Scribd moved from a collection of bespoke moderation tools to a single Gemini Enterprise batch inference pipeline that ingests PDFs directly, enabling classification of more than 400 million user‑uploaded documents (over 12 billion pages) in a few months. This change matters to engineers because it proves that large‑scale, multimodal LLM classification can be done cost‑effectively, without building OCR or rendering layers, and with a predictable, linear cost model.
PDF batch classification with Gemini Enterprise
The core shift was to treat the corpus exactly as it exists: as PDF files. Gemini’s native PDF support reads each page as both text and image, allowing a single multimodal model to evaluate dense documents and image‑heavy presentations in one pass. Scribd selected the Gemini 2.5 Flash Lite model for the primary classification pass and used Gemini 2.5 Pro for a secondary consistency check. Because the model processes pages at a fixed token count, costs scale linearly, and the discount on batch pricing made the operation financially viable at corpus scale.
Architecture and data flow
Documents are first staged in Cloud Storage. A batch job submits the stored PDFs to Gemini Enterprise’s batch prediction endpoint. The service returns classification results, which are then written back to Cloud Storage or a downstream data store for further processing. This pattern eliminates the need for an OCR pipeline, page rendering service, or screenshot generation, simplifying the data path and reducing surface area for operational failures.
Operational considerations
Running a corpus‑wide backfill required orchestrating high‑throughput batch jobs. Google Cloud automatically scaled the batch service to meet the timeline, but practitioners must still monitor job queuing, storage I/O, and result ingestion. The 50 % discount on batch pricing relative to interactive pricing is a key economic factor; teams should evaluate eligibility and compare batch versus real‑time costs for their own workloads. Model selection (Flash Lite for bulk work, Pro for validation) introduces a two‑stage quality gate, which implies additional storage for intermediate results and a verification step in the pipeline.
- Ensure Cloud Storage buckets have appropriate IAM policies to restrict access to raw PDFs and classification outputs.
- Instrument batch job metrics (throughput, latency, error rates) to detect scaling bottlenecks early.
- Plan for storage lifecycle rules to archive or delete intermediate artifacts after validation.
- Validate a sample of batch results against human review to confirm model quality before full backfill.
Related CloudNinjas coverage: Google Cloud.
What This Means For Practitioners
Engineers building large‑scale content moderation or data‑labeling pipelines can consider a native PDF multimodal model combined with batch prediction as a way to cut preprocessing complexity and control costs. The Scribd case shows that a simple storage‑to‑batch‑prediction workflow, paired with a discounted batch pricing tier, can handle billions of pages without custom OCR infrastructure. Teams should assess their document formats, evaluate batch pricing models, and design monitoring around the batch job lifecycle to replicate this pattern safely and efficiently.
