The retrieval stack is becoming agentic

The failure was never the retriever

In a preprint, the authors diagnose multi-hop RAG failures and land on an uncomfortable finding: a significant portion of missing-evidence failures are caused by search drift, not retriever limitations. Agents stop searching before finding what they need, or drift off the query trail. The industry has spent years upgrading embeddings while the actual defect sat in the agent's query generation. m:…3331

Worse, recovering the missing evidence does not produce a proportional reduction in total system error, because downstream reasoning issues like premature resolution persist even when the evidence arrives. The authors argue the primary bottleneck is shifting from search guidance to verifying that retrieved evidence is sufficient. The RAG roadmap flips: less retriever tuning, more requirement tracking. m:…3331

Think less, retrieve better

A second preprint, introducing AURA, attacks the cost of reasoning-conditioned retrieval. Instead of running full reasoning trajectories for every query, AURA supervises intermediate reasoning states based on how well they separate relevant documents from hard distractors, and stops adaptively when a lightweight utility estimator says further thinking is not worth it. Adaptive stopping cuts inference latency from 4.6 seconds to 0.76 seconds relative to full-trajectory reasoning. m:…1335

Quality holds: AURA-4B achieves 34.7 average nDCG@10 on the BRIGHT benchmark. For operators, the message is that reasoning depth is a budget to be allocated, not a dial to be maxed. Production RAG gets cheaper and faster without getting dumber. m:…1335

The design detail worth copying is where the supervision lands: on intermediate reasoning states, not final outcomes. Judging a retriever only by its final answer hides which steps separated signal from distractors. Instrumenting the middle of the trajectory is what makes the stopping decision trustworthy, and it is the difference between a heuristic timeout and a learned budget. m:…1335

Build the library before you query

The third preprint inverts the retrieval order entirely. Librarian Agents shift document retrieval from reactive to anticipatory: a builder agent, trained by reinforcement learning with workspace construction formulated as a Markov decision process, proactively curates unstructured corpora into hierarchical, task-specific workspaces before any question arrives. The curated workspaces outperform standard retrieval and precomputation baselines by 33 percentage points on BrowseComp-Plus. m:…4787

The builders also generalize: trained one way, they improve accuracy by 31 points on average on unseen BRIGHT domains, zero-shot. The implication, and this is reasoning, is that the expensive part of enterprise RAG moves offline. Curate once, answer cheaply forever, instead of re-reasoning over raw documents at query time. m:…4787

Throw away the API

The fourth preprint goes after the tool layer itself. Instead of giving agents a rigid, predefined graph API, the Direct Knowledge Graph Interaction framework lets agents write Python code directly against node and edge files, with multi-turn programming and execution feedback letting them adapt search strategies on the fly. The result exceeds the strongest published baseline by at least 6.7 Hit@1 points on every graph tested, outperforming shell-only or predefined-tool approaches. m:…3087

The authors' diagnosis is pointed: predefined graph interfaces constrain agent access to evidence even when agents use code to combine tool calls. Put the four preprints together, and this is reasoning, and the stack inverts: retrieval stops being a service the agent calls and becomes a skill the agent exercises. Teams still buying retrieval as a black box are buying the layer their competitors are about to replace with an agent. m:…3087