Google Gen AI SDK

---

Documentation:https://googleapis.github.io/python-genai/

---

Google Gen AI Python SDK provides an interface for developers to integrate Google's generative models into their Python applications. It supports the Gemini Developer API and Vertex AI APIs.

Installation

pip install google-genai

Imports

fromgoogleimportgenaifromgoogle.genaiimporttypes

Create a client

Please run one of the following code blocks to create a client for different services (Gemini Developer API or Vertex AI).

fromgoogleimportgenai# Only run this block for Gemini Developer APIclient=genai.Client(api_key='GEMINI_API_KEY')

fromgoogleimportgenai# Only run this block for Vertex AI APIclient=genai.Client(
vertexai=True, project='your-project-id', location='us-central1'
)

(Optional) Using environment variables:

You can create a client by configuring the necessary environment variables. Configuration setup instructions depends on whether you're using the Gemini Developer API or the Gemini API in Vertex AI.

Gemini Developer API: Set GOOGLE_API_KEY as shown below:

export GOOGLE_API_KEY='your-api-key'

Gemini API on Vertex AI: Set GOOGLE_GENAI_USE_VERTEXAI, GOOGLE_CLOUD_PROJECT and GOOGLE_CLOUD_LOCATION, as shown below:

export GOOGLE_GENAI_USE_VERTEXAI=true
export GOOGLE_CLOUD_PROJECT='your-project-id'export GOOGLE_CLOUD_LOCATION='us-central1'

fromgoogleimportgenaiclient=genai.Client()

API Selection

By default, the SDK uses the beta API endpoints provided by Google to support preview features in the APIs. The stable API endpoints can be selected by setting the API version to v1.

To set the API version use http_options. For example, to set the API version to v1 for Vertex AI:

fromgoogleimportgenaifromgoogle.genaiimporttypesclient=genai.Client(
vertexai=True,
project='your-project-id',
location='us-central1',
http_options=types.HttpOptions(api_version='v1')
)

To set the API version to v1alpha for the Gemini Developer API:

fromgoogleimportgenaifromgoogle.genaiimporttypesclient=genai.Client(
api_key='GEMINI_API_KEY',
http_options=types.HttpOptions(api_version='v1alpha')
)

Faster async client option: Aiohttp

By default we use httpx for both sync and async client implementations. In order to have faster performance, you may install google-genai[aiohttp]. In Gen AI SDK we configure trust_env=True to match with the default behavior of httpx. Additional args of aiohttp.ClientSession.request() (see _RequestOptions args) can be passed through the following way:

http_options=types.HttpOptions(
async_client_args={'cookies': ..., 'ssl': ...},
)
client=Client(..., http_options=http_options)

Proxy

Both httpx and aiohttp libraries use urllib.request.getproxies from environment variables. Before client initialization, you may set proxy (and optional SSL_CERT_FILE) by setting the environment variables:

export HTTPS_PROXY='http://username:password@proxy_uri:port'export SSL_CERT_FILE='client.pem'

If you need socks5 proxy, httpx supportssocks5 proxy if you pass it via args to httpx.Client(). You may install httpx[socks] to use it. Then, you can pass it through the following way:

http_options=types.HttpOptions(
client_args={'proxy': 'socks5://user:pass@host:port'},
async_client_args={'proxy': 'socks5://user:pass@host:port'},,
)
client=Client(..., http_options=http_options)

Types

Parameter types can be specified as either dictionaries(TypedDict) or Pydantic Models. Pydantic model types are available in the types module.

Models

The client.models module exposes model inferencing and model getters. See the 'Create a client' section above to initialize a client.

Generate Content

with text content

response=client.models.generate_content(
model='gemini-2.0-flash-001', contents='Why is the sky blue?'
)
print(response.text)

with uploaded file (Gemini Developer API only)

download the file in console.

!wget -q https://storage.googleapis.com/generativeai-downloads/data/a11.txt

python code.

file=client.files.upload(file='a11.txt')
response=client.models.generate_content(
model='gemini-2.0-flash-001',
contents=['Could you summarize this file?', file]
)
print(response.text)

How to structure contents argument for generate_content

The SDK always converts the inputs to the contents argument into list[types.Content]. The following shows some common ways to provide your inputs.

Provide a list[types.Content]

This is the canonical way to provide contents, SDK will not do any conversion.

Provide a types.Content instance

fromgoogle.genaiimporttypescontents=types.Content(
role='user',
parts=[types.Part.from_text(text='Why is the sky blue?')]
)

SDK converts this to

[
types.Content(
role='user',
parts=[types.Part.from_text(text='Why is the sky blue?')]
)
]

Provide a string

contents='Why is the sky blue?'

The SDK will assume this is a text part, and it converts this into the following:

[
types.UserContent(
parts=[
types.Part.from_text(text='Why is the sky blue?')
]
)
]

Where a types.UserContent is a subclass of types.Content, it sets the role field to be user.

Provide a list of string

contents=['Why is the sky blue?', 'Why is the cloud white?']

The SDK assumes these are 2 text parts, it converts this into a single content, like the following:

[
types.UserContent(
parts=[
types.Part.from_text(text='Why is the sky blue?'),
types.Part.from_text(text='Why is the cloud white?'),
]
)
]

Where a types.UserContent is a subclass of types.Content, the role field in types.UserContent is fixed to be user.

Provide a function call part

fromgoogle.genaiimporttypescontents=types.Part.from_function_call(
name='get_weather_by_location',
args={'location': 'Boston'}
)

The SDK converts a function call part to a content with a model role:

[
types.ModelContent(
parts=[
types.Part.from_function_call(
name='get_weather_by_location',
args={'location': 'Boston'}
)
]
)
]

Where a types.ModelContent is a subclass of types.Content, the role field in types.ModelContent is fixed to be model.

Provide a list of function call parts

fromgoogle.genaiimporttypescontents= [
types.Part.from_function_call(
name='get_weather_by_location',
args={'location': 'Boston'}
),
types.Part.from_function_call(
name='get_weather_by_location',
args={'location': 'New York'}
),
]

The SDK converts a list of function call parts to the a content with a model role:

[
types.ModelContent(
parts=[
types.Part.from_function_call(
name='get_weather_by_location',
args={'location': 'Boston'}
),
types.Part.from_function_call(
name='get_weather_by_location',
args={'location': 'New York'}
)
]
)
]

Where a types.ModelContent is a subclass of types.Content, the role field in types.ModelContent is fixed to be model.

Provide a non function call part

fromgoogle.genaiimporttypescontents=types.Part.from_uri(
file_uri: 'gs://generativeai-downloads/images/scones.jpg',
mime_type: 'image/jpeg',
)

The SDK converts all non function call parts into a content with a user role.

[
types.UserContent(parts=[
types.Part.from_uri(
file_uri: 'gs://generativeai-downloads/images/scones.jpg',
mime_type: 'image/jpeg',
)
])
]

Provide a list of non function call parts

fromgoogle.genaiimporttypescontents= [
types.Part.from_text('What is this image about?'),
types.Part.from_uri(
file_uri: 'gs://generativeai-downloads/images/scones.jpg',
mime_type: 'image/jpeg',
)
]

The SDK will convert the list of parts into a content with a user role

[
types.UserContent(
parts=[
types.Part.from_text('What is this image about?'),
types.Part.from_uri(
file_uri: 'gs://generativeai-downloads/images/scones.jpg',
mime_type: 'image/jpeg',
)
]
)
]

Mix types in contents

You can also provide a list of types.ContentUnion. The SDK leaves items of types.Content as is, it groups consecutive non function call parts into a single types.UserContent, and it groups consecutive function call parts into a single types.ModelContent.

If you put a list within a list, the inner list can only contain types.PartUnion items. The SDK will convert the inner list into a single types.UserContent.

System Instructions and Other Configs

The output of the model can be influenced by several optional settings available in generate_content's config parameter. For example, increasing max_output_tokens is essential for longer model responses. To make a model more deterministic, lowering the temperature parameter reduces randomness, with values near 0 minimizing variability. Capabilities and parameter defaults for each model is shown in the Vertex AI docs and Gemini API docs respectively.

fromgoogle.genaiimporttypesresponse=client.models.generate_content(
model='gemini-2.0-flash-001',
contents='high',
config=types.GenerateContentConfig(
system_instruction='I say high, you say low',
max_output_tokens=3,
temperature=0.3,
),
)
print(response.text)

Typed Config

All API methods support Pydantic types for parameters as well as dictionaries. You can get the type from google.genai.types.

fromgoogle.genaiimporttypesresponse=client.models.generate_content(
model='gemini-2.0-flash-001',
contents=types.Part.from_text(text='Why is the sky blue?'),
config=types.GenerateContentConfig(
temperature=0,
top_p=0.95,
top_k=20,
candidate_count=1,
seed=5,
max_output_tokens=100,
stop_sequences=['STOP!'],
presence_penalty=0.0,
frequency_penalty=0.0,
),
)
print(response.text)

List Base Models

To retrieve tuned models, see list tuned models.

formodelinclient.models.list():
print(model)

pager=client.models.list(config={'page_size': 10})
print(pager.page_size)
print(pager[0])
pager.next_page()
print(pager[0])

List Base Models (Asynchronous)

asyncforjobinawaitclient.aio.models.list():
print(job)

async_pager=awaitclient.aio.models.list(config={'page_size': 10})
print(async_pager.page_size)
print(async_pager[0])
awaitasync_pager.next_page()
print(async_pager[0])

Safety Settings

fromgoogle.genaiimporttypesresponse=client.models.generate_content(
model='gemini-2.0-flash-001',
contents='Say something bad.',
config=types.GenerateContentConfig(
safety_settings=[
types.SafetySetting(
category='HARM_CATEGORY_HATE_SPEECH',
threshold='BLOCK_ONLY_HIGH',
)
]
),
)
print(response.text)

Function Calling

Automatic Python function Support

You can pass a Python function directly and it will be automatically called and responded by default.

fromgoogle.genaiimporttypesdefget_current_weather(location: str) ->str:
"""Returns the current weather. Args: location: The city and state, e.g. San Francisco, CA """return'sunny'response=client.models.generate_content(
model='gemini-2.0-flash-001',
contents='What is the weather like in Boston?',
config=types.GenerateContentConfig(tools=[get_current_weather]),
)
print(response.text)

Disabling automatic function calling

If you pass in a python function as a tool directly, and do not want automatic function calling, you can disable automatic function calling as follows:

fromgoogle.genaiimporttypesresponse=client.models.generate_content(
model='gemini-2.0-flash-001',
contents='What is the weather like in Boston?',
config=types.GenerateContentConfig(
tools=[get_current_weather],
automatic_function_calling=types.AutomaticFunctionCallingConfig(
disable=True
),
),
)

With automatic function calling disabled, you will get a list of function call parts in the response:

function_calls: Optional[List[types.FunctionCall]] =response.function_calls

Manually declare and invoke a function for function calling

If you don't want to use the automatic function support, you can manually declare the function and invoke it.

The following example shows how to declare a function and pass it as a tool. Then you will receive a function call part in the response.

fromgoogle.genaiimporttypesfunction=types.FunctionDeclaration(
name='get_current_weather',
description='Get the current weather in a given location',
parameters=types.Schema(
type='OBJECT',
properties={
'location': types.Schema(
type='STRING',
description='The city and state, e.g. San Francisco, CA',
),
},
required=['location'],
),
)
tool=types.Tool(function_declarations=[function])
response=client.models.generate_content(
model='gemini-2.0-flash-001',
contents='What is the weather like in Boston?',
config=types.GenerateContentConfig(tools=[tool]),
)
print(response.function_calls[0])

After you receive the function call part from the model, you can invoke the function and get the function response. And then you can pass the function response to the model. The following example shows how to do it for a simple function invocation.

fromgoogle.genaiimporttypesuser_prompt_content=types.Content(
role='user',
parts=[types.Part.from_text(text='What is the weather like in Boston?')],
)
function_call_part=response.function_calls[0]
function_call_content=response.candidates[0].contenttry:
function_result=get_current_weather(
**function_call_part.function_call.args
)
function_response= {'result': function_result}
except (
Exception
) ase: # instead of raising the exception, you can let the model handle itfunction_response= {'error': str(e)}
function_response_part=types.Part.from_function_response(
name=function_call_part.name,
response=function_response,
)
function_response_content=types.Content(
role='tool', parts=[function_response_part]
)
response=client.models.generate_content(
model='gemini-2.0-flash-001',
contents=[
user_prompt_content,
function_call_content,
function_response_content,
],
config=types.GenerateContentConfig(
tools=[tool],
),
)
print(response.text)

Function calling with ANY tools config mode

If you configure function calling mode to be ANY, then the model will always return function call parts. If you also pass a python function as a tool, by default the SDK will perform automatic function calling until the remote calls exceed the maximum remote call for automatic function calling (default to 10 times).

If you'd like to disable automatic function calling in ANY mode:

fromgoogle.genaiimporttypesdefget_current_weather(location: str) ->str:
"""Returns the current weather. Args: location: The city and state, e.g. San Francisco, CA """return"sunny"response=client.models.generate_content(
model="gemini-2.0-flash-001",
contents="What is the weather like in Boston?",
config=types.GenerateContentConfig(
tools=[get_current_weather],
automatic_function_calling=types.AutomaticFunctionCallingConfig(
disable=True
),
tool_config=types.ToolConfig(
function_calling_config=types.FunctionCallingConfig(mode='ANY')
),
),
)

If you'd like to set x number of automatic function call turns, you can configure the maximum remote calls to be x + 1. Assuming you prefer 1 turn for automatic function calling.

fromgoogle.genaiimporttypesdefget_current_weather(location: str) ->str:
"""Returns the current weather. Args: location: The city and state, e.g. San Francisco, CA """return"sunny"response=client.models.generate_content(
model="gemini-2.0-flash-001",
contents="What is the weather like in Boston?",
config=types.GenerateContentConfig(
tools=[get_current_weather],
automatic_function_calling=types.AutomaticFunctionCallingConfig(
maximum_remote_calls=2
),
tool_config=types.ToolConfig(
function_calling_config=types.FunctionCallingConfig(mode='ANY')
),
),
)

JSON Response Schema

However you define your schema, don't duplicate it in your input prompt, including by giving examples of expected JSON output. If you do, the generated output might be lower in quality.

Pydantic Model Schema support

Schemas can be provided as Pydantic Models.

frompydanticimportBaseModelfromgoogle.genaiimporttypesclassCountryInfo(BaseModel):
name: strpopulation: intcapital: strcontinent: strgdp: intofficial_language: strtotal_area_sq_mi: intresponse=client.models.generate_content(
model='gemini-2.0-flash-001',
contents='Give me information for the United States.',
config=types.GenerateContentConfig(
response_mime_type='application/json',
response_schema=CountryInfo,
),
)
print(response.text)

fromgoogle.genaiimporttypesresponse=client.models.generate_content(
model='gemini-2.0-flash-001',
contents='Give me information for the United States.',
config=types.GenerateContentConfig(
response_mime_type='application/json',
response_schema={
'required': [
'name',
'population',
'capital',
'continent',
'gdp',
'official_language',
'total_area_sq_mi',
],
'properties': {
'name': {'type': 'STRING'},
'population': {'type': 'INTEGER'},
'capital': {'type': 'STRING'},
'continent': {'type': 'STRING'},
'gdp': {'type': 'INTEGER'},
'official_language': {'type': 'STRING'},
'total_area_sq_mi': {'type': 'INTEGER'},
},
'type': 'OBJECT',
},
),
)
print(response.text)

Enum Response Schema

Text Response

You can set response_mime_type to 'text/x.enum' to return one of those enum values as the response.

classInstrumentEnum(Enum):
PERCUSSION='Percussion'STRING='String'WOODWIND='Woodwind'BRASS='Brass'KEYBOARD='Keyboard'response=client.models.generate_content(
model='gemini-2.0-flash-001',
contents='What instrument plays multiple notes at once?',
config={
'response_mime_type': 'text/x.enum',
'response_schema': InstrumentEnum,
},
)
print(response.text)

JSON Response

You can also set response_mime_type to 'application/json', the response will be identical but in quotes.

fromenumimportEnumclassInstrumentEnum(Enum):
PERCUSSION='Percussion'STRING='String'WOODWIND='Woodwind'BRASS='Brass'KEYBOARD='Keyboard'response=client.models.generate_content(
model='gemini-2.0-flash-001',
contents='What instrument plays multiple notes at once?',
config={
'response_mime_type': 'application/json',
'response_schema': InstrumentEnum,
},
)
print(response.text)

Generate Content (Synchronous Streaming)

Generate content in a streaming format so that the model outputs streams back to you, rather than being returned as one chunk.

Streaming for text content

forchunkinclient.models.generate_content_stream(
model='gemini-2.0-flash-001', contents='Tell me a story in 300 words.'
):
print(chunk.text, end='')

Streaming for image content

If your image is stored in Google Cloud Storage, you can use the from_uri class method to create a Part object.

fromgoogle.genaiimporttypesforchunkinclient.models.generate_content_stream(
model='gemini-2.0-flash-001',
contents=[
'What is this image about?',
types.Part.from_uri(
file_uri='gs://generativeai-downloads/images/scones.jpg',
mime_type='image/jpeg',
),
],
):
print(chunk.text, end='')

If your image is stored in your local file system, you can read it in as bytes data and use the from_bytes class method to create a Part object.

fromgoogle.genaiimporttypesYOUR_IMAGE_PATH='your_image_path'YOUR_IMAGE_MIME_TYPE='your_image_mime_type'withopen(YOUR_IMAGE_PATH, 'rb') asf:
image_bytes=f.read()
forchunkinclient.models.generate_content_stream(
model='gemini-2.0-flash-001',
contents=[
'What is this image about?',
types.Part.from_bytes(data=image_bytes, mime_type=YOUR_IMAGE_MIME_TYPE),
],
):
print(chunk.text, end='')

Generate Content (Asynchronous Non Streaming)

client.aio exposes all the analogous async methods that are available on client. Note that it applies to all the modules.

For example, client.aio.models.generate_content is the async version of client.models.generate_content

response=awaitclient.aio.models.generate_content(
model='gemini-2.0-flash-001', contents='Tell me a story in 300 words.'
)
print(response.text)

Generate Content (Asynchronous Streaming)

asyncforchunkinawaitclient.aio.models.generate_content_stream(
model='gemini-2.0-flash-001', contents='Tell me a story in 300 words.'
):
print(chunk.text, end='')

Count Tokens and Compute Tokens

response=client.models.count_tokens(
model='gemini-2.0-flash-001',
contents='why is the sky blue?',
)
print(response)

Compute Tokens

Compute tokens is only supported in Vertex AI.

response=client.models.compute_tokens(
model='gemini-2.0-flash-001',
contents='why is the sky blue?',
)
print(response)

Async

response=awaitclient.aio.models.count_tokens(
model='gemini-2.0-flash-001',
contents='why is the sky blue?',
)
print(response)

Embed Content

response=client.models.embed_content(
model='text-embedding-004',
contents='why is the sky blue?',
)
print(response)

fromgoogle.genaiimporttypes# multiple contents with configresponse=client.models.embed_content(
model='text-embedding-004',
contents=['why is the sky blue?', 'What is your age?'],
config=types.EmbedContentConfig(output_dimensionality=10),
)
print(response)

Imagen

Generate Images

Support for generate images in Gemini Developer API is behind an allowlist

fromgoogle.genaiimporttypes# Generate Imageresponse1=client.models.generate_images(
model='imagen-3.0-generate-002',
prompt='An umbrella in the foreground, and a rainy night sky in the background',
config=types.GenerateImagesConfig(
number_of_images=1,
include_rai_reason=True,
output_mime_type='image/jpeg',
),
)
response1.generated_images[0].image.show()

Upscale Image

Upscale image is only supported in Vertex AI.

fromgoogle.genaiimporttypes# Upscale the generated image from aboveresponse2=client.models.upscale_image(
model='imagen-3.0-generate-001',
image=response1.generated_images[0].image,
upscale_factor='x2',
config=types.UpscaleImageConfig(
include_rai_reason=True,
output_mime_type='image/jpeg',
),
)
response2.generated_images[0].image.show()

Edit Image

Edit image uses a separate model from generate and upscale.

Edit image is only supported in Vertex AI.

# Edit the generated image from abovefromgoogle.genaiimporttypesfromgoogle.genai.typesimportRawReferenceImage, MaskReferenceImageraw_ref_image=RawReferenceImage(
reference_id=1,
reference_image=response1.generated_images[0].image,
)
# Model computes a mask of the backgroundmask_ref_image=MaskReferenceImage(
reference_id=2,
config=types.MaskReferenceConfig(
mask_mode='MASK_MODE_BACKGROUND',
mask_dilation=0,
),
)
response3=client.models.edit_image(
model='imagen-3.0-capability-001',
prompt='Sunlight and clear sky',
reference_images=[raw_ref_image, mask_ref_image],
config=types.EditImageConfig(
edit_mode='EDIT_MODE_INPAINT_INSERTION',
number_of_images=1,
include_rai_reason=True,
output_mime_type='image/jpeg',
),
)
response3.generated_images[0].image.show()

Veo

Generate Videos

Support for generate videos in Vertex and Gemini Developer API is behind an allowlist

fromgoogle.genaiimporttypes# Create operationoperation=client.models.generate_videos(
model='veo-2.0-generate-001',
prompt='A neon hologram of a cat driving at top speed',
config=types.GenerateVideosConfig(
number_of_videos=1,
fps=24,
duration_seconds=5,
enhance_prompt=True,
),
)
# Poll operationwhilenotoperation.done:
time.sleep(20)
operation=client.operations.get(operation)
video=operation.result.generated_videos[0].videovideo.show()

Chats

Create a chat session to start a multi-turn conversations with the model. Then, use chat.send_message function multiple times within the same chat session so that it can reflect on its previous responses (i.e., engage in an ongoing conversation). See the 'Create a client' section above to initialize a client.

Send Message (Synchronous Non-Streaming)

chat=client.chats.create(model='gemini-2.0-flash-001')
response=chat.send_message('tell me a story')
print(response.text)
response=chat.send_message('summarize the story you told me in 1 sentence')
print(response.text)

Send Message (Synchronous Streaming)

chat=client.chats.create(model='gemini-2.0-flash-001')
forchunkinchat.send_message_stream('tell me a story'):
print(chunk.text)

Send Message (Asynchronous Non-Streaming)

chat=client.aio.chats.create(model='gemini-2.0-flash-001')
response=awaitchat.send_message('tell me a story')
print(response.text)

Send Message (Asynchronous Streaming)

chat=client.aio.chats.create(model='gemini-2.0-flash-001')
asyncforchunkinawaitchat.send_message_stream('tell me a story'):
print(chunk.text)

Files

Files are only supported in Gemini Developer API. See the 'Create a client' section above to initialize a client.

!gsutil cp gs://cloud-samples-data/generative-ai/pdf/2312.11805v3.pdf .
!gsutil cp gs://cloud-samples-data/generative-ai/pdf/2403.05530.pdf .

Upload

file1=client.files.upload(file='2312.11805v3.pdf')
file2=client.files.upload(file='2403.05530.pdf')
print(file1)
print(file2)

Get

file1=client.files.upload(file='2312.11805v3.pdf')
file_info=client.files.get(name=file1.name)

Delete

file3=client.files.upload(file='2312.11805v3.pdf')
client.files.delete(name=file3.name)

Caches

client.caches contains the control plane APIs for cached content. See the 'Create a client' section above to initialize a client.

Create

fromgoogle.genaiimporttypesifclient.vertexai:
file_uris= [
'gs://cloud-samples-data/generative-ai/pdf/2312.11805v3.pdf',
'gs://cloud-samples-data/generative-ai/pdf/2403.05530.pdf',
]
else:
file_uris= [file1.uri, file2.uri]
cached_content=client.caches.create(
model='gemini-2.0-flash-001',
config=types.CreateCachedContentConfig(
contents=[
types.Content(
role='user',
parts=[
types.Part.from_uri(
file_uri=file_uris[0], mime_type='application/pdf'
),
types.Part.from_uri(
file_uri=file_uris[1],
mime_type='application/pdf',
),
],
)
],
system_instruction='What is the sum of the two pdfs?',
display_name='test cache',
ttl='3600s',
),
)

Get

cached_content=client.caches.get(name=cached_content.name)

Generate Content with Caches

fromgoogle.genaiimporttypesresponse=client.models.generate_content(
model='gemini-2.0-flash-001',
contents='Summarize the pdfs',
config=types.GenerateContentConfig(
cached_content=cached_content.name,
),
)
print(response.text)

Tunings

client.tunings contains tuning job APIs and supports supervised fine tuning through tune. See the 'Create a client' section above to initialize a client.

Tune

• Vertex AI supports tuning from GCS source
• Gemini Developer API supports tuning from inline examples

fromgoogle.genaiimporttypesifclient.vertexai:
model='gemini-2.0-flash-001'training_dataset=types.TuningDataset(
gcs_uri='gs://cloud-samples-data/ai-platform/generative_ai/gemini-1_5/text/sft_train_data.jsonl',
)
else:
model='models/gemini-2.0-flash-001'training_dataset=types.TuningDataset(
examples=[
types.TuningExample(
text_input=f'Input text {i}',
output=f'Output text {i}',
)
foriinrange(5)
],
)

fromgoogle.genaiimporttypestuning_job=client.tunings.tune(
base_model=model,
training_dataset=training_dataset,
config=types.CreateTuningJobConfig(
epoch_count=1, tuned_model_display_name='test_dataset_examples model'
),
)
print(tuning_job)

Get Tuning Job

tuning_job=client.tunings.get(name=tuning_job.name)
print(tuning_job)

importtimerunning_states=set(
[
'JOB_STATE_PENDING',
'JOB_STATE_RUNNING',
]
)
whiletuning_job.stateinrunning_states:
print(tuning_job.state)
tuning_job=client.tunings.get(name=tuning_job.name)
time.sleep(10)

Use Tuned Model

response=client.models.generate_content(
model=tuning_job.tuned_model.endpoint,
contents='why is the sky blue?',
)
print(response.text)

Get Tuned Model

tuned_model=client.models.get(model=tuning_job.tuned_model.model)
print(tuned_model)

List Tuned Models

To retrieve base models, see list base models.

formodelinclient.models.list(config={'page_size': 10, 'query_base': False}):
print(model)

pager=client.models.list(config={'page_size': 10, 'query_base': False})
print(pager.page_size)
print(pager[0])
pager.next_page()
print(pager[0])

Async

asyncforjobinawaitclient.aio.models.list(config={'page_size': 10, 'query_base': False}):
print(job)

async_pager=awaitclient.aio.models.list(config={'page_size': 10, 'query_base': False})
print(async_pager.page_size)
print(async_pager[0])
awaitasync_pager.next_page()
print(async_pager[0])

Update Tuned Model

fromgoogle.genaiimporttypesmodel=pager[0]
model=client.models.update(
model=model.name,
config=types.UpdateModelConfig(
display_name='my tuned model', description='my tuned model description'
),
)
print(model)

List Tuning Jobs

forjobinclient.tunings.list(config={'page_size': 10}):
print(job)

pager=client.tunings.list(config={'page_size': 10})
print(pager.page_size)
print(pager[0])
pager.next_page()
print(pager[0])

Async

asyncforjobinawaitclient.aio.tunings.list(config={'page_size': 10}):
print(job)

async_pager=awaitclient.aio.tunings.list(config={'page_size': 10})
print(async_pager.page_size)
print(async_pager[0])
awaitasync_pager.next_page()
print(async_pager[0])

Batch Prediction

Only supported in Vertex AI. See the 'Create a client' section above to initialize a client.

Create

# Specify model and source file only, destination and job display name will be auto-populatedjob=client.batches.create(
model='gemini-2.0-flash-001',
src='bq://my-project.my-dataset.my-table',
)
job

# Get a job by namejob=client.batches.get(name=job.name)
job.state

completed_states=set(
[
'JOB_STATE_SUCCEEDED',
'JOB_STATE_FAILED',
'JOB_STATE_CANCELLED',
'JOB_STATE_PAUSED',
]
)
whilejob.statenotincompleted_states:
print(job.state)
job=client.batches.get(name=job.name)
time.sleep(30)
job

List

forjobinclient.batches.list(config=types.ListBatchJobsConfig(page_size=10)):
print(job)

pager=client.batches.list(config=types.ListBatchJobsConfig(page_size=10))
print(pager.page_size)
print(pager[0])
pager.next_page()
print(pager[0])

Async

asyncforjobinawaitclient.aio.batches.list(
config=types.ListBatchJobsConfig(page_size=10)
):
print(job)

async_pager=awaitclient.aio.batches.list(
config=types.ListBatchJobsConfig(page_size=10)
)
print(async_pager.page_size)
print(async_pager[0])
awaitasync_pager.next_page()
print(async_pager[0])

Delete

# Delete the job resourcedelete_job=client.batches.delete(name=job.name)
delete_job

Error Handling

To handle errors raised by the model service, the SDK provides this APIError class.

fromgoogle.genaiimporterrorstry:
client.models.generate_content(
model="invalid-model-name",
contents="What is your name?",
)
excepterrors.APIErrorase:
print(e.code) # 404print(e.message)