RAGFlow 1.0 preview introduces agent-native database concept based on knowledge compilation paper

Details

RAGFlow 1.0 is positioned as an agent-native database rather than a traditional RAG pipeline, based on the paper "Context as Infrastructure: Knowledge Compilation for Retrieval in Agentic Systems with Harnesses".
The core thesis is that retrieval is a harness problem before it is an index problem; the index determines what can be found, while the harness determines whether finding becomes an answer.
Testing of structural artifacts including summary trees, outlines, wikis, and graphs under a competent harness showed no detectable accuracy gain, leading to the conclusion that structure is not additional evidence.
Knowledge compilation is reframed as a supply layer that enriches source documents into typed artifacts (claims, entity relations, timelines, ontologies, wikis, memory) at ingestion time.
The system supplies provenance-bearing context that is compiled once and reused, governed by the harness to provide the smallest sufficient set of verified facts.
Deployment outcomes for claims, graph, and wiki are currently marked as "no measured effect" because none have met pre-registered targets, with cost savings noted as being bought with retrieval rounds.

Key point