Kosha: search infrastructure for AI agents

Why it surfaced

This is unusually concrete open-source infrastructure: the site provides a Docker quickstart, an OpenSearch-compatible client, a design document, and 10-million-document benchmarks. It also reports the uncomfortable details—a vector-search recall of 0.947 and 34 cold-phase errors—rather than presenting performance as pure launch-day theater.

Kosha is an Apache-2.0, Rust-built search engine that combines BM25 full-text search and HNSW vector search while using S3 as its source of truth and local NVMe as a cache.