# laya-jev-GraphRAG

An Agentic GraphRAG framework that separates the AI decision layer from the graph database, running the same ingestion, retrieval and evaluation pipeline across Neo4j, Memgraph, Apache AGE and Kuzu.

laya-jev-GraphRAG is a graph-database-agnostic framework that adds an agentic retrieval-augmented generation layer on top of an existing graph database rather than replacing it. The same four-phase pipeline of ingestion, pre-retrieval, traversal and post-retrieval, along with continuous evaluation, runs identically across Neo4j, Memgraph, Apache AGE and Kuzu, selected through one environment variable.

The framework evaluates edges and relationships throughout the pipeline: during ingestion it scores and prunes extracted relationships, during traversal it uses a custom A* search guided by those scores, and during synthesis it applies a hallucination gate before producing an answer. Each of these decisions is handled by swappable local or cloud judgment models rather than a full LLM call, using three primitives: a numeric edge score, a binary yes/no judgment, and a categorical choice, so that routing and verification decisions do not require a slow generative call for every edge.

It is written in Python under the Apache-2.0 license.

Website: https://github.com/bodepudimuneendra-netizen/laya-jev-GraphRAG