A research system called SemJoin takes aim at one of the messier problems in modern data engineering: joining database tables using plain-English conditions without burning a fortune on LLM calls.
The core problem is arithmetic. A naive semantic join — asking a language model whether each row in table A matches each row in table B — scales at O(M × N) LLM calls. For anything larger than a toy dataset, that's cost-prohibitive. Existing systems cut that bill by locking into a single strategy: embedding similarity, batched comparisons, or some fixed heuristic. SemJoin's authors argue that single-strategy systems leave accuracy and efficiency on the table because different join predicates have different shapes. Their fix is a routing layer: an LLM advisor that inspects each join and dispatches it to one of two execution strategies — a Cluster Join that prunes candidates through unsupervised embedding clustering, or a Classifier strategy for predicates that collapse to a discrete label set.
The benchmark results are specific enough to take seriously. Across three test datasets — IMDb reviews, email contradictions, and Stack Overflow tags — the advisor picked the optimal strategy every time. Against adaptive block join (ABJ), SemJoin improved F1 scores by 20 to 33 points across all three datasets, and used fewer tokens on two of the three. Against featurized-decomposition join (FDJ), it matched or beat accuracy at one to two orders of magnitude lower token cost.
The broader signal here is that LLM-powered databases are moving from proof-of-concept to engineering discipline — and that smart routing, not raw model power, may be where the real efficiency gains live.