Context as Infrastructure: Knowledge Compilation for Retrieval in Agent Harness (arxiv.org)

A new study titled "Context as Infrastructure: Knowledge Compilation for Retrieval in Agentic Systems with Harnesses" explores the potential of retrieval-augmented generation (RAG) in enhancing the capabilities of AI agents. The research reveals that leveraging a search harness with simple tools like BM25 and MiniMax M3 yields impressive accuracy rates—85% on the BrowseComp-Plus benchmark and 91% on multi-hop FRAMES. This approach highlights an emerging problem in AI context management: the inefficiency caused by repetitive querying and document rereading, which often results in cluttered information without enhancing precision. The authors propose a solution through "knowledge compilation," a structured ingestion pipeline that enriches documents with essential contextual information such as navigation hierarchies, entity graphs, and timelines. This not only streamlines the querying process—reducing token usage by up to 91%—but also significantly decreases latency from 140 seconds to just 30 seconds during retrieval. While the structural artifacts introduced do not improve accuracy directly, they play a vital role in operational contexts, influencing governance and integration. The study aims to redefine how AI agents manage and utilize context, presenting RAGFlow's open-source pipeline that transforms conventional retrieval methods into a more efficient, agent-native database system.