Muves at Berlin Buzzwords 2022
This year’s Berlin Buzzwords is now behind us and all-in-all it was a great event with lots of interesting talks on search and discussions with likeminded search professionals. For those who have not heard about Berlin Buzzwords before:
Berlin Buzzwords is Germany’s most exciting conference on storing, processing, streaming and searching large amounts of digital data, with a focus on open source software projects.
In this post I will try to summarise some of the takeaways from the event.
Vector Search field is buzzing!!!
There were many interesting talks about vector search in the event. Some of the talks we would like to mention are the following:
- Jo Bergum: AI-powered Semantic Search; A story of broken promises?
- Max Irwin: The Race to the Bottom — Low Latency in the age of the Transformer
- Florian Hoenicke & Maximilian Werk: Neural Search — Let’s talk about quality
- Alessandro Benedetti: Neural Search Comes to Apache Solr: Approximate Nearest Neighbor, BERT & more
- Lester Solbakken: Hybrid search: Greater than the sum of its parts?
We were also very happy to present our own talk to the audience:
- Aarne Talman & Dmitry Kan: Muves: Multimodal and multilingual vector search with Hardware Acceleration
Introducing Muves!
One of the most exciting things for us in the conference was that we introduced our multilingual and multimodal vector search application Muves to the public for the first time. Although Muves has already been in the making for some time, the recent months have been instrumental in defining the direction of travel for Muves.
In May 2022 we wrote a blog post about our collaboration with GSI Technology Inc. The blog post Multilingual and Multimodal Vector Search with Hardware Acceleration gives an overview of our demo implementation utilising GSI’s APU for hardware accelerated similarity search. We spoke also about this project at the conference and gave a live demonstration of the capabilities of Muves and GSI APU.
Some of the key takeaways from our talk were:
- Building multilingual and multimodal search is easy — but fine-tuning is required for domain adaptation.
- With Muves, implementation of vector search application is much faster than figuring out all the details from scratch.
- It is not the case that vector search is not high-performing and does not scale easily. We demonstrated during our talk how we can process 20 parallel queries in just 220ms using the APU backend.
- In a production setting you most likely will need filtering support. APU allows neural search with symbolic filtering at scale.
Our talk
We really enjoyed presenting Muves! The audience was really engaged and asked many great and important questions.
You can watch the full talk below. The slides for our talk are available here.
What’s next?
Our journey is just beginning! We already have some interesting follow-ups lined up after the conference. If you’d like to discuss with us or arrange a demo of Muves for your organisation, you can shoot us an email at info@muves.io.
