🌲 Agentic RAG System

Vectorless • Reasoning-Based • Human-Like Retrieval

🔗 Visit this

---

📌 Overview

A scalable, agentic document intelligence system built on top of a locally deployed PageIndex instance for document parsing, designed to process long documents and enable reasoning-driven retrieval instead of vector similarity search.

---

🧠 What This Project Actually Does

❌ Most RAG Systems	✅ This System
Chunk documents	Builds a structured tree index (like a Table of Contents)
Store embeddings	Uses LLM reasoning to navigate the tree
Retrieve by similarity	Fetches only what is needed, when needed

---

⚡ Core Idea

Instead of asking:

"Which chunk is similar?"

We ask:

"Where should I look, and why?"

---

🏗️ System Architecture

1. User Layer

• Upload document (PDF, long text)
• Ask questions via chat

2. Validation Layer

• File validation
• Security checks
• Format normalization

3. Async Processing (Queue System)

Upload goes into queue → Background processing:

• Parsing — handled by locally deployed PageIndex
• Structuring
• Index generation

4. 🌲 PageIndex Tree Generation (Locally Deployed)

Documents are parsed and indexed by a self-hosted PageIndex instance, which converts them into a hierarchical tree structure:

{
  "title": "Section",
  "summary": "...",
  "nodes": []
}

Storage:

• Tree Nodes → MongoDB
• Raw Pages → S3

5. 🧠 Agentic Query System (LLM Reasoning)

Step 1: Intent Understanding

LLM decides:

• Is this a simple question?
• Does it require document reasoning?

Step 2: 🌲 Tree Navigation (Core Innovation)

Instead of vector search:

• Traverse tree like a human
• Section → Subsection → Page
• Use summaries to guide decisions

Step 3: ⚡ Smart Retrieval Strategy

Scenario	Action
Simple query	Answer directly
Node-level sufficient	Fetch structured nodes
Deep reasoning needed	Fetch raw pages from S3

Step 4: 🔍 Cross-Node Reasoning

• Combine multiple nodes
• Use cross-page context
• Perform multi-step reasoning

6. Response Generation

Relevant Nodes + Raw Context → LLM → Final Answer

---

🚀 Why This Is Different

❌ Traditional RAG Problems	✅ This System Solves That
Chunking breaks context	No Vector DB
Embeddings miss true relevance	No Chunking
Hard to explain retrieval	Reasoning-Based Retrieval
Expensive at scale	Explainable (traceable path in tree)
	Human-like navigation

---

🧩 Key Features

• 🌲 Tree-Based Indexing
• 🧠 LLM as Decision Engine
• ⚡ Adaptive Data Fetching
• 🔄 Cross-Page Reasoning
• 📦 Scalable Processing

---

🛠️ Tech Stack

Layer	Technology
Backend	Node.js / Express
Document Parser	PageIndex (locally deployed)
Queue	BullMQ / RabbitMQ
Storage	S3 (raw documents), MongoDB (tree index)
LLM	OpenAI / local models
Architecture	Agentic workflow

---

📊 Conceptual Flow

User Upload → Queue → PageIndex (Local) → Tree Index Creation
                                                    ↓
                  User Query → LLM Reasoning → Tree Traversal
                                                    ↓
                                        Fetch Nodes / Raw Pages
                                                    ↓
                                              Final Answer

---

🔬 Inspiration & Core Dependency

• PageIndex — used as the local document parsing engine
• Agentic retrieval systems
• Human expert document navigation patterns

💡 Positioning

❌ This is NOT	✅ This IS
A chatbot	🧠 A reasoning-first retrieval system
A simple RAG pipeline	for long, complex documents