Show HN: VeritasGraph,An on-premise Graph RAG with verifiable source attribution(github.com)
github.com
Show HN: VeritasGraph,An on-premise Graph RAG with verifiable source attribution
https://github.com/bibinprathap/VeritasGraph
1 comments
Very nice project. Consider looking into a graph db with low latency for nice user experience for the question/answer portion of the project. For example: https://github.com/FalkorDB/falkordb
I'm the creator of VeritasGraph. I've been working with RAG pipelines and became frustrated with two major limitations of the standard vector-search approach: its inability to answer complex questions that require connecting multiple pieces of information (multi-hop reasoning), and the "black box" nature of responses that makes them hard to trust or verify.
VeritasGraph is my attempt to solve this. It's an end-to-end framework that runs entirely on your own hardware using local LLMs like Llama 3.1 via Ollama, ensuring complete data privacy.
How it works:
Indexing: It ingests your documents (PDFs, text, etc.) and uses an LLM to automatically build a knowledge graph, extracting key entities and their relationships.
Retrieval: When you ask a question, it uses a hybrid approach. First, a vector search finds the most relevant starting points in the graph. Then, it traverses the graph's relationships to gather the full context needed to answer multi-hop questions.
Generation: The rich, graph-based context is fed to a LoRA-tuned LLM to generate an answer. Crucially, the system is designed to trace every part of the answer back to the original source text, providing verifiable attribution.
The goal was to build something that enterprises could actually trust for high-stakes use cases, where data privacy and factual accuracy are non-negotiable.
The repo includes a Gradio UI for chatting with your data and a guide on how to set everything up with Ollama. I've also included the code for fine-tuning the LLM with LoRA for efficiency.
This is still a work in progress, and I'd love to get your feedback, suggestions, or even contributions. Happy to answer any questions about the architecture or the challenges I ran into.