SimpleRAG - Upload Documents

Upload Documents

Upload documents to index them using your preferred RAG mode.

Current Default RAG Mode: Normal RAG
You can override this setting for individual uploads below.

Select Document

Supported formats: PDF, TXT, DOCX, HTML

RAG Mode for this Document

Normal RAG - Fast semantic search

Graph RAG - Extract entities and relationships

⚠️ Graph RAG takes longer but provides richer context

Neo4j Only - Store entities/relationships in Neo4j graph database

📊 Extracts and stores knowledge graph in Neo4j without vector embeddings

PageIndex RAG NEW Vectorless - Reasoning-based retrieval, no chunking or vector DB

🌲 Builds a hierarchical tree index (like a smart Table of Contents) from your PDF. Best for long professional documents: annual reports, research papers, legal filings.
⏱ Indexing time: ~2–5 min per 100 pages (one-time, LLM-based tree construction). No embeddings generated.

Normal RAG Process

Text extraction from document
Split into overlapping chunks
Generate embeddings for chunks
Store in vector database

Graph RAG Process

Text extraction from document
Extract entities and relationships
Build knowledge graph
Generate embeddings for graph elements
Store both chunks and graph in database

Note: Graph RAG processing takes significantly longer than Normal RAG due to entity extraction and relationship mapping, but provides much richer contextual understanding.

🌲 PageIndex RAG — How It Works NEW — Vectorless

Index Phase (one-time, ~2–5 min)

LLM reads the full document
Builds a hierarchical tree (like a Table of Contents)
Each node gets a title, page range, and summary
Tree saved as JSON — no vector DB needed

Query Phase (~10–20 s)

LLM agent inspects the tree structure
Reasons about which sections are relevant
Fetches exact page text (tight ranges only)
Synthesises answer with traceable page citations

Best for: Annual reports, SEC filings, research papers, legal manuals, technical specifications — any long document where exact section navigation matters. PageIndex achieved 98.7% accuracy on the FinanceBench benchmark.