OmniRetrieval: Unified Retrieval across Heterogeneous Knowledge Sources

OmniRetrieval, from KAIST and DeepAuto.ai, argues that retrieval should not collapse text, SQL, SPARQL, and Cypher into one shared embedding space; instead, it uses a three-stage LLM controller that first selects relevant knowledge sources from their native descriptors, then generates source-specific executable queries, and finally consolidates heterogeneous outputs into one evidence set. The system is evaluated on 13 datasets spanning 309 knowledge bases across unstructured corpora, relational databases, RDF knowledge graphs, and labeled property graphs, with 300 questions sampled per dataset. Using GPT-5.4, Gemini-3.1 (Pro), Sonnet-4.6, Qwen-3.5 (27B), and Gemma-4 (31B), OmniRetrieval outperforms KB Routing on the macro-averaged main metrics: source selection rises from 61.65 to 65.71, retrieval accuracy from 39.98 to 44.34, and LLM-as-a-Judge from 57.99 to 65.88. The strongest gains come from deferring commitment until evidence selection, because the gold source is often included among candidates even when the top-1 route is wrong; in multi-candidate questions, evidence-selection accuracy reaches 67.5% at k=3 but falls to 62.8% at k=10, showing the trade-off between broader exploration and selector quality. The paper also reports that a constrained unified-representation baseline remains well below OmniRetrieval, supporting the claim that structural operators such as joins, graph traversals, and property paths must be preserved rather than flattened.