canonical: https://jentic.com/apis/googleapis.com/aiplatform # Google Vertex AI API The Vertex AI API is Google Cloud's unified surface for training, tuning, deploying, and serving machine learning models, including Google's foundation models such as Gemini and PaLM and customer-trained models. It exposes operations on datasets, training pipelines, models, endpoints, batch prediction jobs, feature stores, indexes, and model lineage. With 202 endpoints, it covers the full MLOps lifecycle from data ingestion through online and batch inference. It is the right tool for teams building production ML systems on Google Cloud rather than just calling a hosted LLM endpoint. ## For AI agents Train, deploy, and call ML models on Google Cloud - including Gemini foundation models - via 202 endpoints covering datasets, pipelines, endpoints, and predictions. ## Scope Does not handle BigQuery analytics queries, Workspace data, or non-ML Cloud resource provisioning - use for Vertex AI model training, deployment, and prediction only. ## Capabilities - Run online predictions against a deployed Vertex AI endpoint, including Gemini and custom models - Submit and monitor training pipelines that produce versioned Vertex AI model resources - Deploy a model to a Vertex AI endpoint and split traffic between deployed model versions - Run batch prediction jobs over BigQuery or Cloud Storage inputs and write results back - Manage datasets, dataset items, and annotation specs that feed AutoML and custom training - Search artifact and context lineage subgraphs to trace how a deployed model was produced - Manage feature stores, feature views, and indexes that back online inference and vector search ## Use cases ### Foundation Model Inference at Scale Teams that want to call Gemini or other Vertex foundation models from production code use the predict and streamGenerateContent endpoints rather than the consumer Gemini API, because Vertex bills against a Google Cloud project and supports IAM, VPC Service Controls, and regional residency. Vertex AI exposes both online prediction for low-latency inference and batch prediction for high-throughput jobs over BigQuery or GCS inputs. Example prompt: POST to projects/{project}/locations/{location}/publishers/google/models/gemini-1.5-pro:generateContent with the prompt and read the response in JSON. ### Custom Training Pipeline Orchestration Data science teams use Vertex AI training pipelines to package their training code in a container, run it on managed compute, and produce a versioned Vertex AI Model resource that can be deployed to an endpoint. The API covers creating, listing, cancelling, and inspecting CustomJobs, TrainingPipelines, and HyperparameterTuningJobs, replacing bespoke Kubernetes setups for ML training. Example prompt: Create a TrainingPipeline under projects/{project}/locations/{location}/trainingPipelines pointing at a container image and Cloud Storage training inputs, then poll its state until it produces a Model resource. ### Vector Search for Retrieval-Augmented Generation Applications using RAG store document embeddings in a Vertex AI Index and query the deployed IndexEndpoint to find nearest neighbours at request time. Vertex AI handles index updates, sharding, and serving, so the application only needs to upsert vectors and call findNeighbors. Indexes integrate with the same IAM and VPC controls as the rest of Vertex AI. Example prompt: Call findNeighbors on a deployed IndexEndpoint with a query embedding and use the returned datapoint IDs to fetch source documents for the LLM context window. ### Model Lineage and Governance Regulated teams need to answer how a deployed model was produced - which dataset, which training run, which evaluation. Vertex AI's metadata store exposes Artifacts, Executions, and Contexts, and the lineage subgraph endpoints walk the graph from a deployed model back to the data that trained it. This produces the audit trail required for ML model risk management. Example prompt: Call queryArtifactLineageSubgraph on the deployed model's artifact resource name and walk the returned graph to surface the training pipeline run and source dataset. ### Agent-Built ML Workflow An agent integrating Vertex AI through Jentic can search for the predict operation, load its schema, and call Gemini or a custom endpoint without writing the OAuth and project-routing boilerplate by hand. Jentic isolates the Google Cloud service account credential and exposes only the operation's inputs, so an agent can chain dataset creation, training, and prediction in one workflow. Example prompt: Use the Jentic search query 'generate text with a Vertex AI Gemini model' to discover the operation, then call generateContent on the chosen publisher model with the prompt and parameters. ## Key endpoints | Method | Path | Description | | --- | --- | --- | | POST | /v1/{endpoint}:predict | Run online prediction against a deployed Vertex AI endpoint | | POST | /v1/{endpoint}:streamGenerateContent | Stream generation from a deployed publisher or custom model | | GET | /v1/datasets | List Vertex AI datasets in a project and location | | POST | /v1/{+context}:queryContextLineageSubgraph | Walk the lineage subgraph for a metadata context | | POST | /v1/{+dataset}:searchDataItems | Search dataset items with filters | ## Key resources - **Endpoints** — Deployed-model serving endpoints for online prediction, including Gemini publisher models - **Models** — Versioned Vertex AI Model resources produced by training or imported from elsewhere - **TrainingPipelines / CustomJobs** — Managed training and tuning runs that produce models - **BatchPredictionJobs** — Bulk inference over Cloud Storage or BigQuery inputs - **Datasets** — Managed datasets, dataset items, and annotation specs - **Indexes / IndexEndpoints** — Vector indexes and serving endpoints for similarity search - **Featurestores** — Online and offline features for ML model input - **Metadata Store** — Artifacts, executions, contexts, and lineage subgraphs for model governance ## Why Jentic - **Setup:** Wiring the Vertex AI API by hand means configuring Google OAuth 2.0 from a service account, rotating short-lived access tokens yourself, and pointing calls at the aiplatform.googleapis.com host across a large operation surface. Through Jentic you install once, import the Vertex AI API from the API Directory, store the OAuth credential once, and your agent calls it. - **Permission scoping:** Vertex AI puts the endpoint and dataset resources in the URL path (/v1/{endpoint}:predict, /v1/{+dataset}:searchDataItems), so a rule can pin your agent to one deployed endpoint or dataset: it can predict or read items there and nothing else. You choose the operations it may call, so training, deployment, or dataset writes are not included unless you add them. - **Credential handling:** Your Google service account credential is stored once, encrypted, by your own Jentic One instance and rotated into short-lived access tokens at execution time. The key never enters the agent's prompt, logs, or context. - **Discovery method:** Agents search Jentic by intent such as 'generate text with a Vertex AI Gemini model' or 'run a batch prediction job', and Jentic returns the matching Vertex AI operation with its input schema so the agent calls the right endpoint without browsing the reference docs. ## Related APIs - **Firebase ML API** — Firebase ML serves models to mobile apps with Firebase auth; Vertex AI is the broader Google Cloud ML control plane. - **AI Platform Training and Prediction API** — The legacy AI Platform API predates Vertex AI and is being retired in favour of Vertex AI's unified surface. - **BigQuery API** — BigQuery is the standard source and destination for Vertex AI training data and batch prediction outputs. - **Cloud Storage API** — Cloud Storage holds the model artifacts, training data, and prediction inputs/outputs that Vertex AI reads and writes. ## FAQ ### What authentication does the Vertex AI API use? Vertex AI uses OAuth 2.0 access tokens minted from a Google Cloud service account or user credential, with the cloud-platform scope. Jentic stores the source credential in its Jentic One instance and gives the agent only short-lived access tokens, so service account keys never enter agent context. ### Can I call Gemini models through the Vertex AI API rather than the consumer Gemini API? Yes. Use the publisher model path projects/{project}/locations/{location}/publishers/google/models/{model}:generateContent or :streamGenerateContent to call Gemini billed against your Google Cloud project, with project-level IAM and VPC controls. ### What are the rate limits for the Vertex AI API? Vertex AI enforces per-region, per-model quotas for online prediction (queries per minute) and concurrent training jobs. Quotas are listed under the Vertex AI service in IAM and Admin, Quotas in the Google Cloud Console and can be raised via quota requests. ### How do I run a Gemini prompt on Vertex AI through Jentic? Search Jentic for 'generate text with a Vertex AI Gemini model', load the schema for the publishers/google/models/{model}:generateContent operation, and execute it. Run pip install jentic and use the async search, load, execute pattern with your project and location. ### Can I deploy a custom-trained model to a Vertex AI endpoint via the API? Yes. Upload or register the model in the Vertex AI Model Registry, then call deployModel on a Vertex AI Endpoint with traffic split percentages to route inference traffic between deployed model versions. ### Is Vertex AI free? Vertex AI is paid: you pay per online prediction request, per training-hour for compute, per node-hour for endpoint serving, and per stored vector for index serving. Pricing varies by model, machine type, and region - see Vertex AI pricing in the Google Cloud Console. ### Can I limit what my agent is allowed to do with the Vertex AI API? Yes. Because you run Jentic One yourself, your own rules decide which Vertex AI operations and which Google service account credential the agent may use. Vertex AI carries the endpoint and dataset in the URL path, such as /v1/{endpoint}:predict and /v1/{+dataset}:searchDataItems, so you can pin the agent to a single deployed endpoint or dataset and let it run predictions or read items there and nothing else. You pick the operations it may call, so training, deployment, and dataset writes stay excluded until you add them.