Why it matters: enterprise OCR failures break agents long before they show up on academic PDF benchmarks. LlamaIndex says ParseBench evaluates about 2,000 human-verified pages with over 167,000 rules across 14 methods on Kaggle.
Why it matters: document agents fail when PDF parsing destroys table and column structure. LiteParse uses a monospace grid projection approach instead of heavy layout models, and the code is open source.
Why it matters: document agents fail when parsers drop tables, chart values, or visual grounding. ParseBench uses about 2,000 enterprise document pages, 167K+ rule-based tests, and 14 evaluated methods.
A detailed engineering write-up resonated on Hacker News because it treated production RAG as a data and operations problem, not a prompt demo.
A LocalLLaMA thread and linked GitHub issues argue that LlamaIndex's OpenAI-by-default behavior can surprise local-first RAG builders when nested components are created without explicit model injection. Maintainers say the behavior is longstanding and documented, but the discussion is pushing for a stricter fail-fast mode for sovereign deployments.