Decoupling Search from Reasoning: A Vendor-Agnostic Grounding Architecture for LLM Agents
For developers of LLM agents who need reliable, cost-effective, and inspectable real-time search grounding, DSG provides a practical solution to the coupling problem inherent in native search integration.
The paper introduces Decoupled Search Grounding (DSG), a vendor-agnostic architecture that separates search grounding from the reasoning model, enabling independent control over retrieval policy, provider routing, caching, and cost. On SimpleQA, DSG achieves 86.1% accuracy vs. 87.7% for native search but at 91% lower search cost, and on an e-commerce workload it matches accuracy while cutting search cost by over 98%.
Production LLM agents increasingly depend on real-time search, yet native search grounding bundles retrieval policy, provider choice, evidence injection, cost, latency, and generation behavior behind a single model-provider boundary. This coupling makes grounding hard to inspect, tune, reuse, or port, and can trigger Search-Induced Verbosity that breaks strict output contracts. We present Decoupled Search Grounding (DSG), a vendor-agnostic boundary that moves grounding outside the reasoning model through an MCP-compatible gateway, exposing provider routing, source-aware context rendering, configured fallback, retrieval-depth control, and exact plus semantic caching as first-class controls. Across five frontier models on SimpleQA, FreshQA, and HotpotQA, native search leads on recency-sensitive FreshQA, but DSG exposes a stronger frontier when control matters: on SimpleQA it nearly matches native accuracy (86.1% vs. 87.7%) at 91% lower search cost, preserves concise answer contracts, and reaches a 99.4% warm-cache hit rate with 68% lower latency. Deployed as a shared production grounding layer for large-scale agentic workloads with interchangeable models, DSG matches or slightly exceeds native-search accuracy on an e-commerce query-understanding (QIU) workload while cutting search cost by over 98%. Real-time grounding is best treated as an optimizable interface boundary, not a fixed model feature.