgramListenPassthroughLoggingHandler().deepgram_listen_passthrough_handler( + websocket_messages=frames, logging_obj=_logging_obj(), upstream_url=upstream_url + ) + response_cost = handler_result["kwargs"]["response_cost"] + assert isinstance(response_cost, float) + return response_cost + + +def test_handler_bills_metadata_duration_at_the_registry_rate_and_names_the_model(): + frames = (_results(0.0, 5.0, "first sentence"), _results(5.0, 7.5, "second sentence"), _metadata(12.5)) + logging_obj = _logging_obj() + + handler_result = DeepgramListenPassthroughLoggingHandler().deepgram_listen_passthrough_handler( + websocket_messages=frames, + logging_obj=logging_obj, + upstream_url=NOVA_3_URL, + kwargs={"litellm_params": {"metadata": {}}}, + ) + + result = handler_result["result"] + assert isinstance(result, TranscriptionResponse) + assert result.text == "first sentence second sentence" + assert result._hidden_params["audio_transcription_duration"] == 12.5 + assert result._hidden_params["response_cost"] == pytest.approx(_registry_cost("streaming/nova-3", 12.5)) + assert handler_result["kwargs"]["response_cost"] == pytest.approx(_registry_cost("streaming/nova-3", 12.5)) + assert handler_result["kwargs"]["model"] == "nova-3" + assert handler_result["kwargs"]["custom_llm_provider"] == "deepgram" + assert handler_result["kwargs"]["litellm_params"] == {"metadata": {}} + assert logging_obj.model == "nova-3" + assert logging_obj.model_call_details["model"] == "nova-3" + assert logging_obj.model_call_details["custom_llm_provider"] == "deepgram" + assert logging_obj.model_call_details["response_cost"] == pytest.approx(_registry_cost("streaming/nova-3", 12.5)) + + +def test_handler_bills_streaming_not_prerecorded_rates(): + """Deepgram prices /v1/listen over a WebSocket separately from pre-recorded transcription, so the streaming entry + must be the one charged; the two registry rows only need to differ for this to matter, whatever their values.""" + streaming = litellm.model_cost["deepgram/streaming/nova-3"]["input_cost_per_second"] + prerecorded = litellm.model_cost["deepgram/nova-3"]["input_cost_per_second"] + assert streaming != prerecorded + + assert _cost(NOVA_3_URL, _metadata(60.0)) == pytest.approx(60.0 * streaming) + + +def test_handler_bills_multilingual_streaming_when_language_is_multi(): + monolingual = _cost(NOVA_3_URL, _metadata(60.0)) + multilingual = _cost(f"{NOVA_3_URL}&language=multi", _metadata(60.0)) + + assert multilingual == pytest.approx(_registry_cost("streaming/nova-3-multilingual", 60.0)) + assert multilingual > monolingual + + +@pytest.mark.parametrize( + ("query", "addons"), + [ + pytest.param("redact=pci", ("redact",), id="redaction"), + pytest.param("redact=pci&redact=numbers", ("redact",), id="redaction counted once"), + pytest.param("keyterm=LiteLLM&keyterm=Deepgram", ("keyterm",), id="keyterm prompting"), + pytest.param("detect_entities=true", ("detect_entities",), id="entity detection"), + pytest.param("diarize=true", ("diarize",), id="diarization"), + pytest.param("diarize_model=v1", ("diarize",), id="diarization via diarize_model"), + pytest.param("diarize=true&diarize_model=v1", ("diarize",), id="diarization counted once"), + pytest.param( + "redact=pci&keyterm=x&detect_entities=true&diarize=true", + ("redact", "keyterm", "detect_entities", "diarize"), + id="every add-on", + ), + pytest.param("detect_entities=false&diarize=False&redact=", (), id="disabled add-ons cost nothing"), + ], +) +def test_handler_adds_each_priced_add_on_once_on_top_of_the_base_rate(query: str, addons: tuple[str, ...]): + base = _cost(NOVA_3_URL, _metadata(60.0)) + expected = base + sum(_registry_cost(f"streaming/{addon}", 60.0) for addon in addons) + + assert _cost(f"{NOVA_3_URL}&{query}", _metadata(60.0)) == pytest.approx(expected) + + +def test_handler_add_ons_scale_with_channels_like_the_base_rate(): + stereo_plain = _cost(f"{NOVA_3_URL}&channels=2", _metadata(60.0, channels=2)) + stereo_redacted = _cost(f"{NOVA_3_URL}&channels=2&redact=pci", _metadata(60.0, channels=2)) + + assert stereo_redacted - stereo_plain == pytest.approx(_registry_cost("streaming/redact", 120.0)) + + +@pytest.mark.parametrize( + "upstream_url", + [ + pytest.param("wss://api.deepgram.com/v1/listen?model=nova-2", id="only a pre-recorded entry"), + pytest.param("wss://api.deepgram.com/v1/listen?model=nova-99-not-in-registry", id="no entry at all"), + ], +) +def test_handler_never_substitutes_another_rate_for_a_missing_streaming_entry(monkeypatch, upstream_url): + """The route refuses these sessions up front; should the registry change under a live one, the spend row + keeps the duration and carries no cost, rather than the pre-recorded rate or any other stand-in.""" + monkeypatch.delitem(litellm.model_cost, "deepgram/streaming/nova-2", raising=False) + assert "deepgram/nova-2" in litellm.model_cost + + handler_result = DeepgramListenPassthroughLoggingHandler().deepgram_listen_passthrough_handler( + websocket_messages=(_metadata(60.0),), logging_obj=_logging_obj(), upstream_url=upstream_url + ) + + assert handler_result["kwargs"]["response_cost"] is None + assert handler_result["result"]._hidden_params["audio_transcription_duration"] == 60.0 + + +def test_handler_falls_back_to_results_frames_when_the_stream_ends_without_metadata(): + frames = (_results(0.0, 30.0, "a"), _results(30.0, 30.0, "b"), _results(60.0, 12.5, "c")) + + handler_result = DeepgramListenPassthroughLoggingHandler().deepgram_listen_passthrough_handler( + websocket_messages=frames, logging_obj=_logging_obj(), upstream_url=NOVA_3_URL + ) + + assert handler_result["result"]._hidden_params["audio_transcription_duration"] == 72.5 + assert handler_result["kwargs"]["response_cost"] == pytest.approx(_registry_cost("streaming/nova-3", 72.5)) + + +def test_handler_charges_more_for_more_audio_on_the_same_model(): + short = DeepgramListenPassthroughLoggingHandler().deepgram_listen_passthrough_handler( + websocket_messages=(_metadata(10.0),), logging_obj=_logging_obj(), upstream_url=NOVA_3_URL + ) + long = DeepgramListenPassthroughLoggingHandler().deepgram_listen_passthrough_handler( + websocket_messages=(_metadata(30.0),), logging_obj=_logging_obj(), upstream_url=NOVA_3_URL + ) + + assert long["kwargs"]["response_cost"] == pytest.approx(3 * short["kwargs"]["response_cost"]) + assert short["kwargs"]["response_cost"] > 0 + + +def test_handler_bills_every_channel_of_a_multichannel_session(): + """Deepgram bills processed audio per channel (deepgram.com/pricing FAQ, 2026-09-17), so a stereo session must be + charged for twice its wall-clock duration or budgets can be bypassed by requesting more channels.""" + mono = DeepgramListenPassthroughLoggingHandler().deepgram_listen_passthrough_handler( + websocket_messages=(_metadata(30.0),), logging_obj=_logging_obj(), upstream_url=NOVA_3_URL + ) + stereo = DeepgramListenPassthroughLoggingHandler().deepgram_listen_passthrough_handler( + websocket_messages=(_metadata(30.0, channels=2),), + logging_obj=_logging_obj(), + upstream_url=f"{NOVA_3_URL}&multichannel=true&channels=2", + ) + + assert stereo["result"]._hidden_params["audio_transcription_duration"] == 60.0 + assert stereo["kwargs"]["response_cost"] == pytest.approx(2 * mono["kwargs"]["response_cost"]) + assert stereo["kwargs"]["response_cost"] == pytest.approx(_registry_cost("streaming/nova-3", 60.0)) + + +def test_handler_bills_the_declared_channels_when_the_stream_dies_before_any_frame_reports_them(): + handler_result = DeepgramListenPassthroughLoggingHandler().deepgram_listen_passthrough_handler( + websocket_messages=(_results(0.0, 10.0, "a"),), + logging_obj=_logging_obj(), + upstream_url=f"{NOVA_3_URL}&multichannel=true&channels=3", + ) + + assert handler_result["result"]._hidden_params["audio_transcription_duration"] == 30.0 + assert handler_result["kwargs"]["response_cost"] == pytest.approx(_registry_cost("streaming/nova-3", 30.0)) + + +class _CapturingLogger(CustomLogger): + def __init__(self) -> None: + super().__init__() + self.payloads: list[StandardLoggingPayload] = [] + + async def async_log_success_event(self, kwargs, response_obj, start_time, end_time): + self.payloads.append(kwargs["standard_logging_object"]) + + +@pytest.mark.asyncio +async def test_success_handler_dispatches_deepgram_listen_and_logs_duration_based_spend(monkeypatch): + """Drives the shared passthrough success handler the way the WebSocket relay does at socket close and reads + what a spend logger receives: Deepgram model and provider, the audio duration billed at the registry rate.""" + capturing_logger = _CapturingLogger() + monkeypatch.setattr(litellm, "_async_success_callback", [capturing_logger]) + monkeypatch.setattr(litellm, "success_callback", []) + monkeypatch.setattr(litellm, "callbacks", []) + logging_obj = _logging_obj("call-dg-e2e") + frames = [_results(0.0, 5.0, "hello world", is_final=False), _results(0.0, 5.0, "hello world"), _metadata(20.0)] + user_api_key_dict = UserAPIKeyAuth(api_key="hashed-key", team_id="team-stt", user_id="user-1") + start_time = datetime.now() + passthrough_logging_payload = PassthroughStandardLoggingPayload( + url=NOVA_3_URL, request_body={}, request_method="WEBSOCKET", cost_per_request=None + ) + logging_obj.update_environment_variables( + model="unknown", + user="unknown", + optional_params={}, + litellm_params={ + "metadata": { + "user_api_key": user_api_key_dict.api_key, + "user_api_key_team_id": user_api_key_dict.team_id, + "user_api_key_user_id": user_api_key_dict.user_id, + } + }, + call_type="pass_through_endpoint", + ) + + await PassThroughEndpointLogging().pass_through_async_success_handler( + httpx_response=SimpleNamespace( + status_code=200, + text="WebSocket connection successful", + headers={}, + request=SimpleNamespace(method="WEBSOCKET", url=NOVA_3_URL), + ), + response_body=frames, + logging_obj=logging_obj, + url_route="/deepgram/v1/listen", + result="websocket_connection_successful", + start_time=start_time, + end_time=datetime.now(), + cache_hit=False, + request_body={}, + passthrough_logging_payload=passthrough_logging_payload, + litellm_params={ + "metadata": { + "user_api_key": user_api_key_dict.api_key, + "user_api_key_team_id": user_api_key_dict.team_id, + "user_api_key_user_id": user_api_key_dict.user_id, + } + }, + ) + + assert len(capturing_logger.payloads) == 1 + payload = capturing_logger.payloads[0] + assert payload["model"] == "nova-3" + assert payload["custom_llm_provider"] == "deepgram" + assert payload["response_cost"] == pytest.approx(_registry_cost("streaming/nova-3", 20.0)) + assert payload["metadata"]["user_api_key_team_id"] == "team-stt" + assert payload["id"] == "call-dg-e2e" diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_deepgram_ws_passthrough_routes.py b/tests/test_litellm/proxy/pass_through_endpoints/test_deepgram_ws_passthrough_routes.py new file mode 100644 index 00000000000..44533f35c72 --- /dev/null +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_deepgram_ws_passthrough_routes.py @@ -0,0 +1,474 @@ +"""Deepgram ``/v1/listen`` passthrough WebSocket route: registration, auth, credential injection, target URL.""" + +import asyncio +from collections.abc import Mapping +from dataclasses import dataclass +from types import MappingProxyType, SimpleNamespace +from typing import Final +from unittest.mock import AsyncMock, MagicMock, patch + +import pytest +from fastapi import FastAPI +from fastapi.testclient import TestClient +from starlette.routing import WebSocketRoute +from starlette.websockets import WebSocketDisconnect + +import litellm +from litellm.caching.dual_cache import DualCache +from litellm.proxy._lazy_features import LAZY_FEATURES +from litellm.proxy._types import LiteLLMRoutes, UserAPIKeyAuth +from litellm.proxy.auth.auth_checks import _cache_key_object +from litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints import ( + _websocket_relay, + deepgram_listen_websocket_route, + router, +) +from litellm.proxy.utils import hash_token + +GET_CREDENTIALS: Final = ( + "litellm.proxy.pass_through_endpoints.llm_passthrough_endpoints.passthrough_endpoint_router.get_credentials" +) +USER_API_KEY_AUTH: Final = "litellm.proxy.auth.user_api_key_auth.user_api_key_auth" +LISTEN_PATHS: Final = ("/deepgram/v1/listen", "/deepgram/listen") +NOVA_2_STREAMING_KEY: Final = "deepgram/streaming/nova-2" + +pytestmark: Final = pytest.mark.usefixtures("local_model_cost_map") + + +def _price_nova_2_streaming(monkeypatch: pytest.MonkeyPatch) -> None: + """An operator-supplied streaming row: the bundled map prices only nova-3 for streaming.""" + monkeypatch.setitem(litellm.model_cost, NOVA_2_STREAMING_KEY, dict(litellm.model_cost["deepgram/streaming/nova-3"])) + + +class _FakeWebSocket: + def __init__(self, path: str, query: str) -> None: + self.url = SimpleNamespace(path=path, query=query) + self.headers = {"authorization": "Bearer sk-litellm-virtual", "x-api-key": "sk-caller-secret"} + self.accepts: list[str | None] = [] + self.closed: tuple[int, str] | None = None + + async def accept(self, subprotocol: str | None = None) -> None: + self.accepts.append(subprotocol) + + async def close(self, code: int = 1000, reason: str = "") -> None: + self.closed = (code, reason) + + +@dataclass(frozen=True, slots=True) +class _RelayCall: + target: str + custom_headers: Mapping[str, str] + user_api_key_dict: UserAPIKeyAuth + forward_headers: bool + endpoint: str + accept_websocket: bool + + +class _FakeRelay: + def __init__(self) -> None: + self.calls: list[_RelayCall] = [] + + async def __call__( + self, + *, + websocket: object, + target: str, + custom_headers: dict[str, str], + user_api_key_dict: UserAPIKeyAuth, + forward_headers: bool, + endpoint: str, + accept_websocket: bool, + ) -> None: + self.calls.append( + _RelayCall( + target=target, + custom_headers=MappingProxyType(dict(custom_headers)), + user_api_key_dict=user_api_key_dict, + forward_headers=forward_headers, + endpoint=endpoint, + accept_websocket=accept_websocket, + ) + ) + + +async def _serve(websocket: _FakeWebSocket, user_api_key_dict: UserAPIKeyAuth | None = None) -> _FakeRelay: + relay = _FakeRelay() + await deepgram_listen_websocket_route( + websocket=websocket, + user_api_key_dict=user_api_key_dict or UserAPIKeyAuth(), + relay=relay, + ) + return relay + + +def test_deepgram_listen_websocket_routes_registered(): + ws_paths = {route.path for route in router.routes if isinstance(route, WebSocketRoute)} + assert set(LISTEN_PATHS) <= ws_paths + + +@pytest.mark.parametrize("path", LISTEN_PATHS) +def test_deepgram_listen_is_a_lazily_loaded_mapped_pass_through_route(path): + """The route must be reachable before the passthrough module is imported and must be authed and + billed as a mapped pass-through route like the other provider prefixes.""" + feature = next(feature for feature in LAZY_FEATURES if feature.name == "llm_passthrough") + assert feature.matches(path) + assert any(path.startswith(prefix) for prefix in LiteLLMRoutes.mapped_pass_through_routes.value) + + +@pytest.mark.asyncio +@pytest.mark.parametrize("path", LISTEN_PATHS) +async def test_deepgram_listen_forwards_query_and_injects_only_provider_auth(path, monkeypatch): + monkeypatch.delenv("DEEPGRAM_API_BASE", raising=False) + websocket = _FakeWebSocket(path, "encoding=linear16&sample_rate=16000&keywords=hi%3A2&keywords=there") + caller = UserAPIKeyAuth(api_key="sk-litellm-virtual", team_id="team-stt") + + with patch(GET_CREDENTIALS, return_value="dg-provider-key") as get_credentials: + relay = await _serve(websocket, caller) + + assert get_credentials.call_args.kwargs == {"custom_llm_provider": "deepgram", "region_name": None} + assert relay.calls == [ + _RelayCall( + target=( + "wss://api.deepgram.com/v1/listen" + "?encoding=linear16&sample_rate=16000&keywords=hi%3A2&keywords=there&model=nova-3" + ), + custom_headers=MappingProxyType({"Authorization": "Token dg-provider-key"}), + user_api_key_dict=caller, + forward_headers=False, + endpoint=path, + accept_websocket=False, + ) + ] + assert websocket.accepts == [None] + assert websocket.closed is None + + +@pytest.mark.asyncio +async def test_deepgram_listen_keeps_caller_chosen_model(monkeypatch): + monkeypatch.delenv("DEEPGRAM_API_BASE", raising=False) + _price_nova_2_streaming(monkeypatch) + websocket = _FakeWebSocket("/deepgram/v1/listen", "model=nova-2&language=en") + + with patch(GET_CREDENTIALS, return_value="dg-provider-key"): + relay = await _serve(websocket) + + assert [call.target for call in relay.calls] == ["wss://api.deepgram.com/v1/listen?model=nova-2&language=en"] + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + ("query", "expected_target"), + [ + ("", "wss://api.deepgram.com/v1/listen?model=nova-3"), + ("model=", "wss://api.deepgram.com/v1/listen?model=nova-3"), + ("model=&language=en", "wss://api.deepgram.com/v1/listen?language=en&model=nova-3"), + ], +) +async def test_deepgram_listen_defaults_to_nova_3_when_no_model_is_named(query, expected_target, monkeypatch): + monkeypatch.delenv("DEEPGRAM_API_BASE", raising=False) + websocket = _FakeWebSocket("/deepgram/listen", query) + + with patch(GET_CREDENTIALS, return_value="dg-provider-key"): + relay = await _serve(websocket) + + assert [call.target for call in relay.calls] == [expected_target] + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + ("api_base", "expected_target"), + [ + ("https://api.eu.deepgram.com/v1/", "wss://api.eu.deepgram.com/v1/listen?model=nova-3"), + ("http://localhost:8080/v1", "ws://localhost:8080/v1/listen?model=nova-3"), + ("wss://deepgram.internal.example/v1", "wss://deepgram.internal.example/v1/listen?model=nova-3"), + ], +) +async def test_deepgram_listen_honours_server_configured_api_base(api_base, expected_target, monkeypatch): + monkeypatch.setenv("DEEPGRAM_API_BASE", api_base) + websocket = _FakeWebSocket("/deepgram/v1/listen", "") + + with patch(GET_CREDENTIALS, return_value="dg-provider-key"): + relay = await _serve(websocket) + + assert [call.target for call in relay.calls] == [expected_target] + + +@pytest.mark.asyncio +async def test_deepgram_listen_ignores_caller_supplied_api_base(monkeypatch): + """V1: the server-configured Deepgram key must only ever go to the server-configured host.""" + monkeypatch.delenv("DEEPGRAM_API_BASE", raising=False) + websocket = _FakeWebSocket("/deepgram/v1/listen", "api_base=wss%3A%2F%2Fattacker.example%2Fv1&model=nova-3") + + with patch(GET_CREDENTIALS, return_value="dg-provider-key"): + relay = await _serve(websocket) + + assert [call.target for call in relay.calls] == [ + "wss://api.deepgram.com/v1/listen?api_base=wss%3A%2F%2Fattacker.example%2Fv1&model=nova-3" + ] + + +@pytest.mark.asyncio +async def test_deepgram_listen_closes_cleanly_when_provider_credentials_missing(): + websocket = _FakeWebSocket("/deepgram/v1/listen", "model=nova-3") + + with patch(GET_CREDENTIALS, return_value=None): + relay = await _serve(websocket) + + assert websocket.closed is not None + assert websocket.closed[0] == 1011 + assert "DEEPGRAM_API_KEY" in websocket.closed[1] + assert websocket.accepts == [] + assert relay.calls == [] + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + "query", + [ + pytest.param("model=nova-3&callback=https%3A%2F%2Fsink.example%2Fdg", id="http callback"), + pytest.param("callback=wss%3A%2F%2Fsink.example&callback_method=put&model=nova-3", id="ws callback"), + ], +) +async def test_deepgram_listen_rejects_callback_delivery_that_would_go_unbilled(query, monkeypatch): + """With ``callback`` set, Deepgram sends every Results and Metadata frame to the caller's URL and only a + request id down this socket, so the proxy would meter zero seconds of audio; refuse before contacting Deepgram.""" + monkeypatch.delenv("DEEPGRAM_API_BASE", raising=False) + websocket = _FakeWebSocket("/deepgram/v1/listen", query) + + with patch(GET_CREDENTIALS, return_value="dg-provider-key"): + relay = await _serve(websocket) + + assert relay.calls == [] + assert websocket.closed is not None + assert websocket.closed[0] == 1008 + assert "callback" in websocket.closed[1] + assert "dg-provider-key" not in websocket.closed[1] + + +@pytest.mark.asyncio +@pytest.mark.parametrize( + ("query", "missing_key"), + [ + pytest.param("model=nova-2", "deepgram/streaming/nova-2", id="model with only a pre-recorded price"), + pytest.param("model=nova-99", "deepgram/streaming/nova-99", id="model unknown to the registry"), + pytest.param( + "model=nova-3&language=multi", + "deepgram/streaming/nova-3-multilingual", + id="multilingual session without its own price", + ), + ], +) +async def test_deepgram_listen_refuses_sessions_it_cannot_price(query, missing_key, monkeypatch): + """A session with no streaming price would be logged at zero (or at the pre-recorded rate), letting a caller run + up unmetered spend, so the proxy closes it before Deepgram is contacted and names the registry row to add.""" + monkeypatch.delenv("DEEPGRAM_API_BASE", raising=False) + monkeypatch.delitem(litellm.model_cost, missing_key, raising=False) + assert "deepgram/nova-2" in litellm.model_cost + websocket = _FakeWebSocket("/deepgram/v1/listen", query) + + with patch(GET_CREDENTIALS, return_value="dg-provider-key"): + relay = await _serve(websocket) + + assert relay.calls == [] + assert websocket.closed is not None + assert websocket.closed[0] == 1008 + assert missing_key in websocket.closed[1] + assert "dg-provider-key" not in websocket.closed[1] + + +@pytest.mark.asyncio +async def test_deepgram_listen_relays_once_the_operator_prices_the_model(monkeypatch): + monkeypatch.delenv("DEEPGRAM_API_BASE", raising=False) + websocket = _FakeWebSocket("/deepgram/v1/listen", "model=nova-2") + with patch(GET_CREDENTIALS, return_value="dg-provider-key"): + assert (await _serve(websocket)).calls == [] + + _price_nova_2_streaming(monkeypatch) + priced_websocket = _FakeWebSocket("/deepgram/v1/listen", "model=nova-2") + with patch(GET_CREDENTIALS, return_value="dg-provider-key"): + relay = await _serve(priced_websocket) + + assert [call.target for call in relay.calls] == ["wss://api.deepgram.com/v1/listen?model=nova-2"] + assert priced_websocket.closed is None + + +def _app_with_relay(relay: _FakeRelay) -> FastAPI: + app = FastAPI() + app.include_router(router) + app.dependency_overrides[_websocket_relay] = lambda: relay + return app + + +def test_deepgram_listen_rejects_connections_without_a_litellm_key(): + relay = _FakeRelay() + client = TestClient(_app_with_relay(relay)) + + with patch(GET_CREDENTIALS, return_value="dg-provider-key") as get_credentials: + with pytest.raises(WebSocketDisconnect) as disconnect: + with client.websocket_connect("/deepgram/v1/listen?model=nova-3"): + pass + + assert disconnect.value.code == 1008 + assert relay.calls == [] + get_credentials.assert_not_called() + + +def test_deepgram_listen_callback_rejection_reaches_the_client_as_a_policy_close(monkeypatch): + monkeypatch.delenv("DEEPGRAM_API_BASE", raising=False) + relay = _FakeRelay() + client = TestClient(_app_with_relay(relay)) + + with ( + patch(GET_CREDENTIALS, return_value="dg-provider-key"), + patch(USER_API_KEY_AUTH, new=AsyncMock(return_value=UserAPIKeyAuth(api_key="hashed"))), + ): + with pytest.raises(WebSocketDisconnect) as disconnect: + with client.websocket_connect( + "/deepgram/v1/listen?model=nova-3&callback=https%3A%2F%2Fsink.example%2Fdg", + headers={"Authorization": "Bearer sk-litellm-virtual"}, + ) as connection: + connection.receive_text() + + assert disconnect.value.code == 1008 + assert "callback" in disconnect.value.reason + assert relay.calls == [] + + +def test_deepgram_listen_authenticates_the_litellm_key_and_relays_to_deepgram(monkeypatch): + monkeypatch.delenv("DEEPGRAM_API_BASE", raising=False) + relay = _FakeRelay() + client = TestClient(_app_with_relay(relay)) + caller = UserAPIKeyAuth(api_key="hashed-sk-litellm", team_id="team-stt") + + with ( + patch(GET_CREDENTIALS, return_value="dg-provider-key"), + patch(USER_API_KEY_AUTH, new=AsyncMock(return_value=caller)) as auth, + ): + with client.websocket_connect( + "/deepgram/v1/listen?model=nova-3&punctuate=true", + headers={"Authorization": "Bearer sk-litellm-virtual"}, + ): + pass + + assert auth.await_args.kwargs["api_key"] == "Bearer sk-litellm-virtual" + assert relay.calls == [ + _RelayCall( + target="wss://api.deepgram.com/v1/listen?model=nova-3&punctuate=true", + custom_headers=MappingProxyType({"Authorization": "Token dg-provider-key"}), + user_api_key_dict=caller, + forward_headers=False, + endpoint="/deepgram/v1/listen", + accept_websocket=False, + ) + ] + + +async def _cache_restricted_key(virtual_key: str, models: list[str]) -> DualCache: + cache = DualCache() + await _cache_key_object( + hashed_token=hash_token(virtual_key), + user_api_key_obj=UserAPIKeyAuth(token=hash_token(virtual_key), models=models), + user_api_key_cache=cache, + proxy_logging_obj=None, + ) + return cache + + +@pytest.mark.parametrize( + ("query", "expect_relay"), + [ + pytest.param("model=nova-2", True, id="allowed model named"), + pytest.param("model=nova-3", False, id="denied model named"), + pytest.param("", False, id="model omitted, default denied"), + pytest.param("model=&language=en", False, id="model blank, default denied"), + ], +) +def test_deepgram_listen_authorizes_the_model_it_will_actually_send_upstream(query, expect_relay, monkeypatch): + """A key allowed only ``nova-2`` must not reach ``nova-3`` by leaving ``model`` out and letting the proxy fill + in its default: the real key auth path must see the same model the upstream target will carry.""" + monkeypatch.delenv("DEEPGRAM_API_BASE", raising=False) + monkeypatch.setattr(litellm, "max_budget", 0.0) + _price_nova_2_streaming(monkeypatch) + cache = asyncio.run(_cache_restricted_key("sk-only-nova-2", ["nova-2"])) + relay = _FakeRelay() + client = TestClient(_app_with_relay(relay)) + + with ( + patch(GET_CREDENTIALS, return_value="dg-provider-key"), + patch.multiple( # test-quality-ok: the real key auth path reads these proxy_server globals and has no injection seam + "litellm.proxy.proxy_server", + master_key="sk-master", + prisma_client=MagicMock(), + user_api_key_cache=cache, + llm_model_list=None, + llm_router=None, + ), + ): + if expect_relay: + with client.websocket_connect( + f"/deepgram/v1/listen?{query}", headers={"Authorization": "Bearer sk-only-nova-2"} + ): + pass + assert [call.target for call in relay.calls] == [f"wss://api.deepgram.com/v1/listen?{query}"] + return + with pytest.raises(WebSocketDisconnect) as disconnect: + with client.websocket_connect( + f"/deepgram/v1/listen?{query}", headers={"Authorization": "Bearer sk-only-nova-2"} + ): + pass + + assert disconnect.value.code == 1008 + assert relay.calls == [] + + +def test_deepgram_listen_strips_a_second_model_that_would_outrank_the_authorized_one(monkeypatch): + """Deepgram honours the last repeated ``model``; auth and pricing read the first. A key allowed only ``nova-2`` + must not smuggle ``nova-3`` past authorization behind an authorized first value.""" + monkeypatch.delenv("DEEPGRAM_API_BASE", raising=False) + monkeypatch.setattr(litellm, "max_budget", 0.0) + _price_nova_2_streaming(monkeypatch) + cache = asyncio.run(_cache_restricted_key("sk-only-nova-2", ["nova-2"])) + relay = _FakeRelay() + client = TestClient(_app_with_relay(relay)) + + with ( + patch(GET_CREDENTIALS, return_value="dg-provider-key"), + patch.multiple( # test-quality-ok: the real key auth path reads these proxy_server globals and has no injection seam + "litellm.proxy.proxy_server", + master_key="sk-master", + prisma_client=MagicMock(), + user_api_key_cache=cache, + llm_model_list=None, + llm_router=None, + ), + ): + with client.websocket_connect( + "/deepgram/v1/listen?model=nova-2&language=en&model=nova-3&language=multi", + headers={"Authorization": "Bearer sk-only-nova-2"}, + ): + pass + + assert [call.target for call in relay.calls] == ["wss://api.deepgram.com/v1/listen?model=nova-2&language=en"] + + +def test_deepgram_listen_echoes_the_browser_subprotocol_that_carries_the_litellm_key(monkeypatch): + """Browsers cannot set headers, so they send the key as a subprotocol and abort the handshake unless the + server echoes that subprotocol back; the key itself must still stay off the upstream connection.""" + monkeypatch.delenv("DEEPGRAM_API_BASE", raising=False) + relay = _FakeRelay() + client = TestClient(_app_with_relay(relay)) + + with ( + patch(GET_CREDENTIALS, return_value="dg-provider-key"), + patch(USER_API_KEY_AUTH, new=AsyncMock(return_value=UserAPIKeyAuth(api_key="hashed"))), + ): + with client.websocket_connect( + "/deepgram/v1/listen?model=nova-3", + subprotocols=["openai-insecure-api-key.sk-litellm-virtual"], + ) as connection: + assert connection.accepted_subprotocol == "openai-insecure-api-key.sk-litellm-virtual" + + assert [call.custom_headers for call in relay.calls] == [ + MappingProxyType({"Authorization": "Token dg-provider-key"}) + ] + assert [call.forward_headers for call in relay.calls] == [False] diff --git a/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py b/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py index 701ff6db3bb..be090a309fc 100644 --- a/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py +++ b/tests/test_litellm/proxy/pass_through_endpoints/test_pass_through_endpoints.py @@ -4942,18 +4942,21 @@ async def test_unusable_upstream_cost_records_zero_not_the_flat_estimate(): class FakeUpstreamWebSocket: - def __init__(self, first_frame: bytes): - self._first_frame = first_frame + """Serves the given frames in order, then closes normally, the way a real websockets connection does""" + + def __init__(self, *frames: str | bytes): + self._frames = iter(frames) self.close = AsyncMock() + self.send = AsyncMock() - async def recv(self, decode: bool = True): - return self._first_frame + async def recv(self, decode: bool | None = None): + from websockets.exceptions import ConnectionClosedOK + from websockets.frames import Close - def __aiter__(self): - return self - - async def __anext__(self): - raise StopAsyncIteration + frame = next(self._frames, None) + if frame is None: + raise ConnectionClosedOK(rcvd=Close(1000, ""), sent=Close(1000, ""), rcvd_then_sent=True) + return frame class FakeUpstreamConnect: @@ -4974,7 +4977,7 @@ async def test_websocket_passthrough_forwards_non_ascii_first_frame(): first_frame = json.dumps( {"type": "session.created", "session": {"instructions": "Hablas español, ¿sí?"}}, ensure_ascii=False, - ).encode("utf-8") + ) upstream_ws = FakeUpstreamWebSocket(first_frame) websocket = MagicMock() @@ -5028,7 +5031,7 @@ async def test_websocket_passthrough_propagates_active_trace_context( from starlette.websockets import WebSocketState captured: dict[str, dict[str, str]] = {} - upstream_ws = FakeUpstreamWebSocket(b"{}") + upstream_ws = FakeUpstreamWebSocket("{}") def fake_connect(target, additional_headers): captured["headers"] = additional_headers @@ -5457,6 +5460,144 @@ async def test_websocket_passthrough_does_not_close_twice_when_success_logging_f websocket.close.assert_awaited_once_with(code=1008, reason=upstream_reason) +DEEPGRAM_LISTEN_TARGET = "wss://api.deepgram.com/v1/listen?model=nova-3&encoding=linear16&sample_rate=16000" +DEEPGRAM_INTERIM_FRAME = json.dumps( + { + "type": "Results", + "start": 0.0, + "duration": 1.02, + "is_final": False, + "channel": {"alternatives": [{"transcript": "hello wor", "confidence": 0.71}]}, + } +) +DEEPGRAM_FINAL_FRAME = json.dumps( + { + "type": "Results", + "start": 0.0, + "duration": 2.5, + "is_final": True, + "speech_final": True, + "channel": {"alternatives": [{"transcript": "hello world, ¿qué tal?", "confidence": 0.98}]}, + }, + ensure_ascii=False, +) +DEEPGRAM_METADATA_FRAME = json.dumps({"type": "Metadata", "request_id": "req-1", "duration": 2.5, "channels": 1}) + + +async def _relay_deepgram_listen(upstream_ws, client_receive): + """Runs the generic relay the way the Deepgram route does and returns (client websocket, success handler mock)""" + websocket = _client_websocket(client_receive) + with ( + _patched_websocket_passthrough_environment(upstream_ws), + patch( # test-quality-ok: pass_through_endpoint_logging is a module global read inside websocket_passthrough_request; there is no injection seam + "litellm.proxy.pass_through_endpoints.pass_through_endpoints." + "pass_through_endpoint_logging.pass_through_async_success_handler", + new=AsyncMock(), + ) as success_handler, + ): + await websocket_passthrough_request( + websocket=websocket, + target=DEEPGRAM_LISTEN_TARGET, + custom_headers={"Authorization": "Token dg-provider-key"}, + user_api_key_dict=UserAPIKeyAuth(), + forward_headers=False, + endpoint="/deepgram/v1/listen", + accept_websocket=False, + ) + return websocket, success_handler + + +@pytest.mark.asyncio +async def test_websocket_passthrough_relays_deepgram_transcript_frames_verbatim_and_keeps_them_for_billing(): + """Interim, final and Metadata frames reach the client byte for byte (no JSON round trip, non-ASCII intact, + a binary frame first) and every JSON object frame is what the success handler gets to bill from.""" + upstream_ws = FakeUpstreamWebSocket( + b"\x00\x01binary-first", + DEEPGRAM_INTERIM_FRAME, + "not json at all", + DEEPGRAM_FINAL_FRAME, + DEEPGRAM_METADATA_FRAME, + ) + + websocket, success_handler = await _relay_deepgram_listen(upstream_ws, _pending_receive) + + assert [call.args[0] for call in websocket.send_bytes.await_args_list] == [b"\x00\x01binary-first"] + assert [call.args[0] for call in websocket.send_text.await_args_list] == [ + DEEPGRAM_INTERIM_FRAME, + "not json at all", + DEEPGRAM_FINAL_FRAME, + DEEPGRAM_METADATA_FRAME, + ] + success_call = success_handler.call_args.kwargs + assert success_call["url_route"] == "/deepgram/v1/listen" + assert success_call["response_body"] == [ + json.loads(DEEPGRAM_INTERIM_FRAME), + json.loads(DEEPGRAM_FINAL_FRAME), + json.loads(DEEPGRAM_METADATA_FRAME), + ] + assert success_call["httpx_response"].request.url == DEEPGRAM_LISTEN_TARGET + assert success_call["logging_obj"].model_call_details.get("custom_llm_provider") is None + websocket.close.assert_awaited_once_with() + + +@pytest.mark.asyncio +async def test_websocket_passthrough_sends_deepgram_audio_bytes_and_control_text_upstream_unchanged(): + upstream_ws = RecordingUpstreamWebSocket() + audio_chunk = bytes(range(256)) * 4 + close_stream = json.dumps({"type": "CloseStream"}) + + await _relay_deepgram_listen( + upstream_ws, + AsyncMock( + side_effect=[ + {"type": "websocket.receive", "bytes": audio_chunk}, + {"type": "websocket.receive", "text": close_stream}, + {"type": "websocket.disconnect"}, + ] + ), + ) + + assert [call.args[0] for call in upstream_ws.send.await_args_list] == [audio_chunk, close_stream] + assert isinstance(upstream_ws.send.await_args_list[0].args[0], bytes) + upstream_ws.close.assert_awaited_once() + + +@pytest.mark.asyncio +async def test_websocket_passthrough_vertex_live_setup_ack_names_the_model_but_is_not_billed_as_usage(): + """Vertex Live keeps its special first frame: the setup acknowledgement is forwarded verbatim, read for the + model, and left out of the frames the usage handler sees; later frames are kept as before.""" + setup_ack = json.dumps( + {"setupComplete": {}, "model": "projects/p/locations/global/publishers/google/models/gemini-live-2.5-flash"} + ) + server_content = json.dumps({"serverContent": {"turnComplete": True}, "usageMetadata": {"totalTokenCount": 12}}) + upstream_ws = FakeUpstreamWebSocket(setup_ack, server_content) + websocket = _client_websocket(_pending_receive) + + with ( + _patched_websocket_passthrough_environment(upstream_ws), + patch( # test-quality-ok: pass_through_endpoint_logging is a module global read inside websocket_passthrough_request; there is no injection seam + "litellm.proxy.pass_through_endpoints.pass_through_endpoints." + "pass_through_endpoint_logging.pass_through_async_success_handler", + new=AsyncMock(), + ) as success_handler, + ): + await websocket_passthrough_request( + websocket=websocket, + target="wss://aiplatform.googleapis.com/ws/google.cloud.aiplatform.v1.LlmBidiService/BidiGenerateContent", + custom_headers={"Authorization": "Bearer token"}, + user_api_key_dict=UserAPIKeyAuth(), + forward_headers=False, + endpoint="/vertex_ai/live", + accept_websocket=False, + ) + + assert [call.args[0] for call in websocket.send_text.await_args_list] == [setup_ack, server_content] + success_call = success_handler.call_args.kwargs + assert success_call["response_body"] == [json.loads(server_content)] + assert success_call["logging_obj"].model == "gemini-live-2.5-flash" + assert success_call["logging_obj"].model_call_details["custom_llm_provider"] == "vertex_ai_language_models" + + def _passthrough_kwargs_for_reservation( user_api_key_dict: UserAPIKeyAuth, parsed_body: dict | None = None, diff --git a/tests/test_litellm/proxy/test_proxy_server.py b/tests/test_litellm/proxy/test_proxy_server.py index 139fe2cce15..f42954f8dbc 100644 --- a/tests/test_litellm/proxy/test_proxy_server.py +++ b/tests/test_litellm/proxy/test_proxy_server.py @@ -7607,8 +7607,9 @@ async def test_deleting_the_stored_pass_through_row_takes_the_route_out_of_servi settings: Final = patch("litellm.proxy.proxy_server.general_settings", {}) # test-quality-ok: the method reads this module global; no injection seam yaml_endpoints: Final = patch("litellm.proxy.proxy_server.config_passthrough_endpoints", None) # test-quality-ok: module global holding the YAML endpoints; this case has none + app_routes: Final = patch("litellm.proxy.pass_through_endpoints.pass_through_endpoints.SafeRouteAdder.add_api_route_if_not_exists") # test-quality-ok: the registry is the observable; a real route would stay on the shared FastAPI app for the rest of the xdist worker try: - with settings, yaml_endpoints: + with settings, yaml_endpoints, app_routes: pc = ProxyConfig() await pc._update_general_settings(db_general_settings={"pass_through_endpoints": [db_endpoint]}) assert live_routes(), "the stored endpoint should be serving before the row is deleted" @@ -7648,8 +7649,9 @@ async def test_a_stored_pass_through_row_never_disturbs_the_config_declared_rout settings: Final = patch("litellm.proxy.proxy_server.general_settings", {"pass_through_endpoints": [config_endpoint]}) # test-quality-ok: the method reads this module global; no injection seam yaml_endpoints: Final = patch("litellm.proxy.proxy_server.config_passthrough_endpoints", [config_endpoint]) # test-quality-ok: module global holding the YAML endpoints the reload merges in + app_routes: Final = patch("litellm.proxy.pass_through_endpoints.pass_through_endpoints.SafeRouteAdder.add_api_route_if_not_exists") # test-quality-ok: the registry is the observable; a real route would stay on the shared FastAPI app for the rest of the xdist worker try: - with settings, yaml_endpoints: + with settings, yaml_endpoints, app_routes: await initialize_pass_through_endpoints(pass_through_endpoints=[config_endpoint]) assert live_paths() == {config_path} diff --git a/tests/test_litellm/test_utils.py b/tests/test_litellm/test_utils.py index dc6a6ad8cef..5adfb2aea4c 100644 --- a/tests/test_litellm/test_utils.py +++ b/tests/test_litellm/test_utils.py @@ -903,6 +903,7 @@ def test_aaamodel_prices_and_context_window_json_is_valid(): "/v1/audio/speech", "/v1/ocr", "/vertex_ai/live", + "/v1/listen", "/v1beta/interactions", ], },