Natural Language Query to Configuration for Retrieval Agents

A new system called BRANE can dynamically select the most cost-effective configuration for retrieval agents on a per-query basis, matching the accuracy of the best static setup while cutting costs by up to 89 percent, according to a paper submitted on 26 May 2026 [1]. Modern retrieval agents, a class of AI agents that use tools and take actions to pursue goals [3], rely on numerous configurable components including the large language model (LLM), the retriever, the number of documents fetched, and the synthesis strategy [1]. These pipelines are typically hand-tuned once for an entire workload, a practice that leaves significant per-query optimization unrealized [1]. The proposed system, BRANE, addresses this by formulating the task as selecting a configuration from a predefined catalog that either minimizes cost for a given accuracy target or maximizes accuracy under a budget constraint [1]. BRANE operates in two stages. It first uses an LLM to convert each natural-language query into workload-specific characteristics. It then trains a lightweight predictor for each pipeline configuration to estimate whether that configuration will answer the query correctly [1]. At inference time, BRANE selects the configuration that maximizes the predicted correctness, penalized by its serving cost, creating a tunable cost-quality trade-off without requiring retraining [1]. The system was evaluated across three datasets: MuSiQue, BrowseComp-Plus, and FinanceBench [1]. Language model benchmarks such as these provide standardized datasets and evaluation metrics to compare model capabilities on tasks like question answering and reasoning [5]. On these benchmarks, BRANE consistently pushed the cost-quality Pareto frontier forward [1]. It matched the accuracy of the best fixed configuration while reducing cost by up to 89 percent and outperformed several baselines, including LLM-routing, rule-based methods, and a fine-tuned Qwen3-4B model [1]. The underlying LLMs used in such retrieval pipelines are built on the transformer architecture, which processes text by converting tokens into vectors and contextualizing them through multi-head attention mechanisms [4]. The results indicate that per-query configuration of the full retrieval pipeline is a practical alternative to static workload-level tuning [1].