DSG: Decoupled Search Grounding for LLM Agents, 98% Cost Cut

Decoupled Search Grounding, or DSG, moves grounding out of the reasoning model and into a vendor-agnostic, MCP-compatible gateway that exposes provider routing, source-aware context rendering, configured fallback, retrieval-depth control, and exact plus semantic caching. The paper, submitted 17 June 2026, evaluates DSG across five frontier models and three QA benchmarks and reports large cost and latency savings while preserving accuracy in many settings.

What is DSG and how does it work?

DSG is a vendor-agnostic grounding boundary that separates search from reasoning by routing queries through an MCP-compatible gateway and making retrieval controls explicit. The gateway surfaces provider routing, source-aware context rendering, configured fallback, retrieval-depth control, and exact plus semantic caching as first-class controls. The architecture treats real-time grounding as an optimizable interface boundary rather than a fixed property of a single model-provider pairing.

How does DSG perform versus native search?

On SimpleQA, DSG nearly matches native-search accuracy at 86.1% versus 87.7%, while reducing search cost by 91% relative to native search. DSG also preserves concise answer contracts and achieved a 99.4% warm-cache hit rate with 68% lower latency. Across the three evaluated QA datasets, the authors note that native search leads on the recency-sensitive FreshQA benchmark, though DSG exposes a stronger frontier when explicit control over grounding matters. Deployed as a shared production grounding layer for large-scale agentic workloads, DSG matched or slightly exceeded native-search accuracy on an e-commerce query-understanding workload while cutting search cost by over 98%.

Why it matters

Decoupling grounding from the reasoning model turns implicit, provider-tied behavior into explicit controls. The paper's core claim is operational: by exposing routing, fallback, retrieval depth, and caching, teams can inspect, tune, reuse, or port grounding without being locked to a single model-provider boundary. The empirical points underline that claim: an accuracy gap of 1.6 percentage points on SimpleQA comes with a 91% drop in search cost and large latency gains, and production deployment yielded over 98% cost savings on an e-commerce workload while maintaining accuracy.

What to watch

Observe whether DSG narrows the gap on recency-sensitive FreshQA cases where the paper reports native search leads. Also watch broader adoption of the MCP-compatible gateway pattern: the next signals will be independent reproductions of the 99.4% warm-cache hit rate, the 68% latency reduction, and multi-workload cost savings outside the paper's reported e-commerce deployment.