​Vector Space Stream is a 4+ hour free virtual stream of the latest research topics in vector search.

​Qdrant engineers, researchers, and devrels will share various topics relating to the space. You will get an open look into what we're studying and how it impacts the worlds of embeddings, similarity search, scalable AI search, and robotics.

​Who Should Join?

​This event is for researchers, engineers, and developers in the AI search space.

​Talks will be fully technical, with no marketing or sales pitches.

​Why Join?

​Get a look into cutting-edge research as well as ask questions and discuss with our researchers. You will definitely come away learning something new. Plus free merch giveaways every 30 minutes.

​Agenda:

​Tune in for the entire time or the topic that interests you most.

• ​A Look into Token Native Storage (7AM PST / 4PM CEST)

  • ​Search and database engines store text as UTF-8, forcing every agent's read to tokenize and every write to detokenize. Our research shows what you get by storing the model's own token IDs instead: 2-3x compression, 10-100x faster reads, and 2-20x faster writes when the primary user is an agent. We'll also talk about a flaw in the BPE algorithm and why the industry needs to agree on a common tokenizer.

• ​The Modern Vector Compression Stack (730AM PST / 430PM CEST)

  • ​This year, Qdrant released TurboQuant and the 'turbo4' datatype for impressive vector compression while maintaining high recall. Learn about the research behind these new features and how they can even be used to improve multivector search.

• ​Robotic Reachy and Qdrant Edge (8AM PST / 5PM CEST)

  • ​What does it take to give a robot memory that lives on the robot itself, with no cloud? This session runs Hugging Face Reachy's memory on Qdrant Edge: embedded, in-process vector search on its own hardware, so it remembers and recognizes what it sees while staying fully offline and private. An open look at moving vector search to the edge and running Qdrant embedded inside a live robot.

• ​New Geometries, New Embeddings (830AM PST / 530PM CEST)

  • ​We're often biased by Euclidean geometry. Even our embeddings can be, too. What if different geometries could create better vector representations for periodic and hierarchical data? Learn about hyperbolic and spherical approaches in this session.

• ​Zero-Compute Dual Encoders (9AM PST / 6PM CEST)

  • ​Asymmetric dual encoders challenge the assumption that queries and documents require comparable encoder capacity. This research examines how retrieval behavior changes when document representations are produced by large cloud-side models while queries are encoded by substantially smaller on-device models. We studied the limits of that asymmetry, identified where retrieval quality begins to fail, and consider what these results imply for the design of efficient edge retrieval architectures.

• ​Autofusion For Adaptive Hybrid Search (930AM PST / 630PM CEST)

  • ​Hybrid search combines sparse and dense rankings, but merging them can bury the better result. Can we predict which retrieval strategy will work best from the query alone? Jenny and Andrei built a dataset of 233K queries and trained a classifier to choose between sparse, dense, and hybrid retrieval. The hard part was building the answer key: measuring which strategy worked best for each query. At Vector Space Stream on September 17, we will walk through the dataset and classifier, then show how the router compared with fixed hybrid search on a collection it had never seen. Link to register in the comments.

• ​Supernova: Extreme Scale Vector Evaluation (10AM PST / 7PM CEST)

  • ​The field continues to push vector search to its limits. At Qdrant, we are seeing increasingly large workloads that far exceed hundreds of millions of embeddings. At the same time, embedding benchmarks have not kept up with this scale. Current benchmarks typically operate at very small scales (<1 million embeddings) or leverage small, non-industry standard dimensionalities (<512 dims). Most importantly, there exists no infrastructure to deal with this. To fill this gap, we created Supernova: a massive, highly parallelizable embedding benchmark framework that enables billion-scale embedding generation, brute-force retrieval, database loading, and benchmarking.

• ​Building Compute and Storage Separation the Hard Way (1030AM PST / 730PM CEST)

  • ​This session shares the path Qdrant has taken to separate compute and storage layers, the hard way but the right way. By transforming from a consensus-based cluster into a multi-deployment engine, you can run vector search anywhere. Join Qdrant's CTO and co-founder to hear about the restrictions encountered and how the team navigated challenges without making any sacrifices.