Using OpenAI libraries with Vertex AI (original) (raw)

The Chat Completions API works as an Open AI-compatible endpoint, designed to make it easier to interface with Gemini on Vertex AI by using the OpenAI libraries for Python and REST. If you're already using the OpenAI libraries, you can use this API as a low-cost way to switch between calling OpenAI models and Vertex AI hosted models to compare output, cost, and scalability, without changing your existing code. If you aren't already using the OpenAI libraries, we recommend that youuse the Google Gen AI SDK.

Supported models

The Chat Completions API supports both Gemini models and select self-deployed models from Model Garden.

Gemini models

The following models provide support for the Chat Completions API:

Self-deployed models from Model Garden

TheHugging Face Text Generation Interface (HF TGI)andVertex AI Model Garden prebuilt vLLMcontainers support the Chat Completions API. However, not every model deployed to these containers supports the Chat Completions API. The following table includes the most popular supported models by container:

HF TGI vLLM
gemma-2-9b-it gemma-2-27b-it Meta-Llama-3.1-8B-Instruct Meta-Llama-3-8B-Instruct Mistral-7B-Instruct-v0.3 Mistral-Nemo-Instruct-2407 Gemma Llama 2 Llama 3 Mistral-7B Mistral Nemo

Supported parameters

For Google models, the Chat Completions API supports the following OpenAI parameters. For a description of each parameter, see OpenAI's documentation onCreating chat completions. Parameter support for third-party models varies by model. To see which parameters are supported, consult the model's documentation.

messages System message User message: The text andimage_url types are supported. Theimage_url type supports images stored a Cloud Storage URI or a base64 encoding in the form"data:;base64,". To learn how to create a Cloud Storage bucket and upload a file to it, seeDiscover object storage. The detail option is not supported. Assistant message Tool message Function message: This field is deprecated, but supported for backwards compatibility.
model
max_completion_tokens Alias for max_tokens.
max_tokens
n
frequency_penalty
presence_penalty
reasoning_effort Configures how much time and how many tokens are used on a response. low: 1024 medium: 8192 high: 24576 As no thoughts are included in the response, only one ofreasoning_effort or extra_body.google.thinking_config may be specified.
response_format json_object: Interpreted as passing "application/json" to the Gemini API. json_schema. Fully recursive schemas are not supported. additional_properties is supported. text: Interpreted as passing "text/plain" to the Gemini API. Any other MIME type is passed as is to the model, such as passing "application/json" directly.
seed Corresponds to GenerationConfig.seed.
stop
stream
temperature
top_p
tools type function name description parameters: Specify parameters by using theOpenAPI specification. This differs from the OpenAI parameters field, which is described as a JSON Schema object. To learn about keyword differences between OpenAPI and JSON Schema, see theOpenAPI guide.
tool_choice none auto required: Corresponds to the mode ANY in theFunctionCallingConfig. validated: Corresponds to the mode VALIDATED in the FunctionCallingConfig. This is Google-specific.
web_search_options Corresponds to the GoogleSearch tool. No sub-options are supported.
function_call This field is deprecated, but supported for backwards compatibility.
functions This field is deprecated, but supported for backwards compatibility.

If you pass any unsupported parameter, it is ignored.

Multimodal input parameters

The Chat Completions API supports select multimodal inputs.

input_audio data: Any URI or valid blob format. We support all blob types, including image, audio, and video. Anything supported by GenerateContent is supported (HTTP, Cloud Storage, etc.). format: OpenAI supports both wav (audio/wav) and mp3 (audio/mp3). Using Gemini, all valid MIME types are supported.
image_url data: Like input_audio, any URI or valid blob format is supported. Note that image_url as a URL will default to the image/* MIME-type and image_url as blob data can be used as any multimodal input. detail: Similar tomedia resolution, this determines the maximum tokens per image for the request. Note that while OpenAI's field is per-image, Gemini enforces the same detail across the request, and passing multiple detail types in one request will throw an error.

In general, the data parameter can be a URI or a combination of MIME type and base64 encoded bytes in the form "data:<MIME-TYPE>;base64,<BASE64-ENCODED-BYTES>". For a full list of MIME types, see GenerateContent. For more information on OpenAI's base64 encoding, see their documentation.

For usage, see our multimodal input examples.

Gemini-specific parameters

There are several features supported by Gemini that are not available in OpenAI models. These features can still be passed in as parameters, but must be contained within anextra_content or extra_body or they will be ignored.

extra_body features

safety_settings This corresponds to Gemini's SafetySetting.
cached_content This corresponds to Gemini's GenerateContentRequest.cached_content.
thinking_config This corresponds to Gemini's GenerationConfig.ThinkingConfig.
thought_tag_marker Used to separate a model's thoughts from its responses for models with Thinking available. If not specified, no tags will be returned around the model's thoughts. If present, subsequent queries will strip the thought tags and mark the thoughts appropriately for context. This helps preserve the appropriate context for subsequent queries.

extra_part lets you specify additional settings at a per-Part level.

extra_content A field for adding Gemini-specific content that shouldn't be ignored.
thought This will explicitly mark if a field is a thought (and take precedence overthought_tag_marker). This should be used to specify whether a tool call is part of a thought or not.

What's next