How Scribd leverages Meta’s AI models to make nuanced book recommendations

The evolution of AI-powered book discovery is transforming how readers find their next great read, with Scribd, Inc.’s Everand service leading the way through innovative use of Meta’s Llama language models.

The transformation of book discovery: Everand’s new Ask AI feature represents a significant upgrade from traditional keyword-based searches, serving a massive library of over 195 million pieces of content to 200 million monthly visitors.

• The previous system relied on pre-generated topics and basic keyword searches, limiting users’ ability to discover specific content
• The new AI-powered system can handle complex, nuanced queries like exploring ancient martial arts elements in modern romance stories
• The service integrates three versions of the Llama model: 8B, 70B, and 405B, each serving specific functions in the recommendation system

Technical implementation: Scribd’s engineering team leveraged advanced AI training techniques and open-source flexibility to create a highly customized recommendation system.

• The team used Llama 3.1 405B to generate synthetic training data that simulates diverse user behaviors
• Parameter-efficient fine-tuning (PEFT) with QLoRA/LoRA and supervised fine-tuning helped optimize the 8B model for accurate, low-latency responses
• The 70B model works in the background to generate metadata for new books, continuously improving discovery accuracy

Infrastructure and deployment: The system’s architecture prioritizes performance and cost optimization while maintaining seamless integration with existing systems.

• Implementation utilizes AWS and Databricks batch inference for large-scale data analysis
• The system outputs responses in JSON format to enhance metadata extraction and real-time performance
• The open-source nature of Llama allowed for deep customization without requiring major infrastructure changes

Key benefits and outcomes: The new AI-powered system delivers significant improvements in content discovery and user experience.

• Users can now explore content through natural language queries rather than rigid keyword searches
• The system demonstrates sophisticated understanding of user intent, even with unusual or complex requests
• Real-time recommendations maintain low latency while delivering highly accurate results

Future trajectory: The implementation of Llama models in Everand’s Ask AI feature appears to be just the beginning of a broader AI integration strategy that could reshape digital reading services.

• Scribd plans to expand Llama integration across more aspects of the user experience
• The company will implement Llama Guard 3 for enhanced content moderation and trust features
• These improvements are expected to boost customer retention and lifetime value

Reading between the lines: While the initial results seem promising, the true test will be whether this AI-powered discovery system can consistently surface relevant content that keeps readers engaged and subscribing to the service over the long term.