LlamaIndex 周报：发布 ParseBench CVPR 成果及 Anthropic Fable 5 测试数据

Hi there, Llama Enthusiasts!

Welcome to this week's edition of the LlamaIndex newsletter! We're excited to bring you major updates including our groundbreaking ParseBench presentation at CVPR 2026, the launch of Parse-Flow for visual document intelligence workflows, Anthropic Fable 5 benchmarking results, and new Granular Bounding Boxes in LlamaParse. Plus, don't miss The Agent Open - AI's first pickleball tournament combining code and competition!

SF Events

• DAIS AI Engineer Happy Hour: Get away from the Moscone crowd and enjoy some good food amongst the brightest devs in SF. Just vibes, no demos. Register now
• The Agent Open Tournament: Join us for AI's first pickleball tournament featuring custom courts, stadium seating, exhibition matches by AI leaders, and every infra startup you love in one place. Register now

The Highlights

• Anthropic Fable 5 ParseBench Results: Day 0 testing shows impressive document understanding with 90.02% content faithfulness and 72.62% semantic formatting - leading competitors by 12+ points in key metrics. Check out the latest leaderboard results at www.parsebench.ai 
• ParseBench at CVPR 2026: Presenting the first doc-parsing benchmark built for AI agents with 2,000+ human-verified pages and 167K+ test rules across 5 dimensions. Dataset, Code, Paper

LlamaParse

• Granular Bounding Boxes: New word, line, and cell-level coordinates for every extracted value, providing complete audit trails from extracted data back to exact source locations in documents. Perfect for compliance and verification workflows. Full announcement
• Parse-Flow Visual Workflow Designer: Open-source project tackling enterprise document processing with four primitives (Parse, Classify, Split, Extract) in a drag-and-drop interface powered by LlamaAgents workflows. Full architecture guide, Source code