Hornet is a retrieval engine purpose-built for AI agents, enabling them to search and process large document corpora with high recall and throughput. It deploys alongside agents and data, either as a managed service or on-premises, and is model-agnostic. The company has demonstrated its technology by building and evaluating a 100-million-document search system.
Funding
Funding not disclosed
Founders
Product
Problem
Legacy retrieval infrastructure was optimized for human users, who issue short keyword queries and expect millisecond responses with top-ten snippets. AI agents, however, run long, structured queries inside reasoning loops and read entire documents, trading latency for throughput and recall. Force-fitting these agent workloads onto legacy infrastructure raises costs and leads to poor results.
Solution
Hornet is a retrieval engine built specifically for AI agents, designed to handle the scale and complexity of agent-driven search. It runs where agents run, beside their data, and can be deployed as a managed service, on-prem, or in a customer's own cloud. Any model, large or small, can use Hornet. The company has validated its approach by building a 100-million-document search engine, where it learned that iteration speed and pipeline quality are as important as index tuning. Hornet's design focuses on improving recall and throughput for long, structured queries while maintaining cost efficiency.
Target Audience
Primary customers are AI application developers and platform teams building agent-based systems that require high-recall, high-throughput retrieval over large document corpora.
Features
- Distributed retrieval cluster deployed on Kubernetes for scalable ingestion and querying.
- Supports hybrid retrieval, combining embedding vectors with keyword matching for robust results.
- ANN index tuning with configurable graph connectivity and build quality to control recall and memory usage.
- Client-side ingestion pipeline optimizations that improved feed throughput from 300 to 1,500 docs/sec.
- Query embedding instruction prefixes to correct content-length distribution bias in retrieved results.
- Full reindexing capability with checkpointing to support iterative experimentation at scale.