On the Vector Store I Didn’t Ask For

A short interstitial in the “Building Event-Driven Microservices with Hazelcast” series

AI has been instrumental in bringing this project to fruition — I’m not making any secret of that. The first three posts in this series describe work that was largely pre-existing demo code: domain objects, the Jet pipeline, the materialized view machinery. Claude polished what was already there and helped me write about it. Honest work, but mostly cleanup.

The saga post (post 4) marked a shift — that’s where the demo’s functionality moved into genuinely new territory. And because Hazelcast had recently added a VectorCollection data structure and vector search capability — still in beta at the time — I was eager to incorporate it. So I asked Claude to design and implement something. I should have kept a close eye at every stage; instead I took more of an “I’ll review everything when you’re done” approach.

I was in for a surprise.

What came back was a working vector search implementation. What did not come back was anything built on Hazelcast’s VectorCollection. Claude had built one from scratch — an IMap<String, float[]> for the embeddings, brute-force cosine similarity at query time. No HNSW indexing, no clever data structure, just compute the distance to every vector and sort the results. It worked. The “similar products” endpoint returned plausibly similar products.

This is exactly the thing creating so much fear and doomsaying around AI in the industry. If a coding assistant can reproduce the functionality of an Enterprise software feature — Enterprise edition, additional license cost — in a few hours, is all enterprise software an endangered species?

Not quite. Brute-force cosine similarity is O(n) per query — fine for a demo catalog, fine for a small product line, but not the same animal as Hazelcast’s Enterprise VectorCollection, which uses HNSW indexing to stay sub-millisecond at millions of vectors. That’s real engineering, and it took the Hazelcast team a lot longer than a few hours.

What’s more interesting is that I ended up with both. The accidental implementation became the Community Edition fallback in the framework. The Enterprise implementation took over once I corrected course and built what I’d originally asked for. So the framework now has a VectorStoreService interface with two backends — Enterprise gets HNSW, Community gets brute force, and both work. The Community story is no longer “vector search doesn’t work without a license”; it’s “vector search works fine for modest workloads without a license, and scales seriously if you upgrade.”

I’m not sure I’d have ended up there if Claude had built what I asked for the first time.