Show HN: VeritasGraph – On-prem Graph RAG (3.3k+ visitors, 130 stars in 5 days) (github.com)

VeritasGraph is an open-source, production-ready on‑prem framework for Retrieval‑Augmented Generation that replaces pure vector search with a hybrid pipeline: LLM-powered entity/relationship extraction builds a knowledge graph (e.g., Neo4j) alongside a vector index, then a hybrid retriever uses vector search for entry points and multi‑hop graph traversal to assemble rich context that a LoRA‑tuned local LLM uses to generate answers with end‑to‑end provenance. The Show HN project (3.3k+ visitors, 130 stars in five days) emphasizes enterprise requirements—100% data sovereignty, auditable reasoning paths, JSON-structured outputs with source IDs, and hardware/software guidance (CPU 16+ cores, 64–128GB RAM, 24GB+ GPU; Docker, Python 3.10+, NVIDIA toolkit). The repo gives a hands‑on stack using Ollama (llama3.1-12k) and nomic-text-embed, plus practical notes about context lengths, model pulls, indexing, reindexing caveats, and a Gradio demo. This matters because traditional vector RAG can surface direct hits but struggles on multi‑hop queries that require connecting disjoint facts; VeritasGraph explicitly combines semantic search, graph traversal, pruning/re‑ranking, and LoRA fine‑tuning to produce attributed, traceable answers—useful for compliance, research synthesis, and complex enterprise QA (e.g., linking engineers across projects/patents). Planned enhancements include broader DB support, graph analytics, agentic workflows, and visualization—positioning VeritasGraph as a practical path to trustworthy, on‑prem knowledge assets and reduced vendor lock‑in.