XNews

PageIndex Flash: fully local PDF tree engine uses layout structure, not visual models

PageIndex Flash: fully local PDF tree engine uses layout structure, not visual models

PageIndex Flash is an open-source, fully local PDF indexing engine that builds a hierarchical table of contents by reading the PDF's own layout information instead of relying on visual model inference. It addresses the similarity≠relevance problem in long documents by letting an LLM navigate a tree index to locate correct pages. The two-stage process builds a tree with headings, page ranges, and summaries, then uses an LLM to search the tree. Indexing costs approximately $0.001 per page with gpt-5.6-luna (a 1000-page textbook costs ~$1), and index time ranges from 13 seconds to 4.5 minutes for documents between 9 and 1098 pages. On 62 lookup questions from MMLongBench-Doc-V2, tree search outperforms full-text feeding: for a 52-page doc, full-text costs 2.1× more; for 420 pages, 16.6×; beyond 800 pages, full-text exceeds the context window. The engine does not work with scanned or image-heavy PDFs; those require PageIndex Cloud. Agent integration provides as_openai_tools() and agent_instructions() for frameworks like OpenAI Agents SDK, Anthropic, LangChain, and PydanticAI.

Source posts