bibinprathap/VeritasGraph
VeritasGraph — open-source Knowledge Graph & GraphRAG framework on GitHub. Build multi-hop reasoning, ontology-aware retrieval, and verifiable attribution over your own data. Nodes, edges, RDF, linked-data — runs locally or in the cloud.
What it solves
VeritasGraph addresses the limitations of traditional Vector RAG, which often relies on simple similarity searches and "blind chunking." It provides a way to perform multi-hop reasoning and verifiable attribution by combining hierarchical document structures (like a Table of Contents) with a semantic Knowledge Graph, allowing AI agents to understand connections between disparate pieces of information rather than just finding similar text.
How it works
The system uses a hybrid retrieval engine that combines vector search for entry points with multi-hop graph traversal to uncover hidden relationships. It follows a four-stage pipeline: indexing documents into text units and extracting entities/relationships into a graph database (like Neo4j), retrieving context through hybrid search, processing it via a reasoning core (often a LoRA-tuned local LLM), and finally producing answers with a structured attribution layer for provenance.
Who it’s for
It is designed for developers and enterprises requiring sovereign, on-premise AI deployments where data privacy is critical and answers must be 100% verifiable with citations. It is particularly useful for those building governed AI agents that need to interact with complex, structured documents and external tools.
Highlights
- Hybrid Retrieval: Combines tree-based navigation (TOC) and graph-based semantic reasoning.
- Governed Agent Studio: A local workspace to build, wire, and test agents with integrated guardrails, memory, and tool catalogs.
- MCP Server: A Model Context Protocol server that allows IDE agents (like Cursor or VS Code) to connect to the GraphRAG engine without data egress.
- VeritasReason Integration: A sister module for deterministic rule-evaluation over structured records to handle compliance and policy violations.
- Local-First: Supports 100% offline operation via Ollama and Neo4j.