Adaptive Retrieval

Adaptive retrieval studies how systems can vary retrieval strategy, depth, and timing instead of using a fixed number of documents or a single retriever for every query. Common approaches include query-complexity routers, adaptive top-k and chunk selection, confidence-based stop-or-fetch policies, and learned controllers that retrieve more when evidence is incomplete. These methods are tested in retrieval-augmented question answering, multi-hop reasoning, long-term memory, and domain tasks such as legal, biomedical, and financial QA, where both answer quality and retrieval cost matter. A central challenge is balancing accuracy, latency, and token use: confidence signals and routing policies can reduce unnecessary retrieval, but their calibration and effectiveness may vary across models, domains, and query types.