Agents need
a new retrieval engine

Legacy retrieval infrastructure was tuned for one user: humans. The stack was optimized for short keyword queries, millisecond response times, and top-ten snippets. Now agents run search, and the workload changes. Agents issue long, structured queries inside reasoning loops. They read entire documents and trade latency for throughput and recall. Force-fitting these workloads onto legacy retrieval infrastructure raises costs and leads to poor results.

Hornet is the retrieval engine for agents

It handles iterative and parallel retrieval loops with a complete toolbox of methods. Hornet's predictable, schema-first APIs cut down on errors and wasted tokens. The API covers multiple agentic workloads, from single agent reasoning loops over naturally scoped data to multi-agent systems supporting extreme query loads over web-scale datasets. Hornet provides the right serving architecture for any agent, any task, any scale.

Run Hornet where you run your agents, beside your data. In your VPC or on-premises. The same control applies to models. Hornet is model-agnostic and open source, so you can use any model and build without lock-in.

Decades of search, distilled into Hornet

We've seen legacy infrastructure buckle under the new user of search: agents. We built and scaled the retrieval technologies that power consumer products and sophisticated search engines for billions of users. We have watched our own creations strain as agents became the primary user, leading to ballooning costs and brittle performance. We know the limits of these systems because we defined them.

That is why we are building Hornet for the new user of search: agents.