Skip to content
Agents tracked: 501 Releases (7d): 777 Agent status: no reported issuesSince last limit reset: Claude 10d · Codex 8d For AI agents: Markdown Updated Nov 21, 2026

PageIndex changelog: what's new each month

Every stable PageIndex release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 4 months covered; the current month updates daily.

November 2026 · so far

5 releases: 0.5.1 → 0.7.0

November brought a major unified SDK release for PageIndex, combining local and cloud capabilities under a single interface with vectorless RAG that costs less and runs faster than traditional vector approaches.

Action needed
  • Flash engine is now the default local mode; use mode="standard" to restore the classic LLM pipeline.

Highlights

  • Unified SDK works identically for both local and cloud deployments, with the same client API for indexing, chat, and agent tools.
  • Vectorless RAG costs less than vector-based RAG for both indexing and queries, with faster tree creation from layout statistics.
  • Flash engine is now the local default, with an option to revert to the classic LLM pipeline.
  • Built-in agent integration supports OpenAI, Anthropic, and Claude MCP tool formats, plus direct function access.
  • Flexible deployment: store documents in your cloud or locally, while choosing any configured chat model to answer queries.

New

  • Vectorless RAG system for document indexing and retrieval.
  • Flash engine as the default local processing mode.
  • Unified PageIndexLocalClient API that works for both local and cloud.
  • Agent integration with support for multiple tool formats: plain functions, OpenAI tools, Anthropic tools, and Claude MCP.
  • One-call agent configuration helpers like openai_agent_config().
  • Page content and tree access through the same client in both local and cloud modes.

Improved

  • Tree creation now uses layout statistics for faster performance.
  • Cost reduced across both indexing and query operations compared to vector RAG.
  • Agent tool contracts brought in-process instead of requiring separate MCP infrastructure.
  • Chat model configuration now decoupled from document storage location.

October 2026

11 releases: 0.2.21 → 0.5.0

October brought a unified vectorless RAG SDK that works identically on local and cloud backends, with the new Flash engine speeding up indexing by deriving trees from layout statistics instead of calling LLMs for structure. The platform is now cheaper and faster than vector-based approaches while integrating seamlessly with agent frameworks.

Highlights

  • Flash engine as local default: indexing trees built from layout stats, with LLMs only writing node summaries.
  • Identical local and cloud API: PageIndexLocalClient and cloud client share the same surface for indexing, chat, and agent tools.
  • Agent-ready out of the box: built-in support for OpenAI, Anthropic, Claude MCP, and generic agent tool formats.
  • Flexible model and backend routing: specify your own models and OpenAI-compatible servers per operation.
  • Streaming chat with visible reasoning: watch thinking and tool calls unfold in real time or via typed events.

New

  • Flash engine: derives document trees from layout statistics without LLM involvement for structure generation.
  • Streaming chat with events: observe reasoning and tool calls as they happen with chat(stream=True) or .events API.
  • Agent tool integration: client.agent_tools() provides tools in OpenAI, Anthropic, and Claude MCP formats.
  • Model and backend flexibility: index_model, chat_model, index_backend, chat_backend accept custom LiteLLM-routed providers and keyless OpenAI-compatible servers.
  • Unified local and cloud client: single PageIndexLocalClient API works for both local storage and cloud backend.
  • Cloud document storage with custom models: api_key controls document location; chat_model controls who answers queries.

September 2026

8 releases: 0.2.13 → 0.2.20

September brought PageIndex's vectorless, reasoning-based RAG SDK to both local and cloud with a unified API. The new Flash engine significantly speeds up indexing by deriving document structure from layout statistics rather than LLM calls.

Highlights

  • Flash engine now the local default, cutting indexing time by skipping LLM calls for structure generation.
  • Unified client API across local and cloud deployments with identical methods for indexing, chat, and agent tools.
  • Built-in agent integration supporting OpenAI, Anthropic, Claude MCP, and generic function-based tools.
  • Streaming chat with real-time visibility into model thinking and tool calls woven into responses.
  • Flexible model and backend configuration supporting LiteLLM providers, OpenAI-compatible servers, and custom endpoints.

New

  • PageIndex Flash engine for layout-stats-based document structure extraction.
  • Unified PageIndexLocalClient and PageIndexClient with identical surface for both modes.
  • Streaming chat mode via chat(stream=True) showing thinking and tool calls as they happen.
  • Typed event streaming via .events for fine-grained run visibility.
  • Agent tool export in multiple formats: plain functions, OpenAI tools, Anthropic tools, Claude MCP.
  • One-call agent configuration via openaiagentconfig() and similar helpers.
  • Embedded bookmark consumption when trustworthy.
  • Tree optimization on by default with configurable merge strategy.

Improved

  • Indexing performance dramatically improved through Flash engine layout-based approach.
  • Tree expansion now handled separately from structure generation for better control.
  • Model and backend configuration passed verbatim to support diverse provider ecosystems.
  • Chat API unified across all backends with consistent parameter handling.
  • Document storage location decoupled from chat model selection in cloud mode.
  • Local storage path now configurable for flexible deployment options.

August 2026

3 releases: 0.2.10 → 0.2.12

August brought PageIndex's unified SDK with local and cloud modes sharing one API, powered by a new Flash engine that dramatically speeds up indexing. The month focused on reasoning-based retrieval without vectors and seamless agent integration across deployment types.

Action needed
  • Python 3.10 now required as minimum version.
  • Anthropic and Claude SDKs moved to optional extras: install with [anthropic] or [claude] flags.

Highlights

  • New PageIndex Flash engine indexes documents 3-5x faster by deriving structure from layout statistics instead of calling LLMs for tree generation.
  • Single SDK works identically in local and cloud modes: PageIndexLocalClient and cloud client share the same methods for indexing, retrieval, chat, and agent tools.
  • Built-in agent integration with OpenAI, Anthropic, and Claude; client generates compatible tool definitions automatically.
  • Chat interface works across any backend with streaming, usage tracking, and document targeting.
  • Flexible model and provider configuration: bring your own OpenAI-compatible server, use LiteLLM routing, or go keyless.

New

  • PageIndex Flash engine: layout-based indexing that skips LLM calls for tree structure generation.
  • Unified client API for local storage and cloud deployments.
  • Agent tools generation for OpenAI, Anthropic, and Claude via client methods.
  • Chat completions protocol with streaming and document ID targeting.
  • Tree optimization on by default with configurable merge strategy.
  • Embedded bookmark consumption when marked trustworthy.
  • Per-call backend override for index and chat operations.
  • Keyless OpenAI-compatible server support.

Improved

  • Tree expansion now handles layout-derived structures more efficiently.
  • Indexing performance significantly faster with Flash engine as local default.
  • Chat responses include honest usage accounting across all backends.
  • Model configuration passed verbatim to index and chat pipelines for vendor flexibility.
  • Bookmark handling improved to validate trustworthiness before consumption.

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. PageIndex: pricing, features and alternatives · All changelogs