OpenAI compat embeddings API

docs/_static/llama-stack-spec.html

@@ -3607,6 +3607,49 @@
                 }
             }
         },
+        "/v1/openai/v1/embeddings": {
+            "post": {
+                "responses": {
+                    "200": {
+                        "description": "An OpenAIEmbeddingsResponse containing the embeddings.",
+                        "content": {
+                            "application/json": {
+                                "schema": {
+                                    "$ref": "#/components/schemas/OpenAIEmbeddingsResponse"
+                                }
+                            }
+                        }
+                    },
+                    "400": {
+                        "$ref": "#/components/responses/BadRequest400"
+                    },
+                    "429": {
+                        "$ref": "#/components/responses/TooManyRequests429"
+                    },
+                    "500": {
+                        "$ref": "#/components/responses/InternalServerError500"
+                    },
+                    "default": {
+                        "$ref": "#/components/responses/DefaultError"
+                    }
+                },
+                "tags": [
+                    "Inference"
+                ],
+                "description": "Generate OpenAI-compatible embeddings for the given input using the specified model.",
+                "parameters": [],
+                "requestBody": {
+                    "content": {
+                        "application/json": {
+                            "schema": {
+                                "$ref": "#/components/schemas/OpenaiEmbeddingsRequest"
+                            }
+                        }
+                    },
+                    "required": true
+                }
+            }
+        },
         "/v1/openai/v1/models": {
             "get": {
                 "responses": {
@@ -11767,6 +11810,139 @@
                 "title": "OpenAICompletionChoice",
                 "description": "A choice from an OpenAI-compatible completion response."
             },
+            "OpenaiEmbeddingsRequest": {
+                "type": "object",
+                "properties": {
+                    "model": {
+                        "type": "string",
+                        "description": "The identifier of the model to use. The model must be an embedding model registered with Llama Stack and available via the /models endpoint."
+                    },
+                    "input": {
+                        "oneOf": [
+                            {
+                                "type": "string"
+                            },
+                            {
+                                "type": "array",
+                                "items": {
+                                    "type": "string"
+                                }
+                            }
+                        ],
+                        "description": "Input text to embed, encoded as a string or array of strings. To embed multiple inputs in a single request, pass an array of strings."
+                    },
+                    "encoding_format": {
+                        "type": "string",
+                        "description": "(Optional) The format to return the embeddings in. Can be either \"float\" or \"base64\". Defaults to \"float\"."
+                    },
+                    "dimensions": {
+                        "type": "integer",
+                        "description": "(Optional) The number of dimensions the resulting output embeddings should have. Only supported in text-embedding-3 and later models."
+                    },
+                    "user": {
+                        "type": "string",
+                        "description": "(Optional) A unique identifier representing your end-user, which can help OpenAI to monitor and detect abuse."
+                    }
+                },
+                "additionalProperties": false,
+                "required": [
+                    "model",
+                    "input"
+                ],
+                "title": "OpenaiEmbeddingsRequest"
+            },
+            "OpenAIEmbeddingData": {
+                "type": "object",
+                "properties": {
+                    "object": {
+                        "type": "string",
+                        "const": "embedding",
+                        "default": "embedding",
+                        "description": "The object type, which will be \"embedding\""
+                    },
+                    "embedding": {
+                        "oneOf": [
+                            {
+                                "type": "array",
+                                "items": {
+                                    "type": "number"
+                                }
+                            },
+                            {
+                                "type": "string"
+                            }
+                        ],
+                        "description": "The embedding vector as a list of floats (when encoding_format=\"float\") or as a base64-encoded string (when encoding_format=\"base64\")"
+                    },
+                    "index": {
+                        "type": "integer",
+                        "description": "The index of the embedding in the input list"
+                    }
+                },
+                "additionalProperties": false,
+                "required": [
+                    "object",
+                    "embedding",
+                    "index"
+                ],
+                "title": "OpenAIEmbeddingData",
+                "description": "A single embedding data object from an OpenAI-compatible embeddings response."
+            },
+            "OpenAIEmbeddingUsage": {
+                "type": "object",
+                "properties": {
+                    "prompt_tokens": {
+                        "type": "integer",
+                        "description": "The number of tokens in the input"
+                    },
+                    "total_tokens": {
+                        "type": "integer",
+                        "description": "The total number of tokens used"
+                    }
+                },
+                "additionalProperties": false,
+                "required": [
+                    "prompt_tokens",
+                    "total_tokens"
+                ],
+                "title": "OpenAIEmbeddingUsage",
+                "description": "Usage information for an OpenAI-compatible embeddings response."
+            },
+            "OpenAIEmbeddingsResponse": {
+                "type": "object",
+                "properties": {
+                    "object": {
+                        "type": "string",
+                        "const": "list",
+                        "default": "list",
+                        "description": "The object type, which will be \"list\""
+                    },
+                    "data": {
+                        "type": "array",
+                        "items": {
+                            "$ref": "#/components/schemas/OpenAIEmbeddingData"
+                        },
+                        "description": "List of embedding data objects"
+                    },
+                    "model": {
+                        "type": "string",
+                        "description": "The model that was used to generate the embeddings"
+                    },
+                    "usage": {
+                        "$ref": "#/components/schemas/OpenAIEmbeddingUsage",
+                        "description": "Usage information"
+                    }
+                },
+                "additionalProperties": false,
+                "required": [
+                    "object",
+                    "data",
+                    "model",
+                    "usage"
+                ],
+                "title": "OpenAIEmbeddingsResponse",
+                "description": "Response from an OpenAI-compatible embeddings request."
+            },
             "OpenAIModel": {
                 "type": "object",
                 "properties": {

docs/_static/llama-stack-spec.yaml

@@ -2520,6 +2520,38 @@ paths:
             schema:
               $ref: '#/components/schemas/OpenaiCompletionRequest'
         required: true
+  /v1/openai/v1/embeddings:
+    post:
+      responses:
+        '200':
+          description: >-
+            An OpenAIEmbeddingsResponse containing the embeddings.
+          content:
+            application/json:
+              schema:
+                $ref: '#/components/schemas/OpenAIEmbeddingsResponse'
+        '400':
+          $ref: '#/components/responses/BadRequest400'
+        '429':
+          $ref: >-
+            #/components/responses/TooManyRequests429
+        '500':
+          $ref: >-
+            #/components/responses/InternalServerError500
+        default:
+          $ref: '#/components/responses/DefaultError'
+      tags:
+        - Inference
+      description: >-
+        Generate OpenAI-compatible embeddings for the given input using the specified
+        model.
+      parameters: []
+      requestBody:
+        content:
+          application/json:
+            schema:
+              $ref: '#/components/schemas/OpenaiEmbeddingsRequest'
+        required: true
   /v1/openai/v1/models:
     get:
       responses:
@@ -8177,6 +8209,118 @@ components:
       title: OpenAICompletionChoice
       description: >-
         A choice from an OpenAI-compatible completion response.
+    OpenaiEmbeddingsRequest:
+      type: object
+      properties:
+        model:
+          type: string
+          description: >-
+            The identifier of the model to use. The model must be an embedding model
+            registered with Llama Stack and available via the /models endpoint.
+        input:
+          oneOf:
+            - type: string
+            - type: array
+              items:
+                type: string
+          description: >-
+            Input text to embed, encoded as a string or array of strings. To embed
+            multiple inputs in a single request, pass an array of strings.
+        encoding_format:
+          type: string
+          description: >-
+            (Optional) The format to return the embeddings in. Can be either "float"
+            or "base64". Defaults to "float".
+        dimensions:
+          type: integer
+          description: >-
+            (Optional) The number of dimensions the resulting output embeddings should
+            have. Only supported in text-embedding-3 and later models.
+        user:
+          type: string
+          description: >-
+            (Optional) A unique identifier representing your end-user, which can help
+            OpenAI to monitor and detect abuse.
+      additionalProperties: false
+      required:
+        - model
+        - input
+      title: OpenaiEmbeddingsRequest
+    OpenAIEmbeddingData:
+      type: object
+      properties:
+        object:
+          type: string
+          const: embedding
+          default: embedding
+          description: >-
+            The object type, which will be "embedding"
+        embedding:
+          oneOf:
+            - type: array
+              items:
+                type: number
+            - type: string
+          description: >-
+            The embedding vector as a list of floats (when encoding_format="float")
+            or as a base64-encoded string (when encoding_format="base64")
+        index:
+          type: integer
+          description: >-
+            The index of the embedding in the input list
+      additionalProperties: false
+      required:
+        - object
+        - embedding
+        - index
+      title: OpenAIEmbeddingData
+      description: >-
+        A single embedding data object from an OpenAI-compatible embeddings response.
+    OpenAIEmbeddingUsage:
+      type: object
+      properties:
+        prompt_tokens:
+          type: integer
+          description: The number of tokens in the input
+        total_tokens:
+          type: integer
+          description: The total number of tokens used
+      additionalProperties: false
+      required:
+        - prompt_tokens
+        - total_tokens
+      title: OpenAIEmbeddingUsage
+      description: >-
+        Usage information for an OpenAI-compatible embeddings response.
+    OpenAIEmbeddingsResponse:
+      type: object
+      properties:
+        object:
+          type: string
+          const: list
+          default: list
+          description: The object type, which will be "list"
+        data:
+          type: array
+          items:
+            $ref: '#/components/schemas/OpenAIEmbeddingData'
+          description: List of embedding data objects
+        model:
+          type: string
+          description: >-
+            The model that was used to generate the embeddings
+        usage:
+          $ref: '#/components/schemas/OpenAIEmbeddingUsage'
+          description: Usage information
+      additionalProperties: false
+      required:
+        - object
+        - data
+        - model
+        - usage
+      title: OpenAIEmbeddingsResponse
+      description: >-
+        Response from an OpenAI-compatible embeddings request.
     OpenAIModel:
       type: object
       properties:

llama_stack/apis/inference/inference.py

@@ -783,6 +783,48 @@ class OpenAICompletion(BaseModel):
     object: Literal["text_completion"] = "text_completion"
 
 
+@json_schema_type
+class OpenAIEmbeddingData(BaseModel):
+    """A single embedding data object from an OpenAI-compatible embeddings response.
+
+    :param object: The object type, which will be "embedding"
+    :param embedding: The embedding vector as a list of floats (when encoding_format="float") or as a base64-encoded string (when encoding_format="base64")
+    :param index: The index of the embedding in the input list
+    """
+
+    object: Literal["embedding"] = "embedding"
+    embedding: list[float] | str
+    index: int
+
+
+@json_schema_type
+class OpenAIEmbeddingUsage(BaseModel):
+    """Usage information for an OpenAI-compatible embeddings response.
+
+    :param prompt_tokens: The number of tokens in the input
+    :param total_tokens: The total number of tokens used
+    """
+
+    prompt_tokens: int
+    total_tokens: int
+
+
+@json_schema_type
+class OpenAIEmbeddingsResponse(BaseModel):
+    """Response from an OpenAI-compatible embeddings request.
+
+    :param object: The object type, which will be "list"
+    :param data: List of embedding data objects
+    :param model: The model that was used to generate the embeddings
+    :param usage: Usage information
+    """
+
+    object: Literal["list"] = "list"
+    data: list[OpenAIEmbeddingData]
+    model: str
+    usage: OpenAIEmbeddingUsage
+
+
 class ModelStore(Protocol):
     async def get_model(self, identifier: str) -> Model: ...
 
@@ -1076,6 +1118,26 @@ class InferenceProvider(Protocol):
         """
         ...
 
+    @webmethod(route="/openai/v1/embeddings", method="POST")
+    async def openai_embeddings(
+        self,
+        model: str,
+        input: str | list[str],
+        encoding_format: str | None = "float",
+        dimensions: int | None = None,
+        user: str | None = None,
+    ) -> OpenAIEmbeddingsResponse:
+        """Generate OpenAI-compatible embeddings for the given input using the specified model.
+
+        :param model: The identifier of the model to use. The model must be an embedding model registered with Llama Stack and available via the /models endpoint.
+        :param input: Input text to embed, encoded as a string or array of strings. To embed multiple inputs in a single request, pass an array of strings.
+        :param encoding_format: (Optional) The format to return the embeddings in. Can be either "float" or "base64". Defaults to "float".
+        :param dimensions: (Optional) The number of dimensions the resulting output embeddings should have. Only supported in text-embedding-3 and later models.
+        :param user: (Optional) A unique identifier representing your end-user, which can help OpenAI to monitor and detect abuse.
+        :returns: An OpenAIEmbeddingsResponse containing the embeddings.
+        """
+        ...
+
 
 class Inference(InferenceProvider):
     """Llama Stack Inference API for generating completions, chat completions, and embeddings.

llama_stack/distribution/routers/inference.py

@@ -45,6 +45,7 @@ from llama_stack.apis.inference.inference import (
     OpenAIChatCompletion,
     OpenAIChatCompletionChunk,
     OpenAICompletion,
+    OpenAIEmbeddingsResponse,
     OpenAIMessageParam,
     OpenAIResponseFormatParam,
 )
@@ -546,6 +547,34 @@ class InferenceRouter(Inference):
                 await self.store.store_chat_completion(response, messages)
             return response
 
+    async def openai_embeddings(
+        self,
+        model: str,
+        input: str | list[str],
+        encoding_format: str | None = "float",
+        dimensions: int | None = None,
+        user: str | None = None,
+    ) -> OpenAIEmbeddingsResponse:
+        logger.debug(
+            f"InferenceRouter.openai_embeddings: {model=}, input_type={type(input)}, {encoding_format=}, {dimensions=}",
+        )
+        model_obj = await self.routing_table.get_model(model)
+        if model_obj is None:
+            raise ValueError(f"Model '{model}' not found")
+        if model_obj.model_type != ModelType.embedding:
+            raise ValueError(f"Model '{model}' is not an embedding model")
+
+        params = dict(
+            model=model_obj.identifier,
+            input=input,
+            encoding_format=encoding_format,
+            dimensions=dimensions,
+            user=user,
+        )
+
+        provider = self.routing_table.get_provider_impl(model_obj.identifier)
+        return await provider.openai_embeddings(**params)
+
     async def list_chat_completions(
         self,
         after: str | None = None,

llama_stack/providers/inline/inference/vllm/vllm.py

@@ -40,6 +40,7 @@ from llama_stack.apis.inference import (
     JsonSchemaResponseFormat,
     LogProbConfig,
     Message,
+    OpenAIEmbeddingsResponse,
     ResponseFormat,
     SamplingParams,
     TextTruncation,
@@ -410,6 +411,16 @@ class VLLMInferenceImpl(
     ) -> EmbeddingsResponse:
         raise NotImplementedError()
 
+    async def openai_embeddings(
+        self,
+        model: str,
+        input: str | list[str],
+        encoding_format: str | None = "float",
+        dimensions: int | None = None,
+        user: str | None = None,
+    ) -> OpenAIEmbeddingsResponse:
+        raise NotImplementedError()
+
     async def chat_completion(
         self,
         model_id: str,

llama_stack/providers/remote/inference/bedrock/bedrock.py

@@ -22,6 +22,7 @@ from llama_stack.apis.inference import (
     Inference,
     LogProbConfig,
     Message,
+    OpenAIEmbeddingsResponse,
     ResponseFormat,
     SamplingParams,
     TextTruncation,
@@ -197,3 +198,13 @@ class BedrockInferenceAdapter(
             response_body = json.loads(response.get("body").read())
             embeddings.append(response_body.get("embedding"))
         return EmbeddingsResponse(embeddings=embeddings)
+
+    async def openai_embeddings(
+        self,
+        model: str,
+        input: str | list[str],
+        encoding_format: str | None = "float",
+        dimensions: int | None = None,
+        user: str | None = None,
+    ) -> OpenAIEmbeddingsResponse:
+        raise NotImplementedError()

llama_stack/providers/remote/inference/cerebras/cerebras.py

@@ -21,6 +21,7 @@ from llama_stack.apis.inference import (
     Inference,
     LogProbConfig,
     Message,
+    OpenAIEmbeddingsResponse,
     ResponseFormat,
     SamplingParams,
     TextTruncation,
@@ -194,3 +195,13 @@ class CerebrasInferenceAdapter(
         task_type: EmbeddingTaskType | None = None,
     ) -> EmbeddingsResponse:
         raise NotImplementedError()
+
+    async def openai_embeddings(
+        self,
+        model: str,
+        input: str | list[str],
+        encoding_format: str | None = "float",
+        dimensions: int | None = None,
+        user: str | None = None,
+    ) -> OpenAIEmbeddingsResponse:
+        raise NotImplementedError()

llama_stack/providers/remote/inference/databricks/databricks.py

@@ -20,6 +20,7 @@ from llama_stack.apis.inference import (
     Inference,
     LogProbConfig,
     Message,
+    OpenAIEmbeddingsResponse,
     ResponseFormat,
     SamplingParams,
     TextTruncation,
@@ -152,3 +153,13 @@ class DatabricksInferenceAdapter(
         task_type: EmbeddingTaskType | None = None,
     ) -> EmbeddingsResponse:
         raise NotImplementedError()
+
+    async def openai_embeddings(
+        self,
+        model: str,
+        input: str | list[str],
+        encoding_format: str | None = "float",
+        dimensions: int | None = None,
+        user: str | None = None,
+    ) -> OpenAIEmbeddingsResponse:
+        raise NotImplementedError()

llama_stack/providers/remote/inference/fireworks/fireworks.py

@@ -37,6 +37,7 @@ from llama_stack.apis.inference.inference import (
     OpenAIChatCompletion,
     OpenAIChatCompletionChunk,
     OpenAICompletion,
+    OpenAIEmbeddingsResponse,
     OpenAIMessageParam,
     OpenAIResponseFormatParam,
 )
@@ -286,6 +287,16 @@ class FireworksInferenceAdapter(ModelRegistryHelper, Inference, NeedsRequestProv
         embeddings = [data.embedding for data in response.data]
         return EmbeddingsResponse(embeddings=embeddings)
 
+    async def openai_embeddings(
+        self,
+        model: str,
+        input: str | list[str],
+        encoding_format: str | None = "float",
+        dimensions: int | None = None,
+        user: str | None = None,
+    ) -> OpenAIEmbeddingsResponse:
+        raise NotImplementedError()
+
     async def openai_completion(
         self,
         model: str,

llama_stack/providers/remote/inference/nvidia/nvidia.py

@@ -29,6 +29,7 @@ from llama_stack.apis.inference import (
     Inference,
     LogProbConfig,
     Message,
+    OpenAIEmbeddingsResponse,
     ResponseFormat,
     SamplingParams,
     TextTruncation,
@@ -238,6 +239,16 @@ class NVIDIAInferenceAdapter(Inference, ModelRegistryHelper):
         #
         return EmbeddingsResponse(embeddings=[embedding.embedding for embedding in response.data])
 
+    async def openai_embeddings(
+        self,
+        model: str,
+        input: str | list[str],
+        encoding_format: str | None = "float",
+        dimensions: int | None = None,
+        user: str | None = None,
+    ) -> OpenAIEmbeddingsResponse:
+        raise NotImplementedError()
+
     async def chat_completion(
         self,
         model_id: str,

llama_stack/providers/remote/inference/ollama/ollama.py

@@ -32,6 +32,7 @@ from llama_stack.apis.inference import (
     JsonSchemaResponseFormat,
     LogProbConfig,
     Message,
+    OpenAIEmbeddingsResponse,
     ResponseFormat,
     SamplingParams,
     TextTruncation,
@@ -370,6 +371,16 @@ class OllamaInferenceAdapter(
 
         return model
 
+    async def openai_embeddings(
+        self,
+        model: str,
+        input: str | list[str],
+        encoding_format: str | None = "float",
+        dimensions: int | None = None,
+        user: str | None = None,
+    ) -> OpenAIEmbeddingsResponse:
+        raise NotImplementedError()
+
     async def openai_completion(
         self,
         model: str,

llama_stack/providers/remote/inference/openai/openai.py

@@ -14,6 +14,9 @@ from llama_stack.apis.inference.inference import (
     OpenAIChatCompletion,
     OpenAIChatCompletionChunk,
     OpenAICompletion,
+    OpenAIEmbeddingData,
+    OpenAIEmbeddingsResponse,
+    OpenAIEmbeddingUsage,
     OpenAIMessageParam,
     OpenAIResponseFormatParam,
 )
@@ -38,6 +41,7 @@ logger = logging.getLogger(__name__)
 # | batch_chat_completion      | LiteLLMOpenAIMixin       |
 # | openai_completion          | AsyncOpenAI              |
 # | openai_chat_completion     | AsyncOpenAI              |
+# | openai_embeddings          | AsyncOpenAI              |
 #
 class OpenAIInferenceAdapter(LiteLLMOpenAIMixin):
     def __init__(self, config: OpenAIConfig) -> None:
@@ -171,3 +175,51 @@ class OpenAIInferenceAdapter(LiteLLMOpenAIMixin):
             user=user,
         )
         return await self._openai_client.chat.completions.create(**params)
+
+    async def openai_embeddings(
+        self,
+        model: str,
+        input: str | list[str],
+        encoding_format: str | None = "float",
+        dimensions: int | None = None,
+        user: str | None = None,
+    ) -> OpenAIEmbeddingsResponse:
+        model_id = (await self.model_store.get_model(model)).provider_resource_id
+        if model_id.startswith("openai/"):
+            model_id = model_id[len("openai/") :]
+
+        # Prepare parameters for OpenAI embeddings API
+        params = {
+            "model": model_id,
+            "input": input,
+        }
+
+        if encoding_format is not None:
+            params["encoding_format"] = encoding_format
+        if dimensions is not None:
+            params["dimensions"] = dimensions
+        if user is not None:
+            params["user"] = user
+
+        # Call OpenAI embeddings API
+        response = await self._openai_client.embeddings.create(**params)
+
+        data = []
+        for i, embedding_data in enumerate(response.data):
+            data.append(
+                OpenAIEmbeddingData(
+                    embedding=embedding_data.embedding,
+                    index=i,
+                )
+            )
+
+        usage = OpenAIEmbeddingUsage(
+            prompt_tokens=response.usage.prompt_tokens,
+            total_tokens=response.usage.total_tokens,
+        )
+
+        return OpenAIEmbeddingsResponse(
+            data=data,
+            model=response.model,
+            usage=usage,
+        )

llama_stack/providers/remote/inference/passthrough/passthrough.py

@@ -19,6 +19,7 @@ from llama_stack.apis.inference import (
     Inference,
     LogProbConfig,
     Message,
+    OpenAIEmbeddingsResponse,
     ResponseFormat,
     SamplingParams,
     TextTruncation,
@@ -210,6 +211,16 @@ class PassthroughInferenceAdapter(Inference):
             task_type=task_type,
         )
 
+    async def openai_embeddings(
+        self,
+        model: str,
+        input: str | list[str],
+        encoding_format: str | None = "float",
+        dimensions: int | None = None,
+        user: str | None = None,
+    ) -> OpenAIEmbeddingsResponse:
+        raise NotImplementedError()
+
     async def openai_completion(
         self,
         model: str,

llama_stack/providers/remote/inference/runpod/runpod.py

@@ -8,6 +8,7 @@ from collections.abc import AsyncGenerator
 from openai import OpenAI
 
 from llama_stack.apis.inference import *  # noqa: F403
+from llama_stack.apis.inference.inference import OpenAIEmbeddingsResponse
 
 # from llama_stack.providers.datatypes import ModelsProtocolPrivate
 from llama_stack.providers.utils.inference.model_registry import ModelRegistryHelper
@@ -134,3 +135,13 @@ class RunpodInferenceAdapter(
         task_type: Optional[EmbeddingTaskType] = None,
     ) -> EmbeddingsResponse:
         raise NotImplementedError()
+
+    async def openai_embeddings(
+        self,
+        model: str,
+        input: str | list[str],
+        encoding_format: str | None = "float",
+        dimensions: int | None = None,
+        user: str | None = None,
+    ) -> OpenAIEmbeddingsResponse:
+        raise NotImplementedError()

llama_stack/providers/remote/inference/tgi/tgi.py

@@ -23,6 +23,7 @@ from llama_stack.apis.inference import (
     Inference,
     LogProbConfig,
     Message,
+    OpenAIEmbeddingsResponse,
     ResponseFormat,
     ResponseFormatType,
     SamplingParams,
@@ -291,6 +292,16 @@ class _HfAdapter(
     ) -> EmbeddingsResponse:
         raise NotImplementedError()
 
+    async def openai_embeddings(
+        self,
+        model: str,
+        input: str | list[str],
+        encoding_format: str | None = "float",
+        dimensions: int | None = None,
+        user: str | None = None,
+    ) -> OpenAIEmbeddingsResponse:
+        raise NotImplementedError()
+
 
 class TGIAdapter(_HfAdapter):
     async def initialize(self, config: TGIImplConfig) -> None:

llama_stack/providers/remote/inference/together/together.py

@@ -23,6 +23,7 @@ from llama_stack.apis.inference import (
     Inference,
     LogProbConfig,
     Message,
+    OpenAIEmbeddingsResponse,
     ResponseFormat,
     ResponseFormatType,
     SamplingParams,
@@ -267,6 +268,16 @@ class TogetherInferenceAdapter(ModelRegistryHelper, Inference, NeedsRequestProvi
         embeddings = [item.embedding for item in r.data]
         return EmbeddingsResponse(embeddings=embeddings)
 
+    async def openai_embeddings(
+        self,
+        model: str,
+        input: str | list[str],
+        encoding_format: str | None = "float",
+        dimensions: int | None = None,
+        user: str | None = None,
+    ) -> OpenAIEmbeddingsResponse:
+        raise NotImplementedError()
+
     async def openai_completion(
         self,
         model: str,

llama_stack/providers/remote/inference/vllm/vllm.py

@@ -38,6 +38,7 @@ from llama_stack.apis.inference import (
     JsonSchemaResponseFormat,
     LogProbConfig,
     Message,
+    OpenAIEmbeddingsResponse,
     ResponseFormat,
     SamplingParams,
     TextTruncation,
@@ -507,6 +508,16 @@ class VLLMInferenceAdapter(Inference, ModelsProtocolPrivate):
         embeddings = [data.embedding for data in response.data]
         return EmbeddingsResponse(embeddings=embeddings)
 
+    async def openai_embeddings(
+        self,
+        model: str,
+        input: str | list[str],
+        encoding_format: str | None = "float",
+        dimensions: int | None = None,
+        user: str | None = None,
+    ) -> OpenAIEmbeddingsResponse:
+        raise NotImplementedError()
+
     async def openai_completion(
         self,
         model: str,

llama_stack/providers/remote/inference/watsonx/watsonx.py

@@ -21,6 +21,7 @@ from llama_stack.apis.inference import (
     Inference,
     LogProbConfig,
     Message,
+    OpenAIEmbeddingsResponse,
     ResponseFormat,
     SamplingParams,
     TextTruncation,
@@ -260,6 +261,16 @@ class WatsonXInferenceAdapter(Inference, ModelRegistryHelper):
     ) -> EmbeddingsResponse:
         raise NotImplementedError("embedding is not supported for watsonx")
 
+    async def openai_embeddings(
+        self,
+        model: str,
+        input: str | list[str],
+        encoding_format: str | None = "float",
+        dimensions: int | None = None,
+        user: str | None = None,
+    ) -> OpenAIEmbeddingsResponse:
+        raise NotImplementedError()
+
     async def openai_completion(
         self,
         model: str,

llama_stack/providers/utils/inference/embedding_mixin.py

@@ -4,7 +4,9 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
+import base64
 import logging
+import struct
 from typing import TYPE_CHECKING
 
 if TYPE_CHECKING:
@@ -15,6 +17,9 @@ from llama_stack.apis.inference import (
     EmbeddingTaskType,
     InterleavedContentItem,
     ModelStore,
+    OpenAIEmbeddingData,
+    OpenAIEmbeddingsResponse,
+    OpenAIEmbeddingUsage,
     TextTruncation,
 )
 from llama_stack.providers.utils.inference.prompt_adapter import interleaved_content_as_str
@@ -43,6 +48,50 @@ class SentenceTransformerEmbeddingMixin:
         )
         return EmbeddingsResponse(embeddings=embeddings)
 
+    async def openai_embeddings(
+        self,
+        model: str,
+        input: str | list[str],
+        encoding_format: str | None = "float",
+        dimensions: int | None = None,
+        user: str | None = None,
+    ) -> OpenAIEmbeddingsResponse:
+        # Convert input to list format if it's a single string
+        input_list = [input] if isinstance(input, str) else input
+        if not input_list:
+            raise ValueError("Empty list not supported")
+
+        # Get the model and generate embeddings
+        model_obj = await self.model_store.get_model(model)
+        embedding_model = self._load_sentence_transformer_model(model_obj.provider_resource_id)
+        embeddings = embedding_model.encode(input_list, show_progress_bar=False)
+
+        # Convert embeddings to the requested format
+        data = []
+        for i, embedding in enumerate(embeddings):
+            if encoding_format == "base64":
+                # Convert float array to base64 string
+                float_bytes = struct.pack(f"{len(embedding)}f", *embedding)
+                embedding_value = base64.b64encode(float_bytes).decode("ascii")
+            else:
+                # Default to float format
+                embedding_value = embedding.tolist()
+
+            data.append(
+                OpenAIEmbeddingData(
+                    embedding=embedding_value,
+                    index=i,
+                )
+            )
+
+        # Not returning actual token usage
+        usage = OpenAIEmbeddingUsage(prompt_tokens=-1, total_tokens=-1)
+        return OpenAIEmbeddingsResponse(
+            data=data,
+            model=model_obj.provider_resource_id,
+            usage=usage,
+        )
+
     def _load_sentence_transformer_model(self, model: str) -> "SentenceTransformer":
         global EMBEDDING_MODELS
 

llama_stack/providers/utils/inference/litellm_openai_mixin.py

@@ -4,6 +4,8 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
+import base64
+import struct
 from collections.abc import AsyncGenerator, AsyncIterator
 from typing import Any
 
@@ -35,6 +37,9 @@ from llama_stack.apis.inference.inference import (
     OpenAIChatCompletion,
     OpenAIChatCompletionChunk,
     OpenAICompletion,
+    OpenAIEmbeddingData,
+    OpenAIEmbeddingsResponse,
+    OpenAIEmbeddingUsage,
     OpenAIMessageParam,
     OpenAIResponseFormatParam,
 )
@@ -264,6 +269,52 @@ class LiteLLMOpenAIMixin(
         embeddings = [data["embedding"] for data in response["data"]]
         return EmbeddingsResponse(embeddings=embeddings)
 
+    async def openai_embeddings(
+        self,
+        model: str,
+        input: str | list[str],
+        encoding_format: str | None = "float",
+        dimensions: int | None = None,
+        user: str | None = None,
+    ) -> OpenAIEmbeddingsResponse:
+        model_obj = await self.model_store.get_model(model)
+
+        # Convert input to list if it's a string
+        input_list = [input] if isinstance(input, str) else input
+
+        # Call litellm embedding function
+        # litellm.drop_params = True
+        response = litellm.embedding(
+            model=self.get_litellm_model_name(model_obj.provider_resource_id),
+            input=input_list,
+            api_key=self.get_api_key(),
+            api_base=self.api_base,
+            dimensions=dimensions,
+        )
+
+        # Convert response to OpenAI format
+        data = []
+        for i, embedding_data in enumerate(response["data"]):
+            # we encode to base64 if the encoding format is base64 in the request
+            if encoding_format == "base64":
+                byte_data = b"".join(struct.pack("f", f) for f in embedding_data["embedding"])
+                embedding = base64.b64encode(byte_data).decode("utf-8")
+            else:
+                embedding = embedding_data["embedding"]
+
+            data.append(OpenAIEmbeddingData(embedding=embedding, index=i))
+
+        usage = OpenAIEmbeddingUsage(
+            prompt_tokens=response["usage"]["prompt_tokens"],
+            total_tokens=response["usage"]["total_tokens"],
+        )
+
+        return OpenAIEmbeddingsResponse(
+            data=data,
+            model=model_obj.provider_resource_id,
+            usage=usage,
+        )
+
     async def openai_completion(
         self,
         model: str,