VeritasGraph

VeritasGraph is an enterprise-grade GraphRAG framework built on the principle "Don't Chunk. Graph." — ingesting whole pages or sections as graph nodes instead of traditional 500-token chunks, preserving document structural integrity. The framework employs a dual Tree + Graph retrieval architecture: PageIndex-style hierarchical TOC navigation runs in parallel with knowledge graph semantic reasoning, supporting cross-section linking and multi-hop reasoning for complex cross-document questions. Every generated claim provides 100% verifiable attribution traceable to exact source document locations, making it suitable for high-compliance domains such as legal, medical, and financial sectors.

Retrieval & Reasoning#

• Tree-based Navigation: PageIndex-style hierarchical TOC navigation with cross-section linking
• Graph-based Semantic Search: Knowledge graph-connected semantic retrieval, not mere vector similarity matching
• Multi-hop Reasoning: Complex reasoning across documents and sections
• Document-Centric Ingestion: Whole pages/sections as nodes, avoiding context loss from chunking

Ingestion Sources#

• PDF: Via pipeline.ingest_pdf() or CLI veritasgraph ingest
• YouTube: Automatic subtitle extraction from URL
• Web Articles: Direct URL ingestion via CLI
• Plain Text: Standard text ingestion
• Charts/Tables: Vision RAG mode converts to knowledge graph nodes

Verifiability & Visualization#

• Verifiable Attribution: Every claim includes a precise attribution path traceable to exact source locations
• Interactive Graph Visualization: PyVis-powered 2D graph browser showing entities, relations, and reasoning paths in real time

Deployment Modes#

LLM/Embedding Compatibility#

Unified through an OpenAI-compatible API abstraction, supporting mixed configurations (e.g., Groq for LLM + Ollama for Embeddings): OpenAI, Azure OpenAI, Groq, Together AI, OpenRouter, LM Studio, vLLM, Ollama.

Architecture#

• Graph Engine Layer: Based on Microsoft GraphRAG for indexing and querying, Neo4j as persistent graph database
• Retrieval Layer: Tree-based navigation and Graph-based semantic search running in parallel
• Document Processing Layer: Document-centric ingestion with whole pages/sections as single retrievable nodes
• LLM Abstraction Layer: OpenAI-compatible API interface unifying multiple local/cloud LLM providers
• Visualization Layer: PyVis interactive 2D graph browser, Gradio Web UI

Installation & Quick Start#

pip install veritasgraph
veritasgraph demo --mode=lite

Optional dependencies: veritasgraph[web] (Gradio UI + visualization), veritasgraph[graphrag] (Microsoft GraphRAG integration), veritasgraph[ingest] (YouTube & web ingestion), veritasgraph[all] (all features).

Docker one-click deployment (full mode):

cd docker/five-minute-magic-onboarding
docker compose up --build
# Ports: Gradio UI :7860, Neo4j Browser :7474, Ollama API :11434

Python API Example#

from veritasgraph import VisionRAGPipeline, VisionRAGConfig

pipeline = VisionRAGPipeline()
doc = pipeline.ingest_pdf("document.pdf")
result = pipeline.query("What are the key findings?")
print(result.answer)

config = VisionRAGConfig(ingest_mode="document-centric")
pipeline = VisionRAGPipeline(config)
doc = pipeline.ingest_pdf("annual_report.pdf")
print(pipeline.get_document_tree())
section = pipeline.navigate_to_section("Methodology")

Key Environment Variables#

Unconfirmed Information#

• Exact PyPI version and release date (JS rendering limitation on PyPI page)
• Independent website/docs URL (README mentions "Live documentation" but no URL provided)
• Formal publication of the accompanying paper (PDF in repo, no arXiv or journal link found)
• Deployed HuggingFace Space address
• Whether GPU is mandatory for local mode
• Performance benchmarks for large-scale document sets