Foundation model REST API reference

Artikkeli
12/18/2024

This article provides general API information for Databricks Foundation Model APIs and the models they support. The Foundation Model APIs are designed to be similar to OpenAI’s REST API to make migrating existing projects easier. Both the pay-per-token and provisioned throughput endpoints accept the same REST API request format.

Endpoints

Each pay-per-token model has a single endpoint, and users can interact with these endpoints using HTTP POST requests. Provisioned throughput endpoints can be created using the API or the Serving UI. These endpoints also support multiple models per endpoint for A/B testing, as long as both served models expose the same API format. For example, both models are chat models. See POST /api/2.0/serving-endpoints for endpoint configuration parameters.

Requests and responses use JSON, the exact JSON structure depends on an endpoint’s task type. Chat and completion endpoints support streaming responses.

Pay-per-token workloads support certain models, see Supported models for pay-per-token for those models and accepted API formats.

Usage

Responses include a usage sub-message which reports the number of tokens in the request and response. The format of this sub-message is the same across all task types.

Field	Type	Description
`completion_tokens`	Integer	Number of generated tokens. Not included in embedding responses.
`prompt_tokens`	Integer	Number of tokens from the input prompt(s).
`total_tokens`	Integer	Number of total tokens.

For models like llama-2-70b-chat a user prompt is transformed using a prompt template before being passed into the model. For pay-per-token endpoints, a system prompt might also be added. prompt_tokens includes all text added by our server.

Chat task

Chat tasks are optimized for multi-turn conversations with a model. Each request describes the conversation so far, where the messages field must alternate between user and assistant roles, ending with a user message. The model response provides the next assistant message in the conversation. See POST /serving-endpoints/{name}/invocations for querying endpoint parameters.

Chat request

Field	Default	Type	Description
`messages`		ChatMessage list	Required. A list of messages representing the current conversation.
`max_tokens`	`null`	`null`, which means no limit, or an integer greater than zero	The maximum number of tokens to generate.
`stream`	`true`	Boolean	Stream responses back to a client in order to allow partial results for requests. If this parameter is included in the request, responses are sent using the Server-sent events standard.
`temperature`	`1.0`	Float in [0,2]	The sampling temperature. 0 is deterministic and higher values introduce more randomness.
`top_p`	`1.0`	Float in (0,1]	The probability threshold used for nucleus sampling.
`top_k`	`null`	`null`, which means no limit, or an integer greater than zero	Defines the number of k most likely tokens to use for top-k-filtering. Set this value to 1 to make outputs deterministic.
`stop`	[]	String or List[String]	Model stops generating further tokens when any one of the sequences in `stop` is encountered.
`n`	1	Integer greater than zero	The API returns `n` independent chat completions when `n` is specified. Recommended for workloads that generate multiple completions on the same input for additional inference efficiency and cost savings. Only available for provisioned throughput endpoints.
`tool_choice`	`none`	String or ToolChoiceObject	Used only in conjunction with the `tools` field. `tool_choice` supports a variety of keyword strings such as `auto`, `required`, and `none`. `auto` means that you are letting the model decide which (if any) tool is relevant to use. With `auto` if the model doesn’t believe any of the tools in `tools` are relevant, the model generates a standard assistant message instead of a tool call. `required` means that the model picks the most relevant tool in `tools` and must generate a tool call. `none` means that the model does not generate any tool calls and instead must generate a standard assistant message. To force a tool call with a specific tool defined in `tools`, use a `ToolChoiceObject`. By default, if the `tools` field is populated `tool_choice = "auto"`. Else, the `tools` field defaults to `tool_choice = "none"`
`tools`	`null`	ToolObject	A list of `tools` that the model can call. Currently, `function` is the only supported `tool` type and a max of 32 functions are supported.
`response_format`	`null`	ResponseFormatObject	An object specifying the format that the model must output. Accepted types are `text`, `json_schema` or `json_object` Setting to `{ "type": "json_schema", "json_schema": {...} }` enables structured outputs which ensures the model follows your supplied JSON schema. Setting to `{ "type": "json_object" }` ensures the responses the model generates is valid JSON, but does not ensure that responses follow a specific schema.
`logprobs`	`false`	Boolean	This parameter indicates whether to provide the log probability of a token being sampled.
`top_logprobs`	`null`	Integer	This parameter controls the number of most likely token candidates to return log probabilities for at each sampling step. Can be 0-20. `logprobs` must be `true` if using this field.

`ChatMessage`

Field	Type	Description
`role`	String	Required. The role of the author of the message. Can be `"system"`, `"user"`, `"assistant"` or `"tool"`.
`content`	String	The content of the message. Required for chat tasks that do not involve tool calls.
`tool_calls`	ToolCall list	The list of `tool_calls` that the model generated. Must have `role` as `"assistant"` and no specification for the `content` field.
`tool_call_id`	String	When `role` is `"tool"`, the ID associated with the `ToolCall` that the message is responding to. Must be empty for other `role` options.

The system role can only be used once, as the first message in a conversation. It overrides the model’s default system prompt.

`ToolCall`

A tool call action suggestion by the model. See Function calling on Azure Databricks.

Field	Type	Description
`id`	String	Required. A unique identifier for this tool call suggestion.
`type`	String	Required. Only `"function"` is supported.
`function`	FunctionCallCompletion	Required. A function call suggested by the model.

`FunctionCallCompletion`

Field	Type	Description
`name`	String	Required. The name of the function the model recommended.
`arguments`	Object	Required. Arguments to the function as a serialized JSON dictionary.

`ToolChoiceObject`

See Function calling on Azure Databricks.

Field	Type	Description
`type`	String	Required. The type of the tool. Currently, only `"function"` is supported.
`function`	Object	Required. An object defining which tool to call of the form `{"type": "function", "function": {"name": "my_function"}}` where `"my_function` is the name of a FunctionObject in the `tools` field.

`ToolObject`

See Function calling on Azure Databricks.

Field	Type	Description
`type`	String	Required. The type of the tool. Currently, only `function` is supported.
`function`	FunctionObject	Required. The function definition associated with the tool.

`FunctionObject`

Field	Type	Description
`name`	String	Required. The name of the function to be called.
`description`	Object	Required. The detailed description of the function. The model uses this description to understand the relevance of the function to the prompt and generate the tool calls with higher accuracy.
`parameters`	Object	The parameters the function accepts, described as a valid JSON schema object. If the tool is called, then the tool call is fit to the JSON schema provided. Omitting parameters defines a function without any parameters. The number of `properties` is limited to 15 keys.
`strict`	Boolean	Whether to enable strict schema adherence when generating the function call. If set to `true`, the model follows the exact schema defined in the schema field. Only a subset of JSON schema is supported when strict is `true`

`ResponseFormatObject`

See Structured outputs on Azure Databricks.

Field	Type	Description
`type`	String	Required. The type of response format being defined. Either `text` for unstructured text, `json_object` for unstructured JSON objects, or `json_schema` for JSON objects adhering to a specific schema.
`json_schema`	JsonSchemaObject	Required. The JSON schema to adhere to if `type` is set to `json_schema`

`JsonSchemaObject`

See Structured outputs on Azure Databricks.

Field	Type	Description
`name`	String	Required. The name of the response format.
`description`	String	A description of what the response format is for, used by the model to determine how to respond in the format.
`schema`	Object	Required. The schema for the response format, described as a JSON schema object.
`strict`	Boolean	Whether to enable strict schema adherence when generating the output. If set to `true`, the model follows the exact schema defined in the schema field. Only a subset of JSON schema is supported when strict is `true`

Chat response

For non-streaming requests, the response is a single chat completion object. For streaming requests, the response is a text/event-stream where each event is a completion chunk object. The top-level structure of completion and chunk objects is almost identical: only choices has a different type.

Field	Type	Description
`id`	String	Unique identifier for the chat completion.
`choices`	List[ChatCompletionChoice] or List[ChatCompletionChunk] (streaming)	List of chat completion texts. `n` choices are returned if the `n` parameter is specified.
`object`	String	The object type. Equal to either `"chat.completions"` for non-streaming or `"chat.completion.chunk"` for streaming.
`created`	Integer	The time the chat completion was generated in seconds.
`model`	String	The model version used to generate the response.
`usage`	Usage	Token usage metadata. Might not be present on streaming responses.

`ChatCompletionChoice`

Field	Type	Description
`index`	Integer	The index of the choice in the list of generated choices.
`message`	ChatMessage	A chat completion message returned by the model. The role will be `assistant`.
`finish_reason`	String	The reason the model stopped generating tokens.

`ChatCompletionChunk`

Field	Type	Description
`index`	Integer	The index of the choice in the list of generated choices.
`delta`	ChatMessage	A chat completion message part of generated streamed responses from the model. Only the first chunk is guaranteed to have `role` populated.
`finish_reason`	String	The reason the model stopped generating tokens. Only the last chunk will have this populated.

Completion task

Text completion tasks are for generating responses to a single prompt. Unlike Chat, this task supports batched inputs: multiple independent prompts can be sent in one request. See POST /serving-endpoints/{name}/invocations for querying endpoint parameters.

Completion request

Field	Default	Type	Description
`prompt`		String or List[String]	Required. The prompts for the model.
`max_tokens`	`null`	`null`, which means no limit, or an integer greater than zero	The maximum number of tokens to generate.
`stream`	`true`	Boolean	Stream responses back to a client in order to allow partial results for requests. If this parameter is included in the request, responses are sent using the Server-sent events standard.
`temperature`	`1.0`	Float in [0,2]	The sampling temperature. 0 is deterministic and higher values introduce more randomness.
`top_p`	`1.0`	Float in (0,1]	The probability threshold used for nucleus sampling.
`top_k`	`null`	`null`, which means no limit, or an integer greater than zero	Defines the number of k most likely tokens to use for top-k-filtering. Set this value to 1 to make outputs deterministic.
`error_behavior`	`"error"`	`"truncate"` or `"error"`	For timeouts and context-length-exceeded errors. One of: `"truncate"` (return as many tokens as possible) and `"error"` (return an error). This parameter is only accepted by pay per token endpoints.
`n`	1	Integer greater than zero	The API returns `n` independent chat completions when `n` is specified. Recommended for workloads that generate multiple completions on the same input for additional inference efficiency and cost savings. Only available for provisioned throughput endpoints.
`stop`	[]	String or List[String]	Model stops generating further tokens when any one of the sequences in `stop` is encountered.
`suffix`	`""`	String	A string that is appended to the end of every completion.
`echo`	`false`	Boolean	Returns the prompt along with the completion.
`use_raw_prompt`	`false`	Boolean	If `true`, pass the `prompt` directly into the model without any transformation.

Completion response

Field	Type	Description
`id`	String	Unique identifier for the text completion.
`choices`	CompletionChoice	A list of text completions. For every prompt passed in, `n` choices are generated if `n` is specified. Default `n` is 1.
`object`	String	The object type. Equal to `"text_completion"`
`created`	Integer	The time the completion was generated in seconds.
`usage`	Usage	Token usage metadata.

`CompletionChoice`

Field	Type	Description
`index`	Integer	The index of the prompt in request.
`text`	String	The generated completion.
`finish_reason`	String	The reason the model stopped generating tokens.

Embedding task

Embedding tasks map input strings into embedding vectors. Many inputs can be batched together in each request. See POST /serving-endpoints/{name}/invocations for querying endpoint parameters.

Embedding request

Field	Type	Description
`input`	String or List[String]	Required. The input text to embed. Can be a string or a list of strings.
`instruction`	String	An optional instruction to pass to the embedding model.

Instructions are optional and highly model specific. For instance the The BGE authors recommend no instruction when indexing chunks and recommend using the instruction "Represent this sentence for searching relevant passages:" for retrieval queries. Other models like Instructor-XL support a wide range of instruction strings.

Embeddings response

Field	Type	Description
`id`	String	Unique identifier for the embedding.
`object`	String	The object type. Equal to `"list"`.
`model`	String	The name of the embedding model used to create the embedding.
`data`	EmbeddingObject	The embedding object.
`usage`	Usage	Token usage metadata.

`EmbeddingObject`

Field	Type	Description
`object`	String	The object type. Equal to `"embedding"`.
`index`	Integer	The index of the embedding in the list of embeddings generated by the model.
`embedding`	List[Float]	The embedding vector. Each model will return a fixed size vector (1024 for BGE-Large)

Jaa