From e22744c4393858357d34b2abfffb0249528880ba Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 29 Aug 2026 11:06:29 -0700 Subject: [PATCH 1/5] fix(embeddings): omit encoding_format when the client omits it on OpenAI-compatible calls When no encoding_format is set on the call, the model config, or LITELLM_DEFAULT_EMBEDDING_ENCODING_FORMAT, leave the field out of the upstream request instead of defaulting to float, and bypass the OpenAI SDK's own base64 default so nothing re-adds it on the wire. Downstreams that reject encoding_format, such as a second LiteLLM proxy fronting Bedrock Titan embeddings, now work when the client omits the field. Fixes #38661 --- litellm/llms/hosted_vllm/embedding/README.md | 5 +- litellm/llms/openai/openai.py | 72 ++++++++++++-------- litellm/main.py | 19 +++--- tests/test_litellm/test_main.py | 61 +++++++++++++++++ 4 files changed, 115 insertions(+), 42 deletions(-) diff --git a/litellm/llms/hosted_vllm/embedding/README.md b/litellm/llms/hosted_vllm/embedding/README.md index 2c58e16fc23..32b7ea5c560 100644 --- a/litellm/llms/hosted_vllm/embedding/README.md +++ b/litellm/llms/hosted_vllm/embedding/README.md @@ -9,8 +9,7 @@ For OpenAI-compatible embedding calls (including `openai/...` with a custom `api 1. Explicit value on the embedding call (`encoding_format=...`). 2. Model config (`litellm_params.encoding_format` on the proxy `model_list` entry). 3. Environment variable `LITELLM_DEFAULT_EMBEDDING_ENCODING_FORMAT` (e.g. in `.env` or container env). -4. Default **`float`**. -That avoids forwarding `encoding_format=None` to the provider/SDK where some servers behave poorly. +If none of those is set, or the winning value is the literal string `none`, the field is omitted from the upstream request entirely (LiteLLM also bypasses the OpenAI SDK's own base64 default), so OpenAI-compatible servers that reject `encoding_format` keep working. -To pass provider-specific parameters, see [provider-specific params](https://docs.litellm.ai/docs/completion/provider_specific_params). \ No newline at end of file +To pass provider-specific parameters, see [provider-specific params](https://docs.litellm.ai/docs/completion/provider_specific_params). diff --git a/litellm/llms/openai/openai.py b/litellm/llms/openai/openai.py index 6e66c998acf..7b71e532b61 100644 --- a/litellm/llms/openai/openai.py +++ b/litellm/llms/openai/openai.py @@ -12,9 +12,14 @@ if TYPE_CHECKING: import openai from openai import AsyncOpenAI, OpenAI +from openai._base_client import make_request_options +from openai._constants import RAW_RESPONSE_HEADER +from openai._legacy_response import LegacyAPIResponse +from openai._types import RequestOptions +from openai.types import CreateEmbeddingResponse from openai.types.beta.assistant_deleted import AssistantDeleted from openai.types.file_deleted import FileDeleted -from pydantic import BaseModel +from pydantic import BaseModel, TypeAdapter from typing_extensions import overload import litellm @@ -322,6 +327,24 @@ class OpenAIChatCompletionResponseIterator(BaseModelResponseIterator): raise e +_EXTRA_HEADERS_ADAPTER: Final = TypeAdapter(dict[str, str] | None) +_EXTRA_QUERY_ADAPTER: Final = TypeAdapter(dict[str, object] | None) + + +def _embedding_request_without_sdk_defaults( + data: Mapping[str, object], timeout: float | httpx.Timeout +) -> tuple[dict[str, object], RequestOptions]: + body: Final = {k: v for k, v in data.items() if k not in ("extra_headers", "extra_query", "extra_body")} + extra_headers: Final = _EXTRA_HEADERS_ADAPTER.validate_python(data.get("extra_headers")) + options: Final = make_request_options( + extra_headers={**(extra_headers or {}), RAW_RESPONSE_HEADER: "true"}, + extra_query=_EXTRA_QUERY_ADAPTER.validate_python(data.get("extra_query")), + extra_body=data.get("extra_body"), + timeout=timeout, + ) + return body, options + + class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM): def __init__(self) -> None: super().__init__() @@ -1148,19 +1171,16 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM): data: dict, timeout: float | httpx.Timeout, logging_obj: LiteLLMLoggingObj, - ): - """ - Helper to: - - call embeddings.create.with_raw_response when litellm.return_response_headers is True - - call embeddings.create by default - """ - try: - raw_response = await openai_aclient.embeddings.with_raw_response.create(**data, timeout=timeout) - headers: Final = dict(raw_response.headers) - response: Final = raw_response.parse() - return headers, response - except Exception as e: - raise e + ) -> tuple[dict[str, str], CreateEmbeddingResponse]: + if "encoding_format" not in data: + body, options = _embedding_request_without_sdk_defaults(data, timeout) + bypass_response: Final = await openai_aclient.post( + "/embeddings", body=body, options=options, cast_to=CreateEmbeddingResponse + ) + assert isinstance(bypass_response, LegacyAPIResponse) + return dict(bypass_response.headers), bypass_response.parse(to=CreateEmbeddingResponse) + raw_response: Final = await openai_aclient.embeddings.with_raw_response.create(**data, timeout=timeout) + return dict(raw_response.headers), raw_response.parse() @track_llm_api_timing() def make_sync_openai_embedding_request( @@ -1169,20 +1189,16 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM): data: dict, timeout: float | httpx.Timeout, logging_obj: LiteLLMLoggingObj, - ): - """ - Helper to: - - call embeddings.create.with_raw_response when litellm.return_response_headers is True - - call embeddings.create by default - """ - try: - raw_response = openai_client.embeddings.with_raw_response.create(**data, timeout=timeout) - - headers: Final = dict(raw_response.headers) - response: Final = raw_response.parse() - return headers, response - except Exception as e: - raise e + ) -> tuple[dict[str, str], CreateEmbeddingResponse]: + if "encoding_format" not in data: + body, options = _embedding_request_without_sdk_defaults(data, timeout) + bypass_response: Final = openai_client.post( + "/embeddings", body=body, options=options, cast_to=CreateEmbeddingResponse + ) + assert isinstance(bypass_response, LegacyAPIResponse) + return dict(bypass_response.headers), bypass_response.parse(to=CreateEmbeddingResponse) + raw_response: Final = openai_client.embeddings.with_raw_response.create(**data, timeout=timeout) + return dict(raw_response.headers), raw_response.parse() async def aembedding( self, diff --git a/litellm/main.py b/litellm/main.py index cafa1e4718f..f40b0e5c227 100644 --- a/litellm/main.py +++ b/litellm/main.py @@ -6289,18 +6289,15 @@ def embedding( if headers is not None and headers != {}: optional_params["extra_headers"] = headers - if encoding_format is not None: - optional_params["encoding_format"] = encoding_format + requested_encoding_format: Final = ( + encoding_format + or optional_params.get("encoding_format") + or get_secret_str("LITELLM_DEFAULT_EMBEDDING_ENCODING_FORMAT") + ) + if requested_encoding_format is None or requested_encoding_format.strip().lower() == "none": + optional_params.pop("encoding_format", None) else: - env_fmt: Final = get_secret_str("LITELLM_DEFAULT_EMBEDDING_ENCODING_FORMAT") - if env_fmt is not None and env_fmt.strip().lower() == "none": - optional_params.pop("encoding_format", None) - else: - _default_fmt: Final = optional_params.get("encoding_format") or env_fmt or "float" - if _default_fmt.strip().lower() == "none": - optional_params.pop("encoding_format", None) - else: - optional_params["encoding_format"] = _default_fmt + optional_params["encoding_format"] = requested_encoding_format api_version = None diff --git a/tests/test_litellm/test_main.py b/tests/test_litellm/test_main.py index 8cf878d05d9..4c591be0f61 100644 --- a/tests/test_litellm/test_main.py +++ b/tests/test_litellm/test_main.py @@ -3181,3 +3181,64 @@ def test_stream_chunk_builder_defers_cost_to_logging_obj_when_usage_cost_absent( assert response is not None assert response._hidden_params.get("response_cost") is None + + +def _mock_openai_embedding_route(respx_mock: respx.MockRouter) -> respx.Route: + return respx_mock.post("https://api.openai.com/v1/embeddings").mock( + return_value=httpx.Response( + 200, + json={ + "object": "list", + "data": [{"object": "embedding", "index": 0, "embedding": [0.1, 0.2, 0.3]}], + "model": "text-embedding-3-small", + "usage": {"prompt_tokens": 2, "total_tokens": 2}, + }, + ) + ) + + +def test_embedding_openai_omits_encoding_format_when_client_omits_it(respx_mock: respx.MockRouter) -> None: + mock_route: Final = _mock_openai_embedding_route(respx_mock) + + response: Final = litellm.embedding(model="openai/text-embedding-3-small", input=["hello"], api_key="sk-test") + + request_body: Final = json.loads(mock_route.calls.last.request.read()) + assert "encoding_format" not in request_body + assert response.data[0]["embedding"] == [0.1, 0.2, 0.3] + + +def test_embedding_openai_forwards_explicit_encoding_format(respx_mock: respx.MockRouter) -> None: + mock_route: Final = _mock_openai_embedding_route(respx_mock) + + litellm.embedding( + model="openai/text-embedding-3-small", input=["hello"], api_key="sk-test", encoding_format="base64" + ) + + request_body: Final = json.loads(mock_route.calls.last.request.read()) + assert request_body["encoding_format"] == "base64" + + +def test_embedding_openai_env_var_sets_default_encoding_format( + respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch +) -> None: + monkeypatch.setenv("LITELLM_DEFAULT_EMBEDDING_ENCODING_FORMAT", "float") + mock_route: Final = _mock_openai_embedding_route(respx_mock) + + litellm.embedding(model="openai/text-embedding-3-small", input=["hello"], api_key="sk-test") + + request_body: Final = json.loads(mock_route.calls.last.request.read()) + assert request_body["encoding_format"] == "float" + + +@pytest.mark.asyncio +async def test_aembedding_openai_omits_encoding_format_when_client_omits_it( + respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch +) -> None: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + mock_route: Final = _mock_openai_embedding_route(respx_mock) + + response: Final = await litellm.aembedding(model="openai/text-embedding-3-small", input=["hello"], api_key="sk-test") + + request_body: Final = json.loads(mock_route.calls.last.request.read()) + assert "encoding_format" not in request_body + assert response.data[0]["embedding"] == [0.1, 0.2, 0.3] From c254605e924455062e72305fdfaa2bd9c9feb089 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 29 Aug 2026 11:10:21 -0700 Subject: [PATCH 2/5] test(embeddings): move encoding_format default coverage to wire-level assertions Consolidate the new regression tests into test_openai_embedding_encoding_format_default.py, replacing mocks that pinned the old float default with respx captures of the request body, and update the stale local_testing default-float test to assert omission --- tests/local_testing/test_embedding.py | 64 +++--- tests/test_litellm/test_main.py | 60 ------ ...penai_embedding_encoding_format_default.py | 188 ++++++++---------- 3 files changed, 109 insertions(+), 203 deletions(-) diff --git a/tests/local_testing/test_embedding.py b/tests/local_testing/test_embedding.py index aed2849f056..ee2ac14f498 100644 --- a/tests/local_testing/test_embedding.py +++ b/tests/local_testing/test_embedding.py @@ -3,6 +3,8 @@ import os import re import traceback +import httpx + import openai import pytest from dotenv import load_dotenv @@ -1255,56 +1257,42 @@ def test_jina_ai_img_embeddings(input_data, expected_payload_input): assert sent_data["input"] == expected_payload_input -def test_encoding_format_defaults_to_float_for_openai_sdk(monkeypatch): +def test_encoding_format_omitted_by_default_for_openai_sdk(monkeypatch): """ - When encoding_format is not provided, LiteLLM sends `float` for OpenAI-path embeddings. + When encoding_format is not provided, LiteLLM leaves it out of the upstream request. Optional global override: `LITELLM_DEFAULT_EMBEDDING_ENCODING_FORMAT`. """ monkeypatch.delenv("LITELLM_DEFAULT_EMBEDDING_ENCODING_FORMAT", raising=False) - with patch( - "litellm.llms.openai.openai.OpenAIChatCompletion._get_openai_client" - ) as mock_get_client: - # Create a mock client instance - mock_client_instance = MagicMock() - mock_get_client.return_value = mock_client_instance + captured_bodies = [] - # Mock the embeddings.with_raw_response.create method - mock_response = MagicMock() - mock_response.parse.return_value = MagicMock( - model_dump=lambda: { - "data": [{"embedding": [0.1, 0.2, 0.3], "index": 0}], - "model": "text-embedding-ada-002", + def handler(request: httpx.Request) -> httpx.Response: + captured_bodies.append(json.loads(request.content)) + return httpx.Response( + 200, + json={ "object": "list", + "data": [{"object": "embedding", "index": 0, "embedding": [0.1, 0.2, 0.3]}], + "model": "text-embedding-ada-002", "usage": {"prompt_tokens": 1, "total_tokens": 1}, - } - ) - mock_response.headers = {} - - mock_client_instance.embeddings.with_raw_response.create.return_value = ( - mock_response + }, ) - # Call the embedding function without encoding_format - response = embedding( - model="text-embedding-ada-002", - input="Hello world", - ) + client = openai.OpenAI( + api_key="sk-test", http_client=httpx.Client(transport=httpx.MockTransport(handler)) + ) - # Get the call arguments to verify what was sent to OpenAI SDK - call_args = mock_client_instance.embeddings.with_raw_response.create.call_args - assert ( - call_args is not None - ), "OpenAI SDK embeddings.create should have been called" + response = embedding( + model="text-embedding-ada-002", + input="Hello world", + api_key="sk-test", + client=client, + ) - call_kwargs = call_args[1] # Get kwargs - - assert "encoding_format" in call_kwargs - assert ( - call_kwargs["encoding_format"] == "float" - ), "encoding_format should default to float when not provided by user" - - print("✅ PASS: encoding_format='float' is correctly passed to OpenAI SDK") + assert response.data[0]["embedding"] == [0.1, 0.2, 0.3] + assert "encoding_format" not in captured_bodies[0], ( + "encoding_format should be omitted from the upstream request when not provided by user" + ) def test_encoding_format_explicit_value_preserved(): diff --git a/tests/test_litellm/test_main.py b/tests/test_litellm/test_main.py index 4c591be0f61..84225c3feb1 100644 --- a/tests/test_litellm/test_main.py +++ b/tests/test_litellm/test_main.py @@ -3182,63 +3182,3 @@ def test_stream_chunk_builder_defers_cost_to_logging_obj_when_usage_cost_absent( assert response is not None assert response._hidden_params.get("response_cost") is None - -def _mock_openai_embedding_route(respx_mock: respx.MockRouter) -> respx.Route: - return respx_mock.post("https://api.openai.com/v1/embeddings").mock( - return_value=httpx.Response( - 200, - json={ - "object": "list", - "data": [{"object": "embedding", "index": 0, "embedding": [0.1, 0.2, 0.3]}], - "model": "text-embedding-3-small", - "usage": {"prompt_tokens": 2, "total_tokens": 2}, - }, - ) - ) - - -def test_embedding_openai_omits_encoding_format_when_client_omits_it(respx_mock: respx.MockRouter) -> None: - mock_route: Final = _mock_openai_embedding_route(respx_mock) - - response: Final = litellm.embedding(model="openai/text-embedding-3-small", input=["hello"], api_key="sk-test") - - request_body: Final = json.loads(mock_route.calls.last.request.read()) - assert "encoding_format" not in request_body - assert response.data[0]["embedding"] == [0.1, 0.2, 0.3] - - -def test_embedding_openai_forwards_explicit_encoding_format(respx_mock: respx.MockRouter) -> None: - mock_route: Final = _mock_openai_embedding_route(respx_mock) - - litellm.embedding( - model="openai/text-embedding-3-small", input=["hello"], api_key="sk-test", encoding_format="base64" - ) - - request_body: Final = json.loads(mock_route.calls.last.request.read()) - assert request_body["encoding_format"] == "base64" - - -def test_embedding_openai_env_var_sets_default_encoding_format( - respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch -) -> None: - monkeypatch.setenv("LITELLM_DEFAULT_EMBEDDING_ENCODING_FORMAT", "float") - mock_route: Final = _mock_openai_embedding_route(respx_mock) - - litellm.embedding(model="openai/text-embedding-3-small", input=["hello"], api_key="sk-test") - - request_body: Final = json.loads(mock_route.calls.last.request.read()) - assert request_body["encoding_format"] == "float" - - -@pytest.mark.asyncio -async def test_aembedding_openai_omits_encoding_format_when_client_omits_it( - respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch -) -> None: - monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) - mock_route: Final = _mock_openai_embedding_route(respx_mock) - - response: Final = await litellm.aembedding(model="openai/text-embedding-3-small", input=["hello"], api_key="sk-test") - - request_body: Final = json.loads(mock_route.calls.last.request.read()) - assert "encoding_format" not in request_body - assert response.data[0]["embedding"] == [0.1, 0.2, 0.3] diff --git a/tests/test_litellm/test_openai_embedding_encoding_format_default.py b/tests/test_litellm/test_openai_embedding_encoding_format_default.py index 94e4e3c81e5..9842bf30585 100644 --- a/tests/test_litellm/test_openai_embedding_encoding_format_default.py +++ b/tests/test_litellm/test_openai_embedding_encoding_format_default.py @@ -1,124 +1,102 @@ -from unittest.mock import MagicMock, patch +import json +from typing import Final +import httpx import pytest +import respx -from litellm import embedding +import litellm -@pytest.mark.parametrize( - "set_env, env_value, expected", - [ - (False, None, "float"), - (True, "base64", "base64"), - ], -) -def test_openai_embedding_encoding_format_default( - monkeypatch, set_env, env_value, expected -): - monkeypatch.delenv("LITELLM_DEFAULT_EMBEDDING_ENCODING_FORMAT", raising=False) - if set_env: - monkeypatch.setenv("LITELLM_DEFAULT_EMBEDDING_ENCODING_FORMAT", env_value) - - mock_response = MagicMock() - mock_response.parse.return_value = MagicMock( - model_dump=lambda: { - "data": [{"embedding": [0.1, 0.2, 0.3], "index": 0}], - "model": "text-embedding-ada-002", - "object": "list", - "usage": {"prompt_tokens": 1, "total_tokens": 1}, - } +def _mock_openai_embedding_route(respx_mock: respx.MockRouter) -> respx.Route: + return respx_mock.post("https://api.openai.com/v1/embeddings").mock( + return_value=httpx.Response( + 200, + json={ + "object": "list", + "data": [{"object": "embedding", "index": 0, "embedding": [0.1, 0.2, 0.3]}], + "model": "text-embedding-3-small", + "usage": {"prompt_tokens": 2, "total_tokens": 2}, + }, + ) ) - mock_response.headers = {} - with patch( - "litellm.llms.openai.openai.OpenAIChatCompletion._get_openai_client" - ) as mock_get_client: - mock_client_instance = MagicMock() - mock_get_client.return_value = mock_client_instance - mock_client_instance.embeddings.with_raw_response.create.return_value = ( - mock_response - ) - embedding( - model="text-embedding-ada-002", - input="Hello world", - ) +@pytest.fixture(autouse=True) +def clear_default_encoding_format_env(monkeypatch: pytest.MonkeyPatch): + monkeypatch.delenv("LITELLM_DEFAULT_EMBEDDING_ENCODING_FORMAT", raising=False) - call_kwargs = ( - mock_client_instance.embeddings.with_raw_response.create.call_args[1] - ) - assert call_kwargs["encoding_format"] == expected + +def test_embedding_openai_omits_encoding_format_when_client_omits_it(respx_mock: respx.MockRouter) -> None: + mock_route: Final = _mock_openai_embedding_route(respx_mock) + + response: Final = litellm.embedding(model="openai/text-embedding-3-small", input=["hello"], api_key="sk-test") + + request_body: Final = json.loads(mock_route.calls.last.request.read()) + assert "encoding_format" not in request_body + assert response.data[0]["embedding"] == [0.1, 0.2, 0.3] + + +def test_embedding_openai_forwards_explicit_encoding_format(respx_mock: respx.MockRouter) -> None: + mock_route: Final = _mock_openai_embedding_route(respx_mock) + + litellm.embedding( + model="openai/text-embedding-3-small", input=["hello"], api_key="sk-test", encoding_format="base64" + ) + + request_body: Final = json.loads(mock_route.calls.last.request.read()) + assert request_body["encoding_format"] == "base64" + + +def test_embedding_openai_explicit_encoding_format_wins_over_env_var( + respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch +) -> None: + monkeypatch.setenv("LITELLM_DEFAULT_EMBEDDING_ENCODING_FORMAT", "float") + mock_route: Final = _mock_openai_embedding_route(respx_mock) + + litellm.embedding( + model="openai/text-embedding-3-small", input=["hello"], api_key="sk-test", encoding_format="base64" + ) + + request_body: Final = json.loads(mock_route.calls.last.request.read()) + assert request_body["encoding_format"] == "base64" + + +@pytest.mark.parametrize("env_value", ["float", "base64"]) +def test_embedding_openai_env_var_sets_default_encoding_format( + respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch, env_value: str +) -> None: + monkeypatch.setenv("LITELLM_DEFAULT_EMBEDDING_ENCODING_FORMAT", env_value) + mock_route: Final = _mock_openai_embedding_route(respx_mock) + + litellm.embedding(model="openai/text-embedding-3-small", input=["hello"], api_key="sk-test") + + request_body: Final = json.loads(mock_route.calls.last.request.read()) + assert request_body["encoding_format"] == env_value @pytest.mark.parametrize("env_none", ["none", "NONE", " none "]) -def test_openai_embedding_encoding_format_env_none_omits_param( - monkeypatch, env_none -): - """LITELLM_DEFAULT_EMBEDDING_ENCODING_FORMAT=none omits encoding_format (provider default).""" +def test_embedding_openai_env_none_omits_encoding_format( + respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch, env_none: str +) -> None: monkeypatch.setenv("LITELLM_DEFAULT_EMBEDDING_ENCODING_FORMAT", env_none) + mock_route: Final = _mock_openai_embedding_route(respx_mock) - mock_response = MagicMock() - mock_response.parse.return_value = MagicMock( - model_dump=lambda: { - "data": [{"embedding": [0.1, 0.2, 0.3], "index": 0}], - "model": "text-embedding-ada-002", - "object": "list", - "usage": {"prompt_tokens": 1, "total_tokens": 1}, - } - ) - mock_response.headers = {} + litellm.embedding(model="openai/text-embedding-3-small", input=["hello"], api_key="sk-test") - with patch( - "litellm.llms.openai.openai.OpenAIChatCompletion._get_openai_client" - ) as mock_get_client: - mock_client_instance = MagicMock() - mock_get_client.return_value = mock_client_instance - mock_client_instance.embeddings.with_raw_response.create.return_value = ( - mock_response - ) - - embedding( - model="text-embedding-ada-002", - input="Hello world", - ) - - call_kwargs = ( - mock_client_instance.embeddings.with_raw_response.create.call_args[1] - ) - assert "encoding_format" not in call_kwargs + request_body: Final = json.loads(mock_route.calls.last.request.read()) + assert "encoding_format" not in request_body -def test_openai_embedding_encoding_format_explicit_overrides_env(monkeypatch): - """Request `encoding_format` wins over LITELLM_DEFAULT_EMBEDDING_ENCODING_FORMAT.""" - monkeypatch.setenv("LITELLM_DEFAULT_EMBEDDING_ENCODING_FORMAT", "float") +@pytest.mark.asyncio +async def test_aembedding_openai_omits_encoding_format_when_client_omits_it( + respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch +) -> None: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + mock_route: Final = _mock_openai_embedding_route(respx_mock) - mock_response = MagicMock() - mock_response.parse.return_value = MagicMock( - model_dump=lambda: { - "data": [{"embedding": [0.1, 0.2, 0.3], "index": 0}], - "model": "text-embedding-ada-002", - "object": "list", - "usage": {"prompt_tokens": 1, "total_tokens": 1}, - } - ) - mock_response.headers = {} + response: Final = await litellm.aembedding(model="openai/text-embedding-3-small", input=["hello"], api_key="sk-test") - with patch( - "litellm.llms.openai.openai.OpenAIChatCompletion._get_openai_client" - ) as mock_get_client: - mock_client_instance = MagicMock() - mock_get_client.return_value = mock_client_instance - mock_client_instance.embeddings.with_raw_response.create.return_value = ( - mock_response - ) - - embedding( - model="text-embedding-ada-002", - input="Hello world", - encoding_format="base64", - ) - - call_kwargs = ( - mock_client_instance.embeddings.with_raw_response.create.call_args[1] - ) - assert call_kwargs["encoding_format"] == "base64" + request_body: Final = json.loads(mock_route.calls.last.request.read()) + assert "encoding_format" not in request_body + assert response.data[0]["embedding"] == [0.1, 0.2, 0.3] From b65592e623805749827d03b1fcacf03173a5d16d Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 29 Aug 2026 11:18:19 -0700 Subject: [PATCH 3/5] test: drop stray trailing blank line in test_main.py --- tests/test_litellm/test_main.py | 1 - 1 file changed, 1 deletion(-) diff --git a/tests/test_litellm/test_main.py b/tests/test_litellm/test_main.py index 84225c3feb1..8cf878d05d9 100644 --- a/tests/test_litellm/test_main.py +++ b/tests/test_litellm/test_main.py @@ -3181,4 +3181,3 @@ def test_stream_chunk_builder_defers_cost_to_logging_obj_when_usage_cost_absent( assert response is not None assert response._hidden_params.get("response_cost") is None - From 8e687f10047ce7f0454204b94ed08e7bac2151b6 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 29 Aug 2026 12:04:44 -0700 Subject: [PATCH 4/5] test(embeddings): move legacy intercepts to the wire for the omitted-format path The omitted-format path deliberately no longer dispatches through embeddings.create, so four legacy tests now intercept at the transport or client.post instead. Also adds a bypass error-path unit test, rewords a stale comment and a README scope note, and ratchets the lint budgets down. --- basedpyright-code-budget.json | 2 +- litellm/llms/hosted_vllm/embedding/README.md | 2 +- litellm/utils.py | 8 +-- ruff-strict-budget.json | 8 +-- test-quality-budget.json | 4 +- .../test_litellm_proxy_provider.py | 71 +++++++++++-------- tests/llm_translation/test_nvidia_nim.py | 48 ++++++++----- tests/local_testing/test_exceptions.py | 17 +++-- tests/local_testing/test_router.py | 4 +- ...penai_embedding_encoding_format_default.py | 19 +++++ type-discipline-budget.json | 2 +- 11 files changed, 117 insertions(+), 68 deletions(-) diff --git a/basedpyright-code-budget.json b/basedpyright-code-budget.json index 962d1266fd7..ba57fd1278b 100644 --- a/basedpyright-code-budget.json +++ b/basedpyright-code-budget.json @@ -105,7 +105,7 @@ "limit": 109 }, "reportUnknownMemberType": { - "limit": 38721 + "limit": 38720 }, "reportUnknownParameterType": { "limit": 19778 diff --git a/litellm/llms/hosted_vllm/embedding/README.md b/litellm/llms/hosted_vllm/embedding/README.md index 32b7ea5c560..50474aabdeb 100644 --- a/litellm/llms/hosted_vllm/embedding/README.md +++ b/litellm/llms/hosted_vllm/embedding/README.md @@ -4,7 +4,7 @@ VLLM is a superset of OpenAI's `embedding` endpoint. ## `encoding_format` -For OpenAI-compatible embedding calls (including `openai/...` with a custom `api_base` pointing at vLLM), LiteLLM resolves `encoding_format` when it is not set on the request: +For OpenAI-compatible embedding calls (including `openai/...` with a custom `api_base` pointing at vLLM), LiteLLM resolves `encoding_format` when it is not set on the request. `hosted_vllm/...` models use a separate handler that never adds the field on its own, so this resolution applies to the `openai/...`-style routes only: 1. Explicit value on the embedding call (`encoding_format=...`). 2. Model config (`litellm_params.encoding_format` on the proxy `model_list` entry). diff --git a/litellm/utils.py b/litellm/utils.py index fa2226dbf2c..66a07d2e2db 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -3542,10 +3542,10 @@ def get_optional_params_embeddings( non_default_params=non_default_params, optional_params={}, kwargs=kwargs ) elif custom_llm_provider == "vertex_ai" or custom_llm_provider == "gemini": - # OpenAI SDKs (and litellm's own client) send encoding_format="float" - # by default; float lists are exactly what the vertex API returns, so - # the param is a no-op — don't reject the provider default. Other - # values (e.g. "base64") stay on the unsupported-param path below. + # OpenAI SDKs send encoding_format="float" by default; float lists are + # exactly what the vertex API returns, so the param is a no-op — don't + # reject the provider default. Other values (e.g. "base64") stay on + # the unsupported-param path below. if non_default_params.get("encoding_format") == "float": non_default_params.pop("encoding_format") supported_params = get_supported_openai_params( diff --git a/ruff-strict-budget.json b/ruff-strict-budget.json index c60988eccc0..73c69a732fe 100644 --- a/ruff-strict-budget.json +++ b/ruff-strict-budget.json @@ -9,7 +9,7 @@ "limit": 827 }, "ANN201": { - "limit": 2003 + "limit": 2001 }, "ANN202": { "limit": 845 @@ -240,13 +240,13 @@ "limit": 96 }, "TRY201": { - "limit": 405 + "limit": 403 }, "TRY203": { - "limit": 113 + "limit": 111 }, "TRY300": { - "limit": 857 + "limit": 855 }, "UP028": { "limit": 2 diff --git a/test-quality-budget.json b/test-quality-budget.json index ee33eb581d6..d834c581609 100644 --- a/test-quality-budget.json +++ b/test-quality-budget.json @@ -3,7 +3,7 @@ "limit": 733 }, "TQ002": { - "limit": 742 + "limit": 741 }, "TQ003": { "limit": 62 @@ -21,6 +21,6 @@ "limit": 117 }, "TQ008": { - "limit": 11139 + "limit": 11135 } } diff --git a/tests/llm_translation/test_litellm_proxy_provider.py b/tests/llm_translation/test_litellm_proxy_provider.py index 1cb805bf9ba..8630259877d 100644 --- a/tests/llm_translation/test_litellm_proxy_provider.py +++ b/tests/llm_translation/test_litellm_proxy_provider.py @@ -5,6 +5,7 @@ from io import BytesIO from unittest.mock import AsyncMock +import httpx import litellm from litellm import completion, embedding import pytest @@ -92,44 +93,54 @@ async def test_litellm_gateway_from_sdk_embedding(is_async): litellm.set_verbose = True litellm._turn_on_debug() + captured_bodies = [] + + def handler(request: httpx.Request) -> httpx.Response: + captured_bodies.append(json.loads(request.content)) + return httpx.Response( + 200, + json={ + "object": "list", + "data": [{"object": "embedding", "index": 0, "embedding": [0.1, 0.2, 0.3]}], + "model": "my-vllm-model", + "usage": {"prompt_tokens": 2, "total_tokens": 2}, + }, + ) + if is_async: from openai import AsyncOpenAI - openai_client = AsyncOpenAI(api_key="fake-key") - mock_method = AsyncMock() - patch_target = openai_client.embeddings.create + openai_client = AsyncOpenAI( + api_key="fake-key", + http_client=httpx.AsyncClient(transport=httpx.MockTransport(handler)), + ) + response = await litellm.aembedding( + model="litellm_proxy/my-vllm-model", + input="Hello world", + client=openai_client, + api_base="my-custom-api-base", + ) else: from openai import OpenAI - openai_client = OpenAI(api_key="fake-key") - mock_method = MagicMock() - patch_target = openai_client.embeddings.create + openai_client = OpenAI( + api_key="fake-key", + http_client=httpx.Client(transport=httpx.MockTransport(handler)), + ) + response = litellm.embedding( + model="litellm_proxy/my-vllm-model", + input="Hello world", + client=openai_client, + api_base="my-custom-api-base", + ) - with patch.object(patch_target.__self__, patch_target.__name__, new=mock_method): - try: - if is_async: - await litellm.aembedding( - model="litellm_proxy/my-vllm-model", - input="Hello world", - client=openai_client, - api_base="my-custom-api-base", - ) - else: - litellm.embedding( - model="litellm_proxy/my-vllm-model", - input="Hello world", - client=openai_client, - api_base="my-custom-api-base", - ) - except Exception as e: - print(e) + request_body = captured_bodies[0] + print("Request body - {}".format(request_body)) - mock_method.assert_called_once() - - print("Call KWARGS - {}".format(mock_method.call_args.kwargs)) - - assert "Hello world" == mock_method.call_args.kwargs["input"] - assert "my-vllm-model" == mock_method.call_args.kwargs["model"] + assert "Hello world" == request_body["input"] + assert "my-vllm-model" == request_body["model"] + assert "encoding_format" not in request_body + assert response.data[0]["embedding"] == [0.1, 0.2, 0.3] @pytest.mark.parametrize("is_async", [False, True]) diff --git a/tests/llm_translation/test_nvidia_nim.py b/tests/llm_translation/test_nvidia_nim.py index 7ee4f347f72..d5942e674d0 100644 --- a/tests/llm_translation/test_nvidia_nim.py +++ b/tests/llm_translation/test_nvidia_nim.py @@ -63,27 +63,39 @@ def test_embedding_nvidia_nim(): litellm.set_verbose = True from openai import OpenAI + captured_bodies = [] + + def handler(request: httpx.Request) -> httpx.Response: + captured_bodies.append(json.loads(request.content)) + return httpx.Response( + 200, + json={ + "object": "list", + "data": [{"object": "embedding", "index": 0, "embedding": [0.1, 0.2, 0.3]}], + "model": "nvidia/nv-embedqa-e5-v5", + "usage": {"prompt_tokens": 6, "total_tokens": 6}, + }, + ) + client = OpenAI( api_key="fake-api-key", + http_client=httpx.Client(transport=httpx.MockTransport(handler)), ) - with patch.object(client.embeddings.with_raw_response, "create") as mock_client: - try: - litellm.embedding( - model="nvidia_nim/nvidia/nv-embedqa-e5-v5", - input="What is the meaning of life?", - input_type="passage", - dimensions=1024, - client=client, - ) - except Exception as e: - print(e) - mock_client.assert_called_once() - request_body = mock_client.call_args.kwargs - print("request_body: ", request_body) - assert request_body["input"] == "What is the meaning of life?" - assert request_body["model"] == "nvidia/nv-embedqa-e5-v5" - assert request_body["extra_body"]["input_type"] == "passage" - assert request_body["dimensions"] == 1024 + response = litellm.embedding( + model="nvidia_nim/nvidia/nv-embedqa-e5-v5", + input="What is the meaning of life?", + input_type="passage", + dimensions=1024, + client=client, + ) + request_body = captured_bodies[0] + print("request_body: ", request_body) + assert request_body["input"] == "What is the meaning of life?" + assert request_body["model"] == "nvidia/nv-embedqa-e5-v5" + assert request_body["input_type"] == "passage" + assert request_body["dimensions"] == 1024 + assert "encoding_format" not in request_body + assert response.data[0]["embedding"] == [0.1, 0.2, 0.3] def test_chat_completion_nvidia_nim_with_tools(): diff --git a/tests/local_testing/test_exceptions.py b/tests/local_testing/test_exceptions.py index 8370046446d..e6392cda406 100644 --- a/tests/local_testing/test_exceptions.py +++ b/tests/local_testing/test_exceptions.py @@ -5,7 +5,7 @@ import traceback from typing import Any import httpx -from openai import AsyncOpenAI, AuthenticationError, BadRequestError, OpenAIError, RateLimitError +from openai import AsyncAzureOpenAI, AsyncOpenAI, AuthenticationError, AzureOpenAI, BadRequestError, OpenAIError, RateLimitError from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler @@ -895,7 +895,12 @@ def _pre_call_utils( ): if call_type == "embedding": data["input"] = "Hello world!" - mapped_target: Any = client.embeddings.with_raw_response + if isinstance(client, (AzureOpenAI, AsyncAzureOpenAI)): + mapped_target: Any = client.embeddings.with_raw_response + patched_attr = "create" + else: + mapped_target = client + patched_attr = "post" if sync_mode: original_function = litellm.embedding else: @@ -905,6 +910,7 @@ def _pre_call_utils( if streaming is True: data["stream"] = True mapped_target = client.chat.completions.with_raw_response # type: ignore + patched_attr = "create" if sync_mode: original_function = litellm.completion else: @@ -914,12 +920,13 @@ def _pre_call_utils( if streaming is True: data["stream"] = True mapped_target = client.completions.with_raw_response # type: ignore + patched_attr = "create" if sync_mode: original_function = litellm.text_completion else: original_function = litellm.atext_completion - return data, original_function, mapped_target + return data, original_function, mapped_target, patched_attr def _pre_call_utils_httpx( @@ -1003,7 +1010,7 @@ async def test_exception_with_headers(sync_mode, provider, model, call_type, str ) data = {"model": model} - data, original_function, mapped_target = _pre_call_utils( + data, original_function, mapped_target, patched_attr = _pre_call_utils( call_type=call_type, data=data, client=openai_client, @@ -1049,7 +1056,7 @@ async def test_exception_with_headers(sync_mode, provider, model, call_type, str with patch.object( mapped_target, - "create", + patched_attr, side_effect=_return_exception, ): new_retry_after_mock_client = MagicMock(return_value=-1) diff --git a/tests/local_testing/test_router.py b/tests/local_testing/test_router.py index 370c43f8f44..c714bb4f9a7 100644 --- a/tests/local_testing/test_router.py +++ b/tests/local_testing/test_router.py @@ -2032,8 +2032,8 @@ def test_router_dynamic_cooldown_correct_retry_after_time(): raise exception with patch.object( - openai_client.embeddings.with_raw_response, - "create", + openai_client, + "post", side_effect=_return_exception, ): new_retry_after_mock_client = MagicMock(return_value=-1) diff --git a/tests/test_litellm/test_openai_embedding_encoding_format_default.py b/tests/test_litellm/test_openai_embedding_encoding_format_default.py index 9842bf30585..7a42eaf0f0a 100644 --- a/tests/test_litellm/test_openai_embedding_encoding_format_default.py +++ b/tests/test_litellm/test_openai_embedding_encoding_format_default.py @@ -100,3 +100,22 @@ async def test_aembedding_openai_omits_encoding_format_when_client_omits_it( request_body: Final = json.loads(mock_route.calls.last.request.read()) assert "encoding_format" not in request_body assert response.data[0]["embedding"] == [0.1, 0.2, 0.3] + + +def test_embedding_openai_omitted_encoding_format_maps_provider_errors( + respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch +) -> None: + respx_mock.post("https://api.openai.com/v1/embeddings").mock( + return_value=httpx.Response( + 429, + headers={"retry-after": "42", "x-should-retry": "false"}, + json={"error": {"message": "rate limited", "type": "rate_limit_error"}}, + ) + ) + + with pytest.raises(litellm.RateLimitError) as exc_info: + litellm.embedding( + model="openai/text-embedding-3-small", input=["hello"], api_key="sk-test", max_retries=0 + ) + + assert int(exc_info.value.litellm_response_headers["retry-after"]) == 42 diff --git a/type-discipline-budget.json b/type-discipline-budget.json index f3f1a7defe7..67ebb2b3b17 100644 --- a/type-discipline-budget.json +++ b/type-discipline-budget.json @@ -27,7 +27,7 @@ "limit": 0 }, "LIT010": { - "limit": 16564 + "limit": 16562 }, "LIT011": { "limit": 5577 From 2a88384e4ec2d2f183fb69f223ed376b2eb24991 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 29 Aug 2026 12:15:32 -0700 Subject: [PATCH 5/5] style(utils): drop an em-dash from the vertex encoding_format comment --- litellm/utils.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/litellm/utils.py b/litellm/utils.py index 66a07d2e2db..c4250dd0e4b 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -3543,9 +3543,9 @@ def get_optional_params_embeddings( ) elif custom_llm_provider == "vertex_ai" or custom_llm_provider == "gemini": # OpenAI SDKs send encoding_format="float" by default; float lists are - # exactly what the vertex API returns, so the param is a no-op — don't - # reject the provider default. Other values (e.g. "base64") stay on - # the unsupported-param path below. + # exactly what the vertex API returns, so the param is a no-op and the + # provider default is not rejected. Other values (e.g. "base64") stay + # on the unsupported-param path below. if non_default_params.get("encoding_format") == "float": non_default_params.pop("encoding_format") supported_params = get_supported_openai_params(