pageindex-rag

Vectorless RAG using LLM-driven hierarchical document indexing — for PDF and Markdown files (including Obsidian vaults).

No embeddings. No vector database. The LLM reads the document structure once, builds a tree index, and then navigates that tree to answer questions.

Based on VectifyAI/PageIndex. Extended with multi-LLM provider support, Obsidian vault indexing, and a dependency security checker.

What was added over the original

| Feature | Original | This fork | |---|---|---| | LLM providers | Anthropic only | Anthropic · OpenAI · Ollama · Claude Code (no API key) | | Markdown indexing | — | Single file and full vault/folder | | Obsidian vault support | — | --vault <dir> CLI flag | | npm security rules | — | npm run security (age + CVE checks) | | Env-var model override | — | PAGEINDEX_MODEL in .env |

Installation

npm install @keleshteri/pageindex-rag

Requires Node.js 18+ (uses built-in fetch).

Quick start

import { PageIndexClient } from '@keleshteri/pageindex-rag';

const client = new PageIndexClient({
  model: 'gpt-4o',           // or 'claude-sonnet-4-6', 'claude-code', 'ollama/llama3'
  workspace: './workspace',  // optional: persists indexes to disk
});

// Index a PDF
const docId = await client.index('./report.pdf');

// Index a Markdown file
const noteId = await client.index('./notes.md');

// Index an entire Obsidian vault
const results = await client.indexVault('./MyVault', { concurrency: 3 });

// Retrieve
const structure = client.getDocumentStructure(docId);
const pages     = await client.getPageContent(docId, '3-5');

LLM options

Set the model in .env or pass it directly to PageIndexClient.

Anthropic

ANTHROPIC_API_KEY=sk-ant-...
PAGEINDEX_MODEL=claude-sonnet-4-6

OpenAI

OPENAI_API_KEY=sk-...
PAGEINDEX_MODEL=gpt-4o

Claude Code — no API key needed

Uses your local Claude Code session. Requires the claude CLI to be installed and logged in.

PAGEINDEX_MODEL=claude-code

Ollama — fully local, no API key needed

ollama pull llama3

PAGEINDEX_MODEL=ollama/llama3
# OLLAMA_BASE_URL=http://localhost:11434/v1  # override if needed

CLI

# Single PDF
npx ts-node src/cli.ts --pdf report.pdf

# Single Markdown file
npx ts-node src/cli.ts --md notes.md

# Entire Obsidian vault
npx ts-node src/cli.ts --vault ~/Documents/MyVault

# Use a specific model
npx ts-node src/cli.ts --pdf report.pdf --model gpt-4o

# Options
#   --no-summary        skip per-node summaries (faster)
#   --add-description   generate a one-sentence doc description
#   --add-text          include raw page text in the output
#   --output <path>     custom output path (single-file mode only)

Output is saved to results/ as JSON.

Programmatic API

const client = new PageIndexClient(options);

// Index
await client.index(filePath)               // auto-detects pdf/md
await client.indexVault(dirPath)           // all .md files in a folder

// Retrieve
client.getDocument(docId)                  // metadata (name, type, page count)
client.getDocumentStructure(docId)         // full tree (no raw text)
await client.getPageContent(docId, pages)  // pages: "1-3", "5,8", "12"
client.listDocuments()                     // all indexed docs

The three retrieve functions return JSON strings — they are designed to be passed directly as LLM tool-use callbacks.

Security

npm run security   # checks all deps: age (< 14 days → blocked) + npm audit

See scripts/check-deps.js for thresholds and rules.

Development

npm run build   # compile TypeScript → dist/
npm run dev     # watch mode

Copy .env.example to .env and configure your LLM provider before running.

Credits

This project is a fork of VectifyAI/PageIndex, licensed under MIT.

License

MIT — see LICENSE.