Hot AI News

Google Cloud Storage can now tag your images and PDFs for AI agents, in preview

Google Cloud's object enrichments preview runs pre-trained models over images and PDFs in Cloud Storage and stores the results as metadata that agents and BigQuery can query.

Google Cloud Storage can now tag your images and PDFs for AI agents, in preview
Image: Google

Google Cloud has introduced object enrichments for Cloud Storage, a preview feature that runs pre-trained models over stored images and PDFs and saves the results as metadata on each object, so agents can find and use files.

Note on sourcing: the announcement post itself did not load when this article was prepared, so the technical details below come from Google’s product documentation for the feature, plus a recent interview with a Google Cloud Storage executive. Details that appear only in the blog post may be missing here.

Diagram from the Google Cloud blog post on GCS object enrichments, placed after the introduction. I could not view the image, so the exact content is unconfirmed.
Image: Google Cloud

Five model types, two file families, a 20 MiB cap

According to Google’s documentation, the preview covers five kinds of enrichment: object detection with bounding boxes, image labeling, face detection, explicit content detection, and PDF classification into business and industry categories with scores. The docs do not list text extraction from documents as an enrichment type, even though the feature is pitched as turning unstructured data into context.

Supported inputs are common image formats (BMP, GIF, ICO, JPEG, PNG, TIFF, WEBP) and PDFs, up to 20 MiB per object. Google decides what to process from the object’s Content-Type metadata; an object with a missing or invalid value is skipped even if the file is valid. Processing is asynchronous.

Second figure from the same post, in the 'Reduce costs with Google models' section. I could not view the image, so the exact content is unconfirmed.
Image: Google Cloud

Results are attached to each object as object contexts under contexts.custom. Teams can inspect single objects in the console or through the APIs, or export metadata through Storage Insights datasets to BigQuery and query the object_contexts_catalog_view view with standard SQL, for example to filter by category or find high-confidence labels.

Advertisements

How to switch it on, and what it excludes

An administrator creates a feature configuration that names the buckets or projects to monitor and the models to run. Buckets must be named explicitly: wildcards and regular expressions are not supported, and new buckets are not picked up automatically. Creating or updating a configuration can take up to 24 hours to take effect.

Processing happens in one of three locations: us, eu or asia. Google tells customers to pick one location per bucket and model, because configuring more than one means being billed for each run. Objects encrypted with customer-supplied keys are not supported, nor are Rapid buckets or buckets with hierarchical namespace. As a Pre-GA feature it is offered as is and may have limited support under Google’s Pre-GA terms.

A bill with five meters, not one

The analysis angle here is cost. Google’s documentation says charges apply for model inference, metadata storage, data access operations and networking. Storage Insights and BigQuery queries add their own costs, and reading from Autoclass buckets can trigger storage class transitions. The pages we read give no price list, so the total cost for a given bucket cannot be estimated from them.

That matters because the pitch is cheaper context extraction than building a pipeline by hand. A custom pipeline has costs of its own, but the managed route is billed per run and per object, so a large archive of images with several models enabled will multiply quickly. The duplicate-location warning shows how an easy configuration slip becomes a repeated charge.

There is also a tension with how Google’s storage team describes the product. In a recent interview reported by aVenture Research, Ronen Schwartz, Google Cloud Storage’s VP of products, said the Smart Storage initiative, anchored by object enrichment, attaches structured metadata to arriving files in seconds and at scale. The documentation instead describes asynchronous processing and a configuration step of up to 24 hours. Both can be true, since one is per-object latency and the other is setup time, but buyers planning an ingest workflow should test the real delay. The same interview put Rapid object storage at more than 15 TB/s for a single bucket, a vendor figure the report did not verify. Rapid buckets are exactly the ones object enrichments cannot cover today, so the fastest storage tier and the new enrichment feature do not yet overlap.

Advertisements

What is established and what is not

Established, per the docs: the feature exists in preview, the five model types, the formats, the size cap, the three processing regions and the billing categories. Not established: pricing, a general availability date, accuracy of the models on customer data, and any benchmark showing cost savings against a self-built pipeline. Claims that the output is agent-ready come from Google and have not been independently tested.

For teams already holding large image and PDF stores in Cloud Storage, the practical change is that labels and categories can now be generated and queried in place instead of through a separate service. Teams with mostly text, audio or video, or with customer-supplied encryption keys, are outside the preview’s scope for now.

Sign Up for the Daily AI Pulse

One email a day. All the stories that matter.

Alex CarterAI News & Big Tech Correspondent

Alex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.

Leave a Reply

Your email address will not be published. Required fields are marked *

You may also like