VectorLiteRAG introduces an adaptive vector index partitioning strategy that optimizes memory allocation between CPU and GPU for faster, more efficient Retrieval Augmented Generation pipelines.

• Reduces end-to-end latency by strategically distributing vector indices between CPU and GPU memory
• Adapts automatically to different hardware configurations and query patterns
• Balances workload effectively between compute resources for improved throughput
• Addresses the engineering challenge of integrating vector search with LLM inference

This research offers immediate practical value for engineers building production RAG systems, enabling more responsive AI applications without requiring hardware upgrades.

An Adaptive Vector Index Partitioning Scheme for Low-Latency RAG Pipeline