hornet.dev

What It Is

Hornet is a retrieval engine built for Agentic Retrieval, with the goal of replacing traditional retrieval infrastructure optimized for “human search.” Its core assumption is that AI Agent queries are now different from short keyword searches: Agents issue longer, more structured queries during reasoning loops, tend to read full documents, and make new trade-offs among latency, throughput, and recall.

Core Capabilities

Based on the available content, Hornet focuses on supporting iterative and parallel retrieval loops, covering scenarios from single-Agent reasoning to high-concurrency retrieval across multiple Agents. Its schema-first API is positioned as a way to reduce call errors and token waste. The product also claims to support data ranging from natural-scale datasets to web-scale datasets, and to provide service architectures tailored to different Agents, tasks, and scales.

Models, API, and Data Deployment

Hornet is explicitly a model-agnostic retrieval layer, not an LLM itself. Users can pair it with any model, reducing the risk of model lock-in. For deployment, it supports running in a VPC or on-premises environment, close to both the Agent and the data, which is appealing for enterprise private data, compliance, and cost control. However, the page does not disclose specific SDKs, framework integrations, access control, auditing, encryption, or operational capabilities.

Pricing and Trial

The crawled text does not provide pricing, a free quota, trial options, or enterprise support plans. Although open source is mentioned, the license, hosted-version pricing model, and boundaries of commercial services are not specified. As a result, its cost-effectiveness can only be judged preliminarily based on its positioning, and it cannot yet be compared precisely with hosted vector databases or search cloud services.

Pros, Cons, and Best Fit

Hornet’s strength is its clear positioning: it is designed specifically for long Agent queries, structured queries, full-document reading, and parallel and iterative retrieval, while supporting private deployment and freedom of model choice. Its limitation is that the public materials are more vision-oriented, with a lack of performance benchmarks, recall metrics, latency data, cost-saving evidence, and real-world case studies. It is better suited to teams with retrieval engineering expertise that are building complex RAG or multi-agent systems, rather than non-technical users looking for an out-of-the-box solution.

Access from China and Alternatives

Access from mainland China cannot be determined from the available text, and payment methods are not disclosed. If access or compliance becomes a constraint, alternatives worth considering include Milvus, Qdrant, Weaviate, OpenSearch, Elasticsearch, and Vespa. Among them, Milvus is usually easier to evaluate in terms of the domestic ecosystem and private deployment in China.