Xerj

XERJ is a search engine designed for AI agents, shipped as a single Rust binary under the Apache 2.0 licence. It indexes code, documents and data, and answers an agent's questions from the index rather than from its context window. The vendor says it replaces separate search, vector database, embedding and agent memory components. It speaks the Elasticsearch 8.13 wire protocol, so existing Elasticsearch clients can connect unchanged, and supports around 50 query types. Retrieval combines BM25 and kNN (HNSW or exact) with reciprocal rank fusion in one query, with SQ8 quantization, columnar log storage and session scoped, time decayed agent memory. A feature the site emphasizes is reference coding: a developer clones open source repositories, indexes them once with the autoindex command, and the agent then looks up exact definitions with search and def commands before writing code. The vendor reports that agents using this loop used 2.7 times fewer output tokens than a grep driven agent in its own benchmark, and publishes the runs on a benchmarks page. It installs with a one line script on Linux, macOS and Windows, on x86_64 and arm64, and the site publishes an llms.txt file with agent instructions. There are three tiers. Community is free and single node, with community support and no SLA. Professional is a three node high availability cluster with a 99.95 percent availability SLA, priced per node on request. Enterprise adds air gapped deployment, 24 hour support, a 99.99 percent SLA and compliance options, with custom pricing.