Scribd, Inc. classifies more than 400 million documents with Gemini batch inference on Gemini Enterprise

Scribd, which hosts one of the world’s largest collections of user-generated documents, needed an efficient way to classify its 400 million documents for trust and safety. The company turned to Google Cloud’s Gemini Enterprise batch prediction, leveraging native PDF understanding to process the entire corpus in months rather than years. By accepting PDFs directly without OCR or rendering pipelines, Scribd avoided costly preprocessing steps while maintaining high accuracy across diverse content types.

The team evaluated multiple moderation tools but found none met their quality and scalability needs. Gemini’s multimodal model allowed a single pipeline to handle both text and images, collapsing what would have required years of specialized engineering into a unified approach. Sachin Sebastian, Senior Engineering Manager at Scribd, noted that Gemini’s native PDF processing eliminated the need for separate solutions for different content categories.

Gemini 2.5 Flash Lite served as the primary classification model, with Gemini 2.5 Pro used for validation in a second consistency pass. The fixed token count per PDF page ensured predictable costs, and batch pricing at a 50% discount to interactive rates made large-scale classification economically viable. The team also utilized prefix caching to further improve efficiency without compromising accuracy.

Google Cloud’s partnership was critical to scaling the backfill, with direct engineering support ensuring capacity and throughput matched project demands. The batch jobs completed faster than projected, shifting bottlenecks to Scribd’s upstream pipeline. The success of the project has since expanded into a continuous classification pipeline for new uploads, with the same approach now applied to additional content-understanding workloads across Scribd’s platforms.