Authors - Shivam Kumar, Dinesh Kumar Saini Abstract - KG-RAG (Knowledge Graph-Retrieval Augmented Generation) is an advanced AI framework that combines structural knowledge graphs with LLMs to make them smarter, more accurate, robust, and less prone to hallucination. However, existing KG-RAG pipelines are often tightly coupled with specific domains. In addition, most of the systems lack proper schema validation and have limited support for temporal knowledge. GenericKG is a modular framework designed to decouple knowledge ingestion, validation, storage and retrieval across domains. The framework includes an agentic ingestion pipeline with schema-driven knowledge graph construction, supported by multi-level validation (L1-L3) to ensure structural, semantic and temporal consistency. Temporal attributes and semantic embeddings are integrated at framework level, enabling time-aware querying and hybrid retrieval without domain-specific reengineering. This paper is evaluated on three benchmarks: the BC5CDR biomedical corpus (87.92% entity F1 with 100% precision), the WebNLG crossdomain dataset (85.6% entity F1 across 15+ relation types on 100 records), and HotpotQA multi-hop question answering (58.0% accuracy on bridge and comparison questions). A raw-LLM baseline without schema guidance scores 0% on all metrics, confirming the importance of the schemadriven pipeline. This framework is implemented in TypeScript and it will be released as open source.
Open Zoom