canonical: https://jentic.com/apis/googleapis.com/recommendations-ai-beta

# Google Recommendations AI (Beta)

Recommendations AI (Beta) is Google Cloud's managed personalization service for predicting catalog items a user is likely to engage with next. The API exposes operations to manage catalogs and catalog items, ingest user events, register prediction API keys, and call serving endpoints that return ranked recommendations. Google strongly recommends new customers use the Retail API, which incorporates the GA version of these capabilities; this beta API remains for existing integrations.

## For AI agents

Manage product catalogs, ingest user events, and call serving endpoints to retrieve personalized recommendations from Google's Recommendations AI Beta service.

## Scope

Does not handle full-text search, fraud scoring, or marketing campaign delivery - use for catalog management, user event ingestion, and recommendation serving only.

## Capabilities

- Create and manage catalogs that hold the items eligible for recommendation
- Import catalog items from Cloud Storage or BigQuery in bulk
- Ingest user events (detail-page-view, add-to-cart, purchase) that train recommendation models
- Register prediction API keys so client apps can call serving endpoints with restricted credentials
- Call predict to retrieve a ranked list of recommended items for a user and event context
- Collect user events through the dedicated collect endpoint for low-latency client-side ingestion

## Use cases

### Product Detail Page Recommendations

On a product detail page, the storefront calls predict with the current user, page event context, and the placement (e.g., others-you-may-like) to receive a ranked list of recommended catalog items. Predictions are personalized by the user-event history Recommendations AI has been trained on.

Example prompt: Call predict on placement projects/p/locations/global/catalogs/default_catalog/eventStores/default_event_store/placements/others_you_may_like with the user event payload for the current visitor.

### Catalog and Event Backfill

Before recommendations are useful, the catalog must be imported and historical user events ingested. Recommendations AI provides import operations on catalogItems and userEvents that can pull from Cloud Storage or BigQuery, making it possible to backfill months of history in a single pipeline run.

Example prompt: Call POST /v1beta1/{+parent}/catalogItems:import with a Cloud Storage URI containing the catalog feed, then POST /v1beta1/{+parent}/userEvents:import for the historical event log.

### Restricted Prediction Keys for Client Apps

Front-end apps should call predict with restricted, per-app API keys rather than full Cloud credentials. The API exposes predictionApiKeyRegistrations so each client app gets its own key tied to a single event store and catalog, reducing blast radius if a key leaks.

Example prompt: Register a prediction API key for our mobile app under the default event store, then ship the key to the app via remote config.

### Agent-Driven Recommendation Operations

A merchandising agent uses Jentic to refresh catalog imports, ingest events, and call predict on demand to populate suggestion modules in marketing emails or off-site placements, without needing direct console access to the Recommendations AI project.

Example prompt: Search Jentic for predict product recommendations, load the schema for the placements predict operation, and execute it with the user and event context for a marketing email recipient.

## Key endpoints

| Method | Path | Description |
| --- | --- | --- |
| POST | /v1beta1/{+name}:predict | Retrieve ranked recommendations for a user event |
| POST | /v1beta1/{+parent}/catalogItems:import | Bulk import catalog items from GCS or BigQuery |
| POST | /v1beta1/{+parent}/userEvents:import | Bulk import historical user events |
| POST | /v1beta1/{+parent}/userEvents:collect | Collect a single user event in real time |
| GET | /v1beta1/{+parent}/catalogItems | List catalog items |
| GET | /v1beta1/{+parent}/predictionApiKeyRegistrations | List registered prediction API keys |

## Key resources

- **catalogs** — Top-level container for items eligible for recommendation
- **catalogItems** — Individual products with create, list, get, patch, delete, and import operations
- **userEvents** — User behavior events with collect, import, list, and purge operations
- **predictionApiKeyRegistrations** — Restricted API keys scoped to an event store for client-side predict calls
- **placements** — Named recommendation placements used as the target of predict calls

## Why Jentic

- **Setup:** Wiring Recommendations AI by hand means configuring Google OAuth 2.0 for control-plane calls, refreshing scoped access tokens, and tracking catalog and user-event resource names across the v1beta1 endpoints on recommendationengine.googleapis.com. Through Jentic you install once, import Recommendations AI from the API Directory, store the OAuth credential once, and your agent calls it.
- **Permission scoping:** This API puts the catalog and event store in the URL path (/v1beta1/{+name} and /v1beta1/{+parent}/catalogItems), so a rule can pin your agent to one catalog: it can serve predictions and read catalog items there and nothing else. You choose the operations it may call, so ones like importing a catalog or ingesting user events are not included unless you add them.
- **Credential handling:** Your Google OAuth credential and prediction key are stored once, encrypted, by your own Jentic One instance and injected at execution time. Neither credential ever enters the agent's prompt, logs, or context.
- **Discovery method:** Agents search Jentic by intent such as 'predict recommendations' or 'import a catalog', and Jentic returns the matching Recommendations AI operation with its input schema so the agent calls the right endpoint without reading the v1beta1 discovery doc.

## Related APIs

- **Discovery Engine API** — GA-era Google Cloud service that incorporates personalization and search capabilities
- **BigQuery API** — Source for catalog and user event imports
- **Cloud Storage API** — Holds JSON or CSV feeds used for bulk imports

## FAQ

### What authentication does the Recommendations AI Beta API use?

The API uses Google OAuth 2.0 with the cloud-platform scope for control-plane and import calls, and registered prediction API keys for client-side predict and collect calls. Through Jentic the OAuth credentials live in the vault and short-lived tokens are minted per call.

### Should I use this API for new projects?

Google recommends new customers use the Retail API, which incorporates the GA version of these capabilities; this v1beta1 surface is maintained primarily for existing integrations. The migration path is documented in the Retail API docs.

### What are the rate limits for Recommendations AI?

Predict QPS, import job, and event ingestion quotas are documented per project in the Recommendations AI quotas page. Import jobs are long-running operations and finish asynchronously, while predict calls are synchronous and sized for serving traffic.

### How do I serve recommendations through Jentic?

Search Jentic for predict product recommendations, load the schema for the placements predict operation which maps to POST /v1beta1/{+name}:predict, and execute it with the userEvent payload and the placement resource name.

### Is the Recommendations AI API free?

No, prediction calls and training data are billed according to the Recommendations AI pricing page. Catalog and event imports also incur Cloud Storage or BigQuery read charges depending on the source.

### How do I ingest user events efficiently?

Use POST /v1beta1/{+parent}/userEvents:import for historical or bulk loads from Cloud Storage or BigQuery, and POST /v1beta1/{+parent}/userEvents:collect for low-latency single-event ingestion from browsers and apps.

### Can I limit what my agent is allowed to do with the Recommendations AI API?

Yes. Because Jentic One is self-hosted, you write the rules that decide which operations and credentials your agent may use. This API carries the catalog and event store in the URL path (like /v1beta1/{+name}:predict and /v1beta1/{+parent}/catalogItems), so a rule can pin the agent to a single catalog and let it only serve predictions and read catalog items there. Higher-impact operations such as catalogItems:import or userEvents:import stay off limits unless you explicitly add them.