v0.1.5 — Smart Hierarchy Extraction for Large Documents

What's New

This release fixes hierarchy extraction accuracy on large (300+ page) documents. Previously, subsections would get flattened to top-level sections — now TreeDex uses multiple strategies to maintain correct depth.

New Features

• PDF ToC extraction — If the PDF has bookmarks/outline, the tree is built directly from them — zero LLM calls needed, perfect hierarchy every time
• Font-size heading detection — Analyzes font sizes across the document and injects [H1]/[H2]/[H3] markers so the LLM knows exactly which level each heading belongs to
• Capped continuation context — For multi-chunk documents, the LLM sees a compact summary (top-level outline + last 30 sections) instead of the full history — 78% fewer tokens wasted on context
• Orphan repair — If the LLM outputs "2.3.1" without a "2.3" parent, synthetic parents are auto-inserted to maintain a valid tree

New Exports

Function	Python	Node.js
Extract PDF ToC	extract_toc(path)	await extractToc(path)
ToC → sections	toc_to_sections(toc)	tocToSections(toc)
Repair orphans	repair_orphans(sections)	repairOrphans(sections)

How It Helps

Before — LLM sees flat text, guesses hierarchy:

1 Introduction  1.1 Background  Large Language Models...

After — LLM sees font-size markers, knows exact depth:

[H2] 1 Introduction
[H3] 1.1 Background
Large Language Models...

Impact on Large Documents (1M+ tokens)

LLM Context	Groups	LLM Calls	Continuation Tokens
20k (default)	56	56	1,336 (was 5,946)
128k (large)	8	8	minimal
PDF with ToC	0	0	N/A

Full Changelog

• pdf_parser / pdf-parser — ToC extraction, heading analysis, annotated text builder
• tree_builder / tree-builder — toc_to_sections(), repair_orphans()
• prompts — Updated to reference [H1]/[H2]/[H3] heading markers
• core — ToC shortcut path, heading detection, capped context, orphan repair
• loaders — Pass-through for detect_headings option
• Tests — New test coverage for all new functionality
• README.md — Updated How It Works section and API reference
• how-treedex-works.svg — Updated pipeline diagram

19 files changed, 1,012 insertions(+), 102 deletions(-)

v0.1.4

What's New

Web Demo — Chat UI + Caching + Vercel Deploy

• Chat-style UI: Two-panel layout with sidebar + chat bubbles (user/AI), typing indicator, collapsible sources
• Main upload zone: File upload front-and-center in the chat area
• Per-file progress: Upload multiple files with real-time indexing status
• Disk cache (.cache/): Re-uploading the same file is instant; cached docs auto-restore on restart
• Auto-retry on 429: Rate-limited Groq calls automatically wait and retry (up to 8 attempts)
• Conversational fallback: Greetings get natural responses instead of "no info found"
• Vercel serverless: Full api/ functions with client-side IndexedDB state
• Mobile responsive: Sidebar hamburger menu, keyboard-aware input bar
• dotenv config: .env.example with GROQ_API_KEY, PORT, LLM_MODEL, LLM_BASE_URL

Install

npm install treedex@0.1.4
pip install treedex==0.1.4

v0.1.2

New Features

• Agentic RAG mode — query(question, agentic=True) retrieves relevant sections then generates a direct LLM answer. Available in both Python and Node.js.
• Multi-document support — Web demo now supports uploading and querying across multiple documents simultaneously
• Answer prompt — New answerPrompt / ANSWER_PROMPT template for answer generation

Bug Fixes

• Fix page range assignment — Sections starting on the same page no longer get inverted ranges (e.g. "pages 10-9"), which caused empty context text
• Fix PDF loading — Convert Buffer to Uint8Array for pdfjs-dist compatibility
• Fix multer file upload — Preserve original file extension for format detection
• Increase LLM timeout — Default 5 min (configurable via timeout option on OpenAICompatibleLLM)

Other

• Shorter retrieval reasoning (one sentence instead of verbose paragraph)
• Improved answer prompt with explicit instructions to extract facts from context
• Web demo switched from NVIDIA Kimi K2.5 to Groq for fast inference
• Updated Colab notebook with agentic mode examples
• Updated README with agentic RAG documentation

Full Changelog: v0.1.1...v0.1.2

v0.1.1

Bug Fixes

• Fix PDF loading — Convert Buffer to Uint8Array for pdfjs-dist compatibility
• Fix web demo file upload — Preserve file extension after multer upload so autoLoader can detect format
• Increase LLM request timeout — Default timeout raised from 2 min to 5 min; now configurable via timeout option on OpenAICompatibleLLM

Other

• Web demo now uses published treedex@0.1.1 from npm instead of local file link

Full Changelog: v0.1.0...v0.1.1

v0.1.0 — Initial Release

TreeDex v0.1.0

Tree-based, vectorless document RAG framework.

Highlights

• Tree-based indexing — preserves document hierarchy (chapters, sections, subsections)
• 18+ LLM backends — Gemini, OpenAI, Claude, Groq, Together AI, Fireworks, DeepSeek, Ollama, and any OpenAI-compatible endpoint
• Zero vector dependencies — no embeddings, no vector DB, just JSON
• Exact page attribution — every answer traces back to source pages
• 4 document formats — PDF, TXT, HTML, DOCX

Install

pip install treedex

Quick Start

from treedex import TreeDex, GeminiLLM

llm = GeminiLLM(api_key="YOUR_KEY")
index = TreeDex.from_file("document.pdf", llm=llm)
result = index.query("What is the main argument?")
print(result.context)
print(result.pages_str)

What's Included

• treedex/ — Core library (pdf_parser, tree_builder, loaders, llm_backends, prompts, core)
• examples/ — Quick start examples + sample index
• tests/ — Full test suite
• benchmarks/ — TreeDex vs ChromaDB vs Naive comparison (auto-run in CI)
• assets/ — SVG charts auto-generated from real benchmarks