How small can a serious semantic search engine be?

Hypothesis

Semantic search is treated as a cloud service you rent. For a company's own corpus — wiki, memory, documents — a tiny self-hosted vector engine on the server you already pay for should be fast enough, private by default, and free at the margin.

What we built

A lightweight vector-search service running on our own server, indexing the workspace corpus and exposed as a simple search API on the private network. It backs the wiki's hybrid retrieval (graph navigation first, vector search as the second hop) and is available to every agent as a skill.

Learnings

• Hybrid beats pure-vector for a structured corpus: following the wiki's own links answers most queries; embeddings catch the phrasing mismatches.
• Self-hosting removes the meter, and removing the meter changes behaviour — agents search liberally when nobody counts queries.
• The private-network-only rule (no public port, reachable over the tailnet) is the entire security model, and it's enough.

Log

• 2026-07-04 — Wired into the wiki's hybrid retrieval as the second hop.
• 2026-06-14 — Service live on the server, workspace corpus indexed.