DynaTree: Dynamic Agentic Tree boosts news recall by 40%

DynaTree tackles a key weakness in agentic RAG: high inference cost and poor suitability for time-sensitive news. Existing methods intertwine semantic expansion and retrieval decisions in short-horizon loops, leading to expensive online reasoning and stale coverage. The new two-stage approach separates concerns. In the offline stage, multiple coordinated agents explore a query topic's semantic space to construct a reusable retrieval tree—essentially materializing all possible retrieval paths without real-time cost. In the online stage, only a lightweight daily subtree selection runs on a time-localized evaluation proxy, requiring no further agentic reasoning, tree modification, or retraining.

Experiments on the multi-day Syft news benchmark and several BEIR datasets show DynaTree achieves strong recall and ranking, consistently beating standard RAG and previous agentic baselines. The real test came in production: DynaTree was deployed in Syft’s system from Jan. 28 to Feb. 6, 2026. Its dynamically adapted variant improved survival rate (a key relevance metric) from a fixed offline subtree's 0.32–0.53 to 0.59–0.73, and outperformed existing production recallers on every single evaluation day. This demonstrates that persistent, structure-aware semantic expansion can translate offline reasoning into practical gains in coverage, freshness, and relevance for real-world news retrieval.