Dope tech: TopK

As a search consultant I’m constantly trying out new systems, building demos, and getting familiar with whatever features vendors ship next. In 2026 that means playing with some genuinely cool products, and I decided to start a series highlighting the tech I reach for when I’m solving a real problem.

To make hev’s dope tech list, a product has to have some fundamental advantage over the other offerings I’ve seen. In a category as crowded as search, that says a lot. This isn’t a paid endorsement, and I call out what’s lacking too.

Alright — first dope tech review. TopK.

What they got right

Object-store-native plus NVMe. This architecture is becoming table stakes in 2026. turbopuffer drastically changed the price-performance paradigm with it, and now other search systems are following suit. Cloud providers offer amazing storage performance and reliability, and combining cheap durable object storage with hot NVMe gives you the best of both worlds. TopK is built on it from the ground up.

Clear consistency. If you’ve worked on search at scale, you’ve grown accustomed to how hard it is to reason about the state of your index. Search systems prioritize query performance over write and indexing throughput — a rational trade, but it leaves a lot of gnarly logic for the developer to deal with. Race conditions, and whole separate databases spun up to solve this one problem. TopK handles it brilliantly: every write returns an LSN — a log sequence number. Send a record in, and you know exactly when it has become available for searching. You can pass that LSN into a query and get read-after-write consistency for free. That takes a whole class of jobs-to-be-done off the search team and hands it to the product.

Postgres-compatible SQL. Every search app has a search rail — the facets and filters that describe the results. Defining those, pre-computing them, and planning around them is usually yet another job for the search team. TopK gives you Postgres-compatible SQL to express it, which makes the whole thing familiar and straightforward.

Thoughtful SDK design. The application-facing SDK is nicely considered, with functionality like .count built right in. Together with the SQL filtering, these are challenges that normally get baked into the Python of the search application and are here abstracted cleanly into the search system. Bravo, TopK.

Multi-vector late interaction. TopK will handle embedding for you, and it ships multi-vector late interaction — an elegant design on paper. I haven’t put it through real testing yet, so I’ll hold the deeper verdict for a follow-up.

They also ship an ask feature that takes on even more of the application team’s work. That one’s especially interesting to me given some of my own experiments in the same space.

Now for some criticism

Compliance is still maturing. TopK has SOC 2 Type 1, with SOC 2 Type 2 and HIPAA reports being finalized — so the story is further along than it first looks, but an enterprise buyer will want those finished. In the same vein, a few more regions beyond a couple of AWS regions would make this a more serious enterprise option.

Missing client languages. They’re short a few SDKs. I’d expect Go (which I’d personally like to see) and Java — another enterprise door they’ll need to unlock.

So you’re buying a pre-enterprise database service. But I see a path for them, which is a challenging feat to pull off in 2026.

I need more time testing and benchmarking before I’ll draw a final conclusion — but TopK earned its spot on the list.