Substrata: Know What You Don't Know

Retrieval-augmented generation (RAG) reduces hallucinations by grounding large language models in external knowledge bases. However, most existing methods focus on improving how knowledge is indexed, retrieved and integrated, neglecting a critical source of hallucination: the lack of unknown awareness. Unaware of the knowledge boundaries of the corpus, a model may stitch together fragmented context to hallucinate unsupported relations. To address this, we introduce Substrata, an unknown-aware RAG framework that makes the unknown locatable. Substrata maintains an evolving knowledge graph via Dempster--Shafer evidence fusion, enabling local graph updates instead of costly full graph reconstruction in graph-based RAG. Crucially, we apply a topological analysis method, persistent homology, to detect long-lived structural holes where the corpus fails to close relational gaps within this graph. These holes are converted into retrievable unknown annotations. During inference, Substrata performs hybrid retrieval over both factual chunks and unknown annotations, allowing the generator to access both supporting context and explicit signals of missing contextual support. Experiments show that Substrata achieves the best performance on three public multi-hop QA benchmarks, with an average of 60.66\% EM and 71.25\% F1 across HotpotQA, MuSiQue, and 2WikiMultiHopQA. On a hallucination-focused domain benchmark, it reaches 95.0\% accuracy on Spurious Relational Premise questions while reducing overall token consumption by 10.3× compared with GraphRAG.