google-cloud-aiplatform

google-cloud-aiplatform (also imported as vertexai) is Google Cloud’s official Python client library for Vertex AI, the company’s unified machine-learning and generative-AI platform. It wraps the underlying gRPC/REST APIs for dataset management, custom and AutoML model training, batch and online prediction, model registry, pipelines, feature store, and explainability behind a Pythonic client interface, and layers a dedicated vertexai module on top for Gemini-based generative AI: chat, evaluation, tuning, agent orchestration (Agent Engine/agentplatform), and RAG retrieval.

As one of the highest-traffic packages in the Google Cloud Python ecosystem, it’s the primary entry point for teams building production ML and GenAI applications on Google Cloud, spanning classic MLOps pipelines through to the newest Gemini agent-runtime features.

What You Get

• A high-level aiplatform.init() + resource-class API for datasets, custom/AutoML training jobs, models, endpoints, and batch prediction
• A vertexai.Client for Gemini chat, text/image/multimodal generation, and prompt-based generative AI use cases
• Built-in Gen AI evaluation (client.evals) for scoring model outputs against reference data with configurable metrics
• Feature Store, Vector Search (matching engine), Pipelines, Model Registry, and Explainable AI integrations
• Agent Engine / agentplatform runtime bindings for deploying and orchestrating LLM agents on Vertex AI
• Optional extras for TensorBoard profiling, LIT, Ray on Vertex, and framework integrations (LangChain, LlamaIndex)

Common Use Cases

• Training and deploying custom or AutoML models to Vertex AI endpoints for online/batch prediction
• Building generative AI applications with Gemini models, including chat, structured output, and multimodal prompts
• Running automated evaluation pipelines to score and compare LLM outputs across metrics and datasets
• Managing ML pipelines, feature stores, and model registries as part of a production MLOps workflow on Google Cloud
• Deploying and orchestrating LLM agents through the Agent Engine runtime for enterprise agent platforms

Under The Hood

Architecture The package is a monorepo-style client spanning several generations of API surface: the oldest layer is auto-generated GAPIC gRPC/REST clients under google/cloud/aiplatform_v1(beta1), wrapped by a hand-written SDK layer in google/cloud/aiplatform/ that exposes resource-oriented classes (Model, Endpoint, Dataset, PipelineJob, etc.) built on a shared base.VertexAiResourceNoun pattern. The newest layer, vertexai/, is a from-scratch redesign centered on a single vertexai.Client object exposing namespaced sub-clients (.models, .evals, .agent_engines) for Gemini-era generative AI, deliberately decoupled from the older aiplatform resource classes to give a cleaner, more Pythonic surface for GenAI workflows. Cross-cutting concerns (auth, project/location resolution, request retries) are centralized in shared base/init modules that both layers call into. Tech Stack Pure Python (99.97% of the codebase) targeting Python 3.8+, built on google-api-core, google-auth, grpcio, and proto-plus for the generated clients, with optional heavy dependencies (pandas, pyarrow, tensorflow, docker) gated behind extras in setup.py so a minimal install stays lightweight. Packaging uses classic setuptools with owlbot.py/synth.py generation tooling typical of Google’s API client generator pipeline, and release automation via release-please. Code Quality The tests/ tree separates unit/ (mirroring the source layout, including dedicated suites per generative-AI integration like vertex_langchain, vertex_llama_index, vertex_rag) from system/ end-to-end tests that hit live Vertex AI resources; this split, combined with noxfile.py-driven multi-version test sessions, indicates a mature CI discipline consistent with Google’s other google-cloud-python client libraries. API Design The API intentionally maintains two coexisting idioms — the older resource-class style (aiplatform.Model.upload(...)) for classic MLOps and the newer vertexai.Client(project=..., location=...) singleton style for GenAI — which lowers boilerplate for new Gemini-focused use cases but means developers must learn which surface applies to a given task; extensive docstrings and the dual README (classic + GenAI quickstarts) mitigate this.