LlamaIndex 发布 ParseBench 及 Opus 4.7 评测周报

Hi there, Llama Enthusiasts! 🦙

Welcome to this week's edition of the LlamaIndex newsletter! We're excited to bring you major updates including the launch of ParseBench (the first document OCR benchmark for AI agents), LiteParse officially joining the LlamaIndex ecosystem, comprehensive benchmarking of Anthropic's Opus 4.7, and an upcoming NYC FinTech Week AI event. Check out these developments along with our deep dives into chart parsing, table evaluation metrics, and content faithfulness testing.

If you haven't explored LlamaParse yet, make sure to sign up and get in touch with us to discuss your specific enterprise use case.

🚀 ParseBench Launch Special

• Introducing ParseBench: The first document OCR benchmark specifically designed for AI agents, featuring comprehensive evaluation metrics for charts, tables, content faithfulness, and more. Full details and GitHub repo
• We also posted deep dive videos in three of the five new accuracy metrics released with ParseBench:
  • Advanced Table Parsing Metrics: Deep dive into TableRecordMatch (GTRM), our new metric for evaluating complex tables as records keyed by column headers—the way your pipelines actually consume them. ParseBench blog
  • Content Faithfulness Testing: ParseBench introduces comprehensive evaluation of three critical failure modes—omissions, hallucinations, and reading order violations—with 167K+ rule-based tests to ensure parsing reliability for agent workflows. ParseBench details
  • Chart Data Point Extraction: New ChartDataPointMatch metric goes beyond OCR'ing captions to extract actual numerical data from charts—bridging the gap between text recognition and true chart comprehension. ParseBench blog

🤩 The Highlights

• LiteParse Gets New Website Page: After hitting 4.3K+ GitHub stars, LiteParse now has its official home with ~500 pages parsed in 2 seconds, 50+ formats supported, and zero cloud dependency. Join the upcoming live workshop to build a fintech due diligence agent. LiteParse page | Workshop signup
• Anthropic Opus 4.7 ParseBench Results: We benchmarked the new Opus 4.7 model showing massive chart parsing improvements (+42.3%) but mixed results across other categories. LlamaParse Agentic still leads with 84.9% overall performance at competitive pricing. ParseBench GitHub

• NYC FinTech Week AI Track: Join us next week for the AI Builders Rooftop Happy Hour, co-hosted with LinkupAPI for developers shipping fintech agents, document intelligence, and agentic workflows. RSVP