AI NewsInfrastructureAnnouncement

Google Cloud Storage now stores AI-extracted labels beside each file

Google Cloud Storage now extracts labels, tags and classification directly from images and PDFs and saves them beside each file. An agent can read that context without running a model itself.

AI News

Editorial2 min read

LinkedInX
Google Cloud Storage object enrichments diagram

Why it mattersA team wiring an agent to a petabyte of PDFs usually builds its own pipeline. Context stored in the bucket removes that pipeline and the queue behind it.

Google Cloud released object enrichments for Cloud Storage on 19 November, in preview. The service runs a first-party AI model over an image or PDF in a bucket. It writes the output as object context, a key-value field stored beside the file itself. An agent or a query can read those labels directly. No model call on the raw bytes.

What it does

Five first-party models cover images and PDFs at launch. For images: object detection, face detection, image labeling and explicit content detection. For PDFs: a classification model for standard business categories. Video and audio are added through upcoming Gemini enrichments, available by invitation at storage-ai@google.com.

A user sets up a job in three steps, as Google describes it. Define the scope by naming a bucket and a processing location. Choose which first-party model to run. Grant the service the IAM permissions it needs. The service then runs the model over the files. It parses the response and writes the result to object context. Retries and throttling are handled for you.

What Google quotes as the price

Google says it costs 700 dollars to enrich 10 million images or PDFs. The post calls that a 75 percent saving against similar services from other cloud providers. The comparison is Google's own, and the post does not name the services it measured against.

Three customer numbers are Google's own, each attributed on the page. L'Oreal reports 100 times faster image labeling, at a fraction of the cost of the pipeline it ran before. Hyperscience reports a 50 percent or higher lift in document classification accuracy, at two to three times lower cost. Vivint says the service cut an AI development schedule from six months to one.

Why context in the bucket matters

Object context has two properties the post calls out and that matter for an agent. Context travels with the file. A copy, a move or a replication carries the labels across. Deleting the file deletes the context. A deleted PDF leaves no stale metadata in some other table.

IAM governs read, write, update and delete permissions at the context level. A team can give an agent read-only access to the enrichments without handing it the file. The context is queryable through fast filters and through Storage Insights datasets in BigQuery. An analyst can run SQL over the labels with no extract step in between.

The release opens a small shift in how a team designs an agent against a large bucket. The pipeline that labelled every file before the agent could read it moves into the storage layer. The agent reads context by key, not by running a model over the raw file.

Source

Transform unstructured data into agent-ready context, Google Cloud.

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project