Instructed Retriever Improves Enterprise Retrieval Accuracy

A new blog post presents the Instructed Retriever, a retrieval architecture that propagates system specifications into both retrieval and generation components to enable precise instruction following for enterprise agents. The blog reports it increases performance by more than 70% over traditional RAG and outperforms a RAG-based multi-step agent by 10%, while preserving low latency and a small model footprint. The architecture supports query decomposition, metadata reasoning, and offline reinforcement learning.