PageIndex's new engine builds RAG trees without an LLM

PageIndex's v0.2.11 release makes Flash the local default for its vectorless RAG approach, replacing the tree-building LLM call with layout statistics. Anyone indexing large document collections paid for an LLM call on every file's tree structure. As of v0.2.11, only the node summaries need one.

Tree expansion used to be a round-trip per node. The Flash engine proposes a wave of nodes in one pass instead. Embedded PDF bookmarks get pulled in directly when they're trustworthy enough to skip inference.

The optimize='merge' option goes further and gives a fully deterministic, LLM-free pass, producing the same tree from the same input every run.

The release adds agent bindings too, exposed as client.agent_tools(), as_openai_tools(), as_anthropic_tools(), and as_claude_mcp(). A prompt written for one of these ports to the others unchanged. That means testing locally before moving to a hosted model, without rewriting the prompt.

What the release notes don't say is how the layout-statistics approach handles documents with inconsistent formatting, scanned PDFs, or non-standard heading structures. Those are the cases an LLM's judgment call used to handle.