Multi-hop retrieval typically requires heavy machinery: dense embeddings, vector indexes, and distributed databases. The new SAG2 architecture, described in an arXiv preprint, argues that strong multi-hop retrieval should be practical to deploy without such expense. Its answer is an embedding-free design built around a single SQLite database.
The paper names two components behind SAG2: BM25, a well-known lexical ranking function, and JEV, whose role is not expanded in the abstract. By avoiding embeddings, the system sidesteps the need for expensive corpus preparation and complex database infrastructure, potentially making multi-hop search far easier to run in ordinary settings.
The abstract is brief and does not include experimental results or a full explanation of JEV, so the practical benefits are asserted rather than demonstrated in the available text. Still, the proposal points toward a leaner direction for retrieval systems that need to answer questions requiring multiple steps of evidence gathering.