Scribd leverages Gemini Enterprise batch inference for 400M+ document classification

Scribd needed to perform trust and safety classification across its 400M+ document corpus, comprising over 12 billion pages, but found traditional methods and existing moderation tools inadequate. They sought a single, cost-effective solution capable of multimodal understanding across diverse content types, languages, and contexts.

Gemini's ability to accept PDFs natively and process each page as both text and image was a key differentiator, eliminating the need for OCR or rendering pipelines for over 99% of Scribd's content. This multimodal approach allowed Gemini 2.5 Flash Lite and Pro models to identify visual policy signals that text-only moderation endpoints often missed.

Gemini Enterprise's batch prediction model, priced at a 50% discount compared to interactive rates, made large-scale LLM classification economically feasible. The execution model was simplified: documents were staged in Cloud Storage, submitted to Gemini, and results flowed back into Scribd's data platform, avoiding serving infrastructure overhead.

Google Cloud partnered with Scribd to plan the backfill, advise on region strategy, and scale batch throughput, resulting in faster-than-projected job completion. This successful project now underpins continuous evaluation of newly uploaded content and is being extended to other content-understanding workloads, significantly streamlining future roadmap items.