KRUTRIM RAG: A Fully Offline Hybrid Retrieval-Augmented Generation Architecture for Enterprise Knowledge Retrieval and Question Answering

Résumé fourni par la source

Retrieval-Augmented Generation (RAG) has emerged as an effective approach for enhancing Large Language Models (LLMs) through the integration of external knowledge sources during response generation. However, most existing RAG frameworks rely on cloud-based infrastructures, resulting in challenges related to data privacy, security, latency, and internet dependency. These limitations hinder their adoption in enterprise, governmental, and research environments where sensitive information requires strict access control and local processing. This paper presents Krutrim RAG, a fully offline hybrid retrieval framework designed to enable secure, scalable, and context-aware knowledge retrieval without reliance on external network services. The proposed architecture combines semantic vector retrieval using Qdrant with graph-based knowledge retrieval using Neo4j, allowing the system to leverage both semantic similarity and structured relational knowledge. In addition, locally deployed embedding models and large language models ensure complete data sovereignty, while a parallel load-balanced ingestion mechanism optimizes resource utilization and processing efficiency. By integrating vector search and knowledge graph reasoning within a unified offline environment, Krutrim RAG improves retrieval accuracy, contextual relevance, and response quality. The key contributions of this work include the design of a fully offline RAG architecture, the integration of synchronized dual-path ingestion with deterministic metadata alignment, the implementation of automated load-balanced parallel ingestion, and the validation of the framework for privacy-preserving enterprise knowledge management and question-answering applications. Research conducted at the Space Applications Centre (SAC), Indian Space Research Organisation (ISRO), Ahmedabad, India.