Vector Podcast: Beyond Hyperspace with Ohad Levi
“We wanted to bring a significant, step function, change to the market. That was our value proposition. We wanted to bring not 20–30% improvement. We wanted to bring 10x improvement.”
A few years ago, I think it was 2022, when I visited Berlin Buzzwords to give a talk about Muves (multimodal vector search) with Aarne Talman, I stopped by booth of Hyperspace. I had a chat with Ohad Levi, starting by me asking a series of basic questions about Hyperspace’s hardware + software technology, and progressing into a live demo. I was quite blown by two facts:
the idea of optimizing for a chosen parameter, like increasing throughput of search (QPS = queries per second), or decreasing overall costs of a search engine;
the complexity of queries it could handle (long boolean queries, similarity search with complex filters, nestedness etc).
Back then I was a TomTom, where latency of a geo-search engine was an accute problem. We dealt with a planet-scale realtime search of billions of objects, that were queried each time a user pressed a character. To put this into a perspective:
a global map can have billions of objects: country names, municipalities, neighbourhoods, POIs (restaurants, hotels etc), parking lots, charging stations — you name it;
if your search engine returns results within p99=500ms, it means that with a debouncing of 300ms on the client side you end up with a roundtrip latency of close to a second;
coupled with a slow hardware in a typical car, you can observe over a second latency, which is painfully slow. I did not even mention over-the-air internet speed that can be flaky.
Making an engine like this measurably faster (say, shave off a few hundred ms) means material impact on the business: more deals, more searches served, less $$$ paid for the infrastructure.
I brought this idea back to my engineering team and they got excited. We did not end up collaborating with Hyperspace, but I at the very least I wanted to learn a bit more about them.
In this episode I (finally!) sat down with Ohad Levi, its CEO and co-founder, to talk through the inception of the idea, what was (and probably still is) missing in the search engine market, theirhardware-accelerated search solution, and their learnings.
Video on YouTube:
You will find this episode in audio form as usual on virtually all the platforms or by adding an RSS: https://rss.com/podcasts/vector-podcast/2916143/
Thanks to Saurabh Rai for the epic episode design!