Hornet
What They Build
Hornet (hornet.dev) is a retrieval infrastructure company focused on high-precision context delivery for agentic AI systems. Their core thesis: the failure mode of agentic retrieval is not missing relevant documents but injecting convincing distractors — text that is semantically adjacent, high-confidence, and wrong.
Their work challenges the assumption that better retrieval always means better end-to-end accuracy. In agentic settings, a retriever that returns more plausible but incorrect documents actively degrades LLM reasoning.
Focus Areas
- Defensive retrieval — precision over recall; dynamic-k over fixed-k; abstention as a first-class retrieval outcome
- Distractor-aware evaluation — UDCG metric that assigns negative utility to passages that cause incorrect answers
- MAD (Mutually Assured Distraction) — framework describing how independently rational improvements to retrieval and reasoning can lead to global system instability
- Agentic query workload — characterizing how agents search differently from humans (long queries, web operators, multi-turn sessions)
- Agentic context quality — treating each retrieved chunk as a liability rather than a free addition
People
- Jo Kristian Bergum — Co-founder; former Vespa AI Chief Scientist; agentic retrieval workload research
- Lester Solbakken — Co-founder; author of the MAD framework and defensive retrieval principles
Articles
- Mutually Assured Distraction — why better retrieval and better reasoning can make agentic systems less reliable
- This Is What Agentic Retrieval Looks Like — GPT-5 query behavior: 24 calls/session, 10-term median, phrase quotes in 98% of sessions
Concepts
Agentic Search · Agentic Retrieval · Clean Context · UDCG · Retrieval Pipeline · LLM as Judge · Search Evaluation · BM25 · Query Operators