# Hornet: the retrieval engine built for agents

Hornet is a retrieval engine built for AI agents, made by Hornet AS. Applications and agents store structured and unstructured data in Hornet and query it with low latency, with explicit control over matching and ranking.

Agents query differently from humans: longer, structured queries issued inside reasoning loops, document-level reads, and throughput over latency. Hornet is retrieval infrastructure designed for the agent workload.

## Why agents need a different retrieval engine

Retrieval infrastructure built for humans was tuned for short keyword queries, millisecond responses, and top-ten snippets, because human users type slow
ly and read shallowly. Agent workloads break those assumptions in specific ways:

- Empty result sets, meaningless status codes, and irrelevant context tokens give an agent nothing to reason over or recover from.
- Overloading the context window with massive, low-relevance token dumps degrades reasoning and wastes compute.
- APIs written for human readers are ambiguous and inconsistent, which makes them unreliable as agent tools.

Hornet addresses these failure modes with scoped, machine-readable context and strongly typed APIs.

## What Hornet provides

* **Deep recall**: retrieves knowledge buried deep in datasets and surfaces everything relevant, not just shallow snippets.
* **Reliable tools**: strongly typed, verifiable APIs with predictable outcomes.
* **Dynamic environments**: read-optimized collections for stable, query-heavy work, and ephemeral scratchpads for planning and iteration, governed by operator permissions.
* **Model-agnostic core**: works with any model, large or small. Runs as a managed service, on-premises, or in a customer's own cloud.

## Where Hornet fits in the AI-infra landscape

```mermaid
graph TD
    subgraph "Application Layer"
        App1[Web Search]
        App2[Answering Engines]
        App3[Coding Agents]
        App4[Deep Research Agents]
        App5[Enterprise Agents]
    end

    subgraph "Agentic Loop"
        A[Loop driven by LLM] -- "Uses" --> T[Reliable Tools and Data APIs]
    end

    subgraph "Retrieval Layer"
        H[Hornet - The Retrieval Engine Built for Agents]
    end

    subgraph "Data Sources"
        D1[Unstructured Data - Text, Video, Images, PDFs]
        D3[Structured Data]
        D2[Databases]
    end

    App1 --> A
    App2 --> A
    App3 --> A
    App4 --> A
    App5 --> A

    T -- "1. API Calls Queries and Writes" --> H
    H -- "2. Returns Relevant Machine-Readable Context" --> A

    %% Ingestion Flow
    D1 -- "Data is Ingested" --> H
    D2 -- "Data is Ingested" --> H
    D3 -- "Data is Ingested" --> H
```

## Company

Hornet AS is based in Trondheim, Norway (Ferjemannsveien 10, 7042 Trondheim). Leadership: Jo Kristian Bergum (CEO), Henning Baldersheim (CTO), Erik Dyrkoren (COO). Open roles are listed at https://hornet.dev/jobs; privacy contact: privacy@hornet.dev.

## Further reading

- [llms.txt](https://hornet.dev/llms.txt): the hornet.dev agent index with raw Markdown sources and entrypoints
- [AGENTS.md](https://hornet.dev/AGENTS.md): notes on consuming hornet.dev programmatically
- [skill.md](https://hornet.dev/skill.md): Hornet capability summary; full details arrive at launch
- [Blog index](https://hornet.dev/blog/llms.txt): every engineering post with descriptions and token counts