repository·main·Indexed 25 days ago
https://github.com/google-gemini/deprecated-generative-ai-pythonLegacy Python SDK for the Gemini API. This repository has reached end-of-life as of November 30, 2025, and is superseded by the unified Google Gen AI SDK. It provides functionality for content generation via GenerativeModel, multi-turn conversations with ChatSession, embeddings, file management, model tuning, and context caching.
This repository is considered legacy and has reached its end-of-life.
Support Status:
The google.generativeai package contains the core API for interacting with Gemini models. Key functional areas include:
GenerativeModel for content generation and ChatSession for multi-turn conversations.configure and managing generation parameters with GenerationConfig.upload_file, list_files, and delete_file.embed_content and embed_content_async.create_tuned_model, list_tuned_models, and update_tuned_model.caching modules and CachedContent for efficient context reuse.get_*, list_*, and delete_* methods.The CodeExecutionResult object represents the outcome of executing ExecutableCode. It is only generated when using the CodeExecution feature and always follows a part containing the ExecutableCode in the model's response.
Use this object to inspect whether the code executed successfully and to retrieve the resulting output or error messages.
By default, all keys defined in a TypedDict are required. You can control key presence using two methods:
total argumentWhen defining the class, you can set total=False to make all keys optional. If total=True (the default), all keys must be present.
class Point2D(TypedDict, total=False):
x: int
y: int
# Any of the keys can be omittedRequired and NotRequired annotationsFor more granular control, use Required and NotRequired (per PEP 655) to specify requirements on a per-key basis within a class.
from typing import NotRequired
class Point2D(TypedDict):
x: int # Required by default
y: NotRequired[int] # Can be omittedThe library provides several types to inspect and manage content safety:
BlockedReason: A list of reasons why content may have been blocked.HarmCategory: Supported harm categories for Gemini-family models.HarmProbability: Represents the probability that a piece of content is harmful.HarmBlockThreshold: Defines the threshold at and beyond which content should be blocked.SafetySettingDict: Used to configure safety-blocking behavior.The Permission class provides asynchronous versions of its primary operations for use in async environments:
get_async(name, client=None): Asynchronous version of get.update_async(updates, client=None): Asynchronous version of update.delete_async(client=None): Asynchronous version of delete.Document is a collection of Chunk objects. Documents are organized within a Corpus, and a single Corpus can contain a maximum of 10,000 Documents.A Tool is an abstraction that enables the Gemini model to interact with external systems or perform actions outside its internal knowledge base. You can provide a Tool to the model to enable one of three primary capabilities:
function_declarations, you define a list of functions the model can request to call. Note that the model does not execute the code itself. Instead, it returns a FunctionCall containing the arguments, which your client-side application must execute. You then provide the result back to the model via a FunctionResponse in the next conversation turn.google_search_retrieval, you enable a retrieval tool powered by Google Search to augment the model's responses.code_execution, you enable the model to execute code as part of its generation process.Each of these attributes is optional.
The google.generativeai.protos.ExecutableCode object represents code generated by the model that is intended to be executed, along with the result returned to the model.
This object is only generated when using the CodeExecution tool. When this tool is active, the code is automatically executed by the system, and a corresponding CodeExecutionResult is generated.
The google.generativeai.protos.CodeExecution tool allows the Gemini model to generate and execute code automatically. When this tool is enabled, the model can write code, run it in a sandboxed environment, and receive the execution results back to inform its next response.
When using this tool, the model's output may include ExecutableCode (the code generated) and CodeExecutionResult (the output/result of that code).
The training_data argument accepts several formats to define the dataset for tuning:
protos.Dataset: A protocol buffer dataset object.Iterable of examples: An iterable containing:protos.TuningExample objects.{'text_input': text_input, 'output': output}.(text_input, output).Mapping of Iterable[str]: A dictionary where keys represent columns. Use input_key and output_key to specify which columns to use as input and output.str or pathlib.Path), a URL, or a Google Sheets URL. These are processed as a Mapping using pd.read_csv.str or pathlib.Path). The contents are handled as either an Iterable or a Mapping based on the structure.google.generativeai.protos.CitationSource object represents a citation to a specific source for a portion of a model response. It is used to attribute segments of text or code to their original origins, providing metadata such as the source URI and licensing information.