A research team has proposed a knowledge-graph retrieval method that stays query-aware at every step — without the memory overhead most competing systems require.
Most graph-based retrieval-augmented generation systems use the question only to pick starting nodes, then wander the graph on structural cues alone, blind to what was actually asked. The dominant query-aware alternative, QAFD-RAG, fixes that with a flow-diffusion solver, but the approach demands loading the entire graph into Python memory and runs an iterative solver with an unpredictable step count — awkward to integrate with a live database. The new method replaces that machinery with a single semantic gate: at each traversal step, a cosine similarity score between a candidate node's description and the original question controls how far activation spreads. The entire retrieval pipeline — seed selection, propagation, top-K ranking, context assembly — runs as one Cypher query in a single round-trip to Neo4j.
The engineering tradeoff matters beyond benchmark scores. Keeping the graph inside the database rather than hoisting it into application memory changes the operational profile of a production system: less RAM, fewer moving parts, and a fixed iteration count that makes latency predictable. On MuSiQue, the method lands within one exact-match point of QAFD-RAG (32.80 vs. 33.50) while beating the structural baseline HippoRAG by 5.3 EM and 3.4 F1; an ablation with the semantic gate disabled shows it accounts for 3.6 to 7.4 F1 points of gain and a 1.5x to 4.9x latency reduction.
The approach does not universally dominate — HippoRAG and QAFD-RAG hold an edge on 2WikiMultiHopQA thanks to their phrase-node designs — so teams working on that benchmark should treat this as one option among several, not a drop-in replacement.