Skip to content
CAI
Software that uses CAICheck a score

VectifyAI/PageIndex

59.1

Adequate · 15 November 2026

29.8k

lines of production code

Python

primary language

5

measurements over time

CAI band scale
CAI trend line
CAI lens gauges

What this system is

Features

Introduce PageIndex Flash for LLM-free PDF tree extraction

Adds the \pageindex.flash\ package, providing a new indexing mode that builds a PageIndex tree structure from PDF layout statistics without requiring an LLM. The \page\_index\_flash\ API accepts a PDF path or stream and returns a hierarchical structure with optional LLM-generated summaries and optimization (merge/expand). It includes logic to detect headers, footers, watermarks, and boilerplate, and supports using embedded PDF bookmarks to frame the detected layout structure. The local client and CLI now support \--mode flash\, and the system enforces rejection policies for unreadable or excessively flat documents. Also in this area: 1 commit (0 fixes) modifying (repo-wide), pageindex; 1 commit (1 fix) fixing pageindex; Token usage reporting and improved error handling for local indexing and chat; Claude Code integration: run indexing and chat on a Claude subscription; Fix: Running Table of Contents links no longer hide headings.

pageindex/flash · low confidence · unverified

New cookbook notebooks for citations, Flash demo, and multimodal use cases

Added three new Jupyter notebooks to the cookbook: \pageindex-citation.ipynb\ demonstrates how to use PageIndex Cloud's block-level citations to highlight specific source regions in original documents; \pageindex-flash-demo.ipynb\ shows how to use the PageIndex local SDK with the default tree-indexing model (PageIndex Flash) to build vectorless RAG; and \pageindex-multimodal.ipynb\ provides a guide for multimodal workflows. Also in this area: 11 commits (0 fixes) modifying cookbook; Improved error messaging for Claude authentication and submission failures; Updated Claude subscription example to stream answers from the Harness² paper.

cookbook · medium confidence · unverified

New framework adapters for agent tool integration

Added new integration modules for Anthropic, Claude Agent SDK, and OpenAI Agents SDK, enabling PageIndex tools to be used directly within these agent frameworks. The Anthropic adapter provides sync and async tool wrappers that proxy calls through MCP, handling errors as tool failures. The Claude Agent SDK adapter creates an in-process MCP server for local use and a cloud relay for CLI sessions, ensuring API keys remain secure. The OpenAI Agents SDK adapter exposes PageIndex tools as MCP servers or hosted tools, allowing seamless integration with OpenAI's agent runtime.

pageindex/integrations · high confidence

PageIndex SDK v0.2.x: Local mode, multi-provider chat, and agent tool integrations

The PageIndex SDK now supports a local indexing and chat mode alongside the existing cloud surface, allowing users to index and query documents entirely on-device using configurable models (via LiteLLM) and a local storage backend. The SDK introduces a unified \PageIndexClient\ that routes to either cloud or local APIs, and adds a \ChatStream\ type for streaming chat responses with support for both text and event views. New agent tool integrations are available for OpenAI Agents SDK, Anthropic SDK, and Claude Agent SDK, enabling document-aware agents to use PageIndex tools natively. The SDK also includes image highlighting helpers for citation bounding boxes, lazy-loading of heavy dependencies, and improved error handling with explicit API key and model validation.

pageindex · high confidence

Behavioural changes

5 commits (0 fixes) modifying cookbook/pageindex-vs-vector-10k

A change to existing behaviour in cookbook/pageindex-vs-vector-10k — 5 commits, 6 files: “docs(cookbook): show the query cost and accuracy chart at the top”; “docs(cookbook): accuracy and query cost, PageIndex vs vector RAG with six embedding models”; “docs(cookbook): query cost and accuracy, PageIndex vs Gemini Embedding 2 on five 10-Ks”; and 2 more.

cookbook/pageindex-vs-vector-10k · medium confidence · unverified

Restructured CLI entry point and unified PDF/Markdown processing

The repository now uses \run\_pageindex.py\ as the primary CLI entry point, replacing the previous \page\_index.py\. This new script unifies processing for both PDF and Markdown documents, introducing a \--mode\ flag (defaulting to 'flash') that controls whether the system uses the optimized Flash indexing path or the standard path. It exposes granular configuration for Flash mode, including \--optimize\ (with 'full', 'merge', or 'off' options), \--embedded-toc\, and \--summary\ toggles, alongside model selection via \--index-model\ and \--summary-model\. The legacy \page\_index.py\ and \utils.py\ modules have been removed in favor of this new structure.

(repo-wide) · high confidence

Test coverage

Added test fixtures and contract data for agent tools and flash layout; Initial test suite for PageIndex SDK and Flash indexing.

Dependencies

Update PageIndex SDK dependencies and add pyproject.toml

The PageIndex Python SDK (v0.2.10) updates its core dependencies to support multi-provider LLM integration and improved retrieval. Key changes include adding \openai-agents\ (\>=0.18.1) as a base dependency, upgrading \litellm\ to 1.97.0, and adding \mcp\ (\>=1.19.0,\<3) for tool-path support. The manifest also introduces optional extras for \claude\ (via \claude-agent-sdk\) and \anthropic\ (\>=0.122.0), while updating \PyPDF2\, \pypdfium2\, \python-dotenv\, and \pyyaml\ to their latest compatible versions. A new \pyproject.toml\ file standardizes the build configuration using Poetry.

(dependencies) · high confidence

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

How this codebase got here

Score

  • CAI 58 → 59 (+0.6)
  • Rubric changed (rubric-2026.11.1 → rubric-2026.11.4) — scores are not directly comparable.

Lenses

  • Code Health 60 → 58 (-1.3)
  • Architecture 88 → 99 (+10.8)
  • Maturity 44 → 45 (+1.1)
  • Readiness 75 → 85 (+10.2)
  • Security 82 → 73 (-8.7)
  • Performance 100 → 100 (+0.0)

Resolved (10)

  • Duplicated block (13 lines × 2) (pageindex/cloud_api.py)
  • Duplicated block (18–20 lines × 2) (pageindex/cloud_api.py)
  • Duplicated block (18–26 lines × 3) (pageindex/cloud_api.py)
  • Duplicated block (8 lines × 5) (pageindex/cloud_api.py)
  • Members sharing a duplicated core (5 members, 50+ identical tokens) (pageindex/cloud_api.py)
  • Off-boarding risk: anonymized user #1
  • Projects may be oversized for their cohesion
  • Repeated repair: pageindex/mcp_bridge.py (pageindex/mcp_bridge.py)
  • utils.llm_acompletion (cognitive 18) (pageindex/utils.py)
  • utils.llm_completion (cognitive 27) (pageindex/utils.py)

New (28)

  • Duplicated block (14 lines × 2) (pageindex/cloud_api.py)
  • Duplicated block (5 lines × 3) (pageindex/client.py)
  • Duplicated block (6–7 lines × 2) (pageindex/agent_tools.py)
  • Duplicated block (7 lines × 2) (pageindex/client.py)
  • Edited copies of one member (3 members, 25+ corresponding lines) (pageindex/cloud_api.py)
  • Edited copies of one member (3 members, 8+ corresponding lines) (pageindex/flash/model/rects.py)
  • Edited copies of one member (5 members, 19+ corresponding lines) (pageindex/cloud_api.py)
  • Edited copy of a member (13 corresponding lines) (pageindex/utils.py)
  • Edited copy of a member (15 corresponding lines) (pageindex/flash/model/numbering.py)
  • Edited copy of a member (15 corresponding lines) (pageindex/utils.py)
  • Edited copy of a member (21 corresponding lines) (pageindex/cloud_api.py)
  • Edited copy of a member (22 corresponding lines) (pageindex/cloud_api.py)
  • Edited copy of a member (22 corresponding lines) (pageindex/flash/parser_pdfium_charlevel/remerge.py)
  • Edited copy of a member (24 corresponding lines) (pageindex/page_index_classic.py)
  • Edited copy of a member (29 corresponding lines) (pageindex/utils.py)
  • Edited copy of a member (9 corresponding lines) (pageindex/utils.py)
  • Members sharing a duplicated core (4 members, 50+ identical tokens) (pageindex/client.py)
  • Off-boarding risk: anonymized user #1
  • Repeated repair: pageindex/integrations/claude_agent_sdk.py (pageindex/integrations/claude_agent_sdk.py)
  • api.page_index_flash (cognitive 16) (pageindex/flash/api.py)
  • …and 8 more

Changes since last survey

  • 17 commits — 9 feature/other, 8 fixes

By area

  • pageindex/client.py — 8 commits
  • (repo) — 2 commits
  • pageindex/cloud_api.py — 2 commits
  • pageindex/integrations — 2 commits
  • tests/test_agent_tools.py — 2 commits
  • (root) — 1 commit

Notable commits

  • fix: fix(chat): own-model chat drops a citation box on a page whose picture the call never saw (#673)
  • fix: fix(cloud): CloudAPI.submit_document docstring drops the OCR claim; a rejected mode is proven to send nothing (#658)
  • fix: fix(cloud): Flash uploads keep node text, and near-miss mode spellings raise
  • fix: fix(cloud): Flash uploads rely on the API keeping node text by default
  • fix: fix(cloud): delete_collection waits up to 600 s for the server to delete every document, and list_documents names a collection among the places path is None
  • fix: fix(cloud): mode="standard" is local-only; cloud takes None or "flash"
  • fix: fix(cloud): submit_document(mode=) accepts only "standard" and "flash"
  • fix: fix: get_citations resolves the library paths the MCP server now cites (#632)
  • change: Merge pull request #657 from VectifyAI/feat/cloud-submit-flash
  • change: Merge pull request #674 from VectifyAI/develop
  • change: docs(cloud): a collection without declared fields learns them from its documents' metadata, its documents take 16 keys at most, and get_metadata_schema(collection_id=) reports counts and common values
  • change: feat(chat): chat(collection_id=) keeps own-model cloud chat in one collection, and citations= takes a format name such as cite_with_boxes (#668)
  • change: feat(cloud): collections: create, list and delete them, upload a document into one, list its documents, and declare its metadata fields
  • change: feat(cloud): submit_document(mode="flash") builds the tree with Flash on the cloud
  • change: perf: summaries think the least each model takes, indexing stops paying for cache writes it never reads, and a chosen reasoning effort arrives as chosen (#659)
  • change: test: the live copy checks skip while the cloud addresses documents by path and local mode does not (#671)
  • change: test: the live envelope check passes on the cloud's text outline and on a picture page's extra blocks (#670)

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

Survey your own repository

VectifyAI/PageIndex was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.

About this page

  • The score is its most recent published measurement, taken on 15 November 2026 at a pinned commit. It is not a live figure and does not change until the project is measured again.
  • Measured at commit cc035e8ddcfa51633309ddf2f794b161a17d8cbc — the exact code this score is about.
  • Scored under rubric-2026.11.4 — the same rubric and the same method as every other entry in this index.
  • Measured by Canine Code Analyzer (cca.canine.dev) using cca measure-2026.11.73+preprod-52de90b74216.