Google Cloud adds object enrichments to extract AI metadata from Cloud Storage files
Google Cloud has launched object enrichments in preview, a managed Cloud Storage feature that runs AI models over stored images and PDFs and writes the results back as metadata on each object. The company says enriching 10 million images or PDFs costs $700.
TL;DR
- Google Cloud Storage now has object enrichments in preview, a managed service that runs AI models over stored files and saves the output as metadata on each object.
- The service targets the cost, plumbing and sync problems teams hit when they pre-process unstructured data for AI agents.
- Google says enriching 10 million images or PDFs costs $700 at published pricing, and Gemini-based enrichment is available on request.
What the feature does
Most enterprise data is unstructured: images, PDFs, video and audio stored as objects. An AI agent cannot reason over raw bytes, so teams usually extract descriptions, labels or categories ahead of time. Google Cloud's new object enrichments feature automates that step. It uses AI models to extract contextual metadata from Cloud Storage objects and stores it in place as what Google calls object context, a set of structured key-value pairs attached to the object itself.
The launch includes five first-party Cloud Storage models for images and PDFs. The image models cover object detection, face detection, image labeling and explicit content detection. A PDF classification model sorts documents into business and industry categories such as Finance, News and Health. Google says the models are the same ones used by internal services such as Google Photos.
Cost and setup
Google's main pitch is price. The company says enriching 10 million images or PDFs costs just $700 according to its published pricing, and it claims savings of 75% compared with similar services from other cloud providers. These are vendor figures and have not been independently verified.
Setup is meant to be simple. According to the Google Cloud Blog, a job needs three steps: define the buckets and processing location, choose one or more models, and grant the service the IAM permissions it needs. The service then handles inference, parsing, retries and throttling. Google presents this as a way to avoid stitching together event triggers, queues and workers on Cloud Run or Google Kubernetes Engine (GKE), which the company describes as a process that can take months.
Gemini is a separate, later option. Google says customers can use its Gemini models for tasks such as summarization, entity extraction, transcription and translation, but that this requires contacting the company for access.
Metadata stays with the object
The design choice that may matter most for developers is where the metadata lives. Typical pipelines write extracted data to a separate database, which then has to be kept in sync when files are copied, moved or deleted. Object context is stored alongside the object. According to Google, it moves with the object during copy, move and replication operations and is removed automatically when the object is deleted.
Access control also follows the object. Administrators can use IAM to decide who can read, add, update or remove object context. Applications can filter object listings by context key-value pairs, and Storage Insights datasets can send the context to BigQuery for SQL queries and joins with structured tables. For teams already using Cloud Storage features like lifecycle rules, the main gain is fewer moving parts rather than new storage behavior.
Early use cases
Google names several scenarios: creative asset search, where descriptive labels can feed a vector database for retrieval-augmented generation (RAG); content moderation of user uploads; curating training data for autonomous driving, smart home and robotics teams; and combining image and document insights with business data for fraud detection or personalization.
Google includes customer quotes. L'Oreal says the feature lets it label its creative asset repository 100 times faster at a fraction of the cost of its existing pipelines. Hyperscience reports a 50%+ lift in document classification accuracy and 2–3x lower classification inference costs. Vivint says it cut AI development time from six months to one. These claims come from the vendor's own blog and customers were not independently tested.
Why it matters
For backend and data teams, object enrichments offers a managed alternative to building custom pipelines for metadata extraction. The main trade-offs to check are the preview status, the per-object cost at your own volume, and whether the five built-in models cover your data. Teams that need the Gemini features must request access first.
The design also changes where governance work sits. If context travels with the object and is deleted with it, teams can drop some sync code, but they will need to review IAM rules for metadata, since read and write permissions on context are now a separate control. For agent builders, the feature offers a path to give agents searchable context over a bucket without streaming every raw file into a model on each request.
Sources
This article was written by Blozhik from the reporting below. Read the originals for full detail.
GoIntroducing object enrichments: Transform unstructured data into agent-ready contextGoogle Cloud Blog