VeritasGraph is an open-source, 100% on-prem GraphRAG framework that combines Tree-Search navigation with Knowledge-Graph reasoning—so your LLM answers multi-hop questions and cites every claim to its source. What's different: it goes beyond retrieval. VeritasGraph Studio wires your knowledge graph into governed AI agents with guardrails, memory, tools, MCP bridges, and context budgeting—running entirely on Ollama, no data egress. Ships an MCP server and a deterministic policy engine.

Hey Product Hunt! 👋 I built VeritasGraph because traditional RAG guesses based on similarity— it chunks your docs blindly and hopes cosine similarity finds the right fragment. VeritasGraph instead reasons over structure: it builds a knowledge graph + hierarchical tree, does true multi-hop reasoning, and attributes every single claim back to [doc#chunk] sources. But retrieval was only half the problem. Enterprises don't just want answers—they want *governed* answers. So VeritasGraph Studio lets you wire your graph into AI agents with guardrails (PII redaction, policy blocks), memory, real HTTP tools, MCP integrations, and headroom-style context budgeting—then watch every stage of the orchestration pipeline live. It runs 100% locally on Ollama, so your data never leaves your infrastructure. It's fully open source (MIT), ships an MCP server for Claude/Cursor/VS Code, and includes VeritasReason for deterministic, auditable policy compliance. 📄 Full guide "Build Governed AI Agents On-Prem": https://github.com/bibinprathap/... 🎬 Walkthrough: https://youtu.be/sA7ReEgdJfg Would love your feedback—especially from anyone building enterprise search, compliance assistants, or agent memory systems. AMA! 🚀

How does the deterministic policy engine actually work under the hood when the underlying LLM is non-deterministic by nature, especially with Ollama models that vary across runs?

how does the deterministic policy engine actually handle conflicts when two cited sources disagree on a fact?

How does the deterministic policy engine actually work under the hood when the underlying LLM is non-deterministic by nature, especially with Ollama models that vary across runs?

how does the deterministic policy engine actually handle conflicts when two cited sources disagree on a fact?

VeritasGraph

Governed on-prem GraphRAG & AI agents with verifiable proof

Governed on-prem GraphRAG & AI agents with verifiable proof