NVIDIA: Agentic Retrieval Boosts nDCG@10 by 8.7 Points but Is 160x Slower
nvidia · hf · 2026-10-07
An NVIDIA study evaluates agentic retrieval—combining LLM reasoning with retrievers in a ReAct loop—against standard dense retrieval for complex tasks.
- Effectiveness: with the same embedding model, nDCG@10 improves by 8.7 points on average; the same pipeline is competitive on both ViDoRe v3 and BRIGHT leaderboards, while specialized retrievers struggle out-of-domain.
- Cost: on average 107.4 seconds per query versus 0.67 seconds for standard retrieval, consuming 764.1K input and 5.8K output tokens per query.
- The takeaway: agentic retrieval clearly wins on effectiveness and generalization, but the cost gap motivates cheaper retrieval agents for large-scale deployment.
Related event: NVIDIA Benchmarks Agentic Retrieval: +8.7 nDCG@10 at 160x the Cost(2 posts)→
More from coding & agent
- 87% of exploited bugs attacked by disclosure day, says Kevin Mandia pitching agent-based defense — a16z · 2026-10-07
- 2 days left: how to claim your $250 Claude Max GitHub credit before it expires — doodlestein · 2026-10-07
- Codex-built Blink 2: an interactive flight sim with a real software depth buffer — MikePFrank · 2026-10-07
- FIXME Turns a Circled Phone Bug into a Full MCP Context for Coding Agents — ZotZot69 · 2026-10-07
- An MCP limerick: great plan, then every tool call asked me to auth again — tdhopper · 2026-10-07
- Solo founder's AI beats Minecraft, outperforming DeepMind and OpenAI attempts 26x — julianweisser · 2026-10-07