Sara Robinson

This commit **adds support for `UnifiedMetric`** objects within the **Vertex Tuning evaluation configuration**, providing a **new capability** for specifying evaluation metrics. It introduces new data models and enums for `UnifiedMetric` in `google/genai/types.py`, updating the `EvaluationConfig` to accept these new metric types. The `t_metrics` function in `google/genai/_transformers.py` is enhanced to process, validate, and generate payloads for `UnifiedMetric` objects. This change **improves the flexibility and standardization** of evaluation metric definitions for Vertex AI tunings, with new tests in `google/genai/tests/tunings/test_tune.py` ensuring correct behavior.

This commit is a **maintenance chore** that integrates **static type checking** into the project's continuous integration via **GitHub Actions**. It introduces two new workflows: `mypy.yml` to run comprehensive `mypy` checks on every pull request, and `mypy-new-errors.yml` to specifically **prevent the introduction of new type errors** by comparing against the `main` branch. This enhancement to the **developer tooling** establishes a robust mechanism for **code quality enforcement**, aiming to improve the overall **reliability and maintainability** of the Python codebase by catching type inconsistencies early.

This commit introduces **support for the `gemini-3-pro-preview` model** within the **local tokenizer** component of the `google.genai` library. It specifically **adds `gemini-3-pro-preview` to the model-to-tokenizer mapping** in `google/genai/_local_tokenizer_loader.py`, enabling the system to correctly identify and load the appropriate tokenizer for this new model. This **new capability** allows developers to perform offline tokenization for `gemini-3-pro-preview` without requiring an API call. The change also includes an update to `google/genai/tests/local_tokenizer/test_local_tokenizer.py` to validate this new model's integration.

This commit **regenerates the documentation** for the **Generative AI (GenAI)** module, specifically updating the **API documentation** to reflect recent changes. It introduces comprehensive documentation for the new `ReplicatedVoiceConfig` type and updates existing definitions for `SpeechConfig` and `VoiceConfig` to include this new configuration. This **documentation update** ensures that users have accurate and up-to-date information regarding voice configuration capabilities within the GenAI module, improving discoverability and understanding of new features.

This commit delivers a **bug fix** to the **`google.genai` library**, specifically enabling the **`APIError` class to be picklable**, thereby resolving issue #1144. Previously, instances of `APIError` could not be serialized, which prevented their use in scenarios requiring **inter-process communication** or **object persistence**. The fix involves adding `__reduce__` and `_rebuild` methods to the `APIError` class within `google/genai/errors.py` to correctly handle its serialization and deserialization. A new test case in `google/genai/tests/errors/test_api_error.py` has been added to verify this critical **serialization capability**.

This commit **fixes a data parsing issue** within the **`google.genai.types` module** by introducing a Pydantic model validator to the `CitationMetadata` class. Specifically, it ensures that if the `citationSources` key is present in incoming data, it is automatically **converted and renamed to `citations`** during instantiation. This **enhances the robustness of `GenerateContentResponse` parsing** by handling legacy or alternative data formats, preventing potential errors and ensuring consistent access to citation information. A new test case has been added to `google/genai/tests/types/test_types.py` to verify this critical data transformation.

This commit **improves user guidance** within the **`google.genai` library** by updating an error message and relevant docstrings in `google/genai/types.py`. It now explicitly recommends using `response_json_schema` as an alternative when `response_schema` fails to process correctly. This **documentation enhancement** helps developers troubleshoot schema-related issues more effectively, particularly concerning response schema processing. The change aims to clarify common pitfalls and guide users towards the correct API usage, thereby improving the overall **developer experience** for the `google.genai` library.

This commit **improves the documentation** for the **`google.genai.types`** module by updating the `README.md` file. It introduces a **new section** that provides clear explanations and examples for effectively using the **Pydantic types** available within this module. Concurrently, an outdated 'Typed Config' section has been removed to streamline the content. This **documentation update** aims to offer clearer and more current guidance for developers working with `google.genai.types`, enhancing usability and understanding.

This commit provides a **bug fix** for the **Live API** by addressing a potential `ImportError` related to `google.auth.transport.requests`. It introduces a `try-except` block within `google/genai/live.py` to gracefully handle scenarios where this module might be unavailable during credential refresh operations. If the necessary module is missing, a more informative `ValueError` is now raised, making the **Live API**'s credential handling more robust and providing clearer diagnostics for users.

This commit delivers a **bug fix** to refine the logging behavior within the **`google.genai` library**, specifically for **accessors** in `GenerateContentResponse` and `LiveServerMessage` objects. It introduces global flags to ensure that warnings related to accessing properties like `text`, `parts`, `function_calls`, and others are logged only once, preventing redundant messages. This change improves the clarity of SDK output by eliminating repetitive warnings when these properties are accessed multiple times. Accompanying test updates verify that this single-logging mechanism functions as intended.

This commit introduces a **new capability** to the **Automatic Function Calling (AFC) utility** by adding **support for abstract types** in function signatures. The `from_function_with_options` function is refactored to handle these complex types more robustly, specifically by falling back to **Pydantic for JSON schema generation** when direct parsing fails. New utility functions are added in `src/google/adk/tools/_function_parameter_parse_util.py` to manage deferred annotations and provide improved error handling for various parameter types. This enhancement significantly **improves the flexibility and robustness of AFC** when processing functions with diverse and abstract type hints.

This commit **enhances the Automatic Function Calling (AFC) utility** by introducing support for **abstract type annotations and complex parameter types**. The `from_function_with_options` function in `src/google/adk/tools/_automatic_function_calling_util.py` now leverages Pydantic-based JSON schema generation as a fallback mechanism to handle these advanced type definitions. This **new capability** significantly improves the robustness and flexibility of AFC, allowing it to correctly parse and generate schemas for functions with more sophisticated parameter and return value signatures. New utility functions in `src/google/adk/tools/_function_parameter_parse_util.py` were added to manage deferred annotations and generate these Pydantic-based schemas.

This commit **updates the documentation** within the **Generative AI Python client library**, specifically in `google/genai/types.py`. It adds detailed descriptions for the new `evaluation_config` field within the `TuningJob` and `TuningJobDict` data structures, which are crucial for configuring **model tuning jobs**. This **documentation update** clarifies the expected format and usage of `EvaluationConfig` when passed to the `tune()` functionality, thereby preparing the Python type definitions for its upcoming functional support. Additionally, it refines the description for the publisher model format in `AutoraterConfig`.

This commit **fixes** an issue in the **`google.genai` library** by refining the **warning mechanism** for accessing response text. It modifies the `_get_text` method within `google/genai/types.py` to ensure warnings are only logged when `response.text` is directly accessed and the content is non-textual or multi-candidate. This change **suppresses unnecessary warnings** when structured responses are being parsed by `_parse_response_content`, preventing log spam. The **response handling** subsystem is improved by providing more targeted and relevant warnings to developers, enhancing the overall developer experience.

This commit introduces a **new capability** to the `google.genai` library by adding **private methods and types** for fetching **project-level operations from Vertex AI**. It primarily modifies `google/genai/operations.py` to include helper functions like `_GetProjectOperationParameters_to_vertex`, `_ProjectOperation_from_vertex`, and the core `_get` method. New types, `_GetProjectOperationParameters` and `ProjectOperation`, are defined in `google/genai/types.py` to support these operations. This enhancement improves the **Vertex AI integration** by enabling programmatic access to project-scoped operations, with new unit tests added in `google/genai/tests/operations/test_get.py` to validate the functionality.

This commit introduces a **new capability** to the **`genai` library**, enabling users to specify `labels` when creating tuning job configurations. An optional `labels` field has been added to the `CreateTuningJobConfig` class in `google/genai/types.py`, allowing for custom metadata attachment. The `google/genai/tunings.py` module now includes validation for these labels and ensures they are correctly mapped for both **Gemini API** and **Vertex AI** tuning job configurations via `_CreateTuningJobConfig_to_mldev` and `_CreateTuningJobConfig_to_vertex`. This enhancement provides better organization and management of tuning jobs, with a new test case in `google/genai/tests/tunings/test_tune.py` verifying the functionality.

This commit performs **code cleanup** by **removing several unused `Metric` types** from the project's **telemetry and metrics collection system**. This **maintenance** task targets definitions that are no longer referenced, streamlining the codebase. The change improves overall code clarity and reduces potential bloat within the **metrics infrastructure**, without affecting any active functionality or data collection.

This commit introduces a **new capability** to support **evaluation configurations** for **Vertex AI model tuning jobs** within the `google.genai` library. It defines new Pydantic models and TypedDicts in `google/genai/types.py` for various evaluation settings, including `GcsDestination` and `AutoraterConfig`, and adds transformation functions like `t_metrics` in `google/genai/_transformers.py` and `_EvaluationConfig_to_vertex` in `google/genai/tunings.py` to convert these into Vertex AI-compatible payloads. The `tune` function in `google/genai/tunings.py` is updated to accept these configurations, allowing users to specify integrated evaluation metrics directly when initiating a tuning job. This significantly enhances the **model tuning workflow** by enabling comprehensive performance assessment as part of the tuning process.

This commit **integrates new type definitions and enums** from a discovery document into the **Generative AI tuning subsystem**, primarily within `google/genai/types.py`. These additions support advanced features such as custom base models, preference optimization, VEO tuning specifications, and Google Maps grounding. Concurrently, the `google/genai/tunings.py` module sees **updates to conversion functions** like `_TuningJob_from_mldev` and `_TuningJob_from_vertex` to accommodate these new fields. This work represents a **feature addition** that expands tuning capabilities, alongside **maintenance** by removing deprecated types like `Environment` and `ToolComputerUse`.

This commit introduces a **feature enhancement** to the **Google GenAI SDK's automatic function calling**, specifically expanding the capabilities of `FunctionDeclaration.from_callable`. It now supports a wider range of Python type hints for function parameters and return values, including union operators, tuple parameters, and fixed-length tuple return types with `unevaluatedItems`. New utility functions in `google/genai/_automatic_function_calling_util.py` improve schema generation and error handling for unsupported types. This allows developers to define more sophisticated and accurately typed functions for **automatic function calling**, with comprehensive new tests validating these expanded capabilities in `google/genai/tests/`.