Co-founder & CEO @ Qdrant. Search AI for humans, agents, and robots.
Berlin
Joined October 2008
- Tweets361
- Following310
- Followers683
- Likes373
The @qdrant_engine v1.20 release is here and brings improvements for full-text search. While we still believe that the sparse vectors are super powerful when using neural sparse models like SPLADE, sometimes it just has to be simple and fast. You asked, we did it.
From now on, Qdrant scores BM25 directly on the full-text index. => qdrant.tech/blog/qdrant-1.20…
Andre Zayarni retweeted
Which German councils are discussing wind power?
Poliscope answers in one search.
4,800 council portals, 130M points, one self-hosted Qdrant node.
Searches went from up to 20 s to 0.87 s on average with hybrid search, tenant filters, and TurboQuant.
qdrant.tech/blog/case-study-…
Reachy Mini, the open-source robot by @huggingface, can now answer: “What did you see today?” With multimodal memory inside powered by @qdrant_engine.
Details on how to do face recognition with multivectors, and how conversations outlive a 1K-token context window. ⤵
qdrant.tech/blog/reachy-robo…
Andre Zayarni retweeted
This assistant remembers where you left your keys. No network connection. No remote server.
Build one yourself in a course built in partnership with Qdrant and taught by Dylan Couzon.
👉 Enroll for free: hubs.la/Q04zGYrF0
#AI #VectorSearch #AIEngineering
Your embeddings default to flat space, not because anyone checked if your data is flat.
Hierarchical data isn't. On WordNet, a 5-dim Poincaré embedding hit 0.823 MAP. A 200-dim Euclidean one hit 0.168, with 40x more dimensions.
That's super power of Hyperbolic Embeddings:
qdrant.tech/articles/hyperbo…
We do a lot of research at @qdrant_engine: about embeddings, storage, quantization, and information retrieval in general. And we want to share our research topics and results with the community.
Eight sessions today, eight actual engineering problems:
Vector Space lifestream: luma.com/vector-space-stream
Contextual relevance: 0.92 out of 1.00.
Incorrect answers: 40%.
Repeated ingestion had increased the number of items from 8,416 to 22,946; consequently, the top five results retrieved by the agent repeatedly contained the same text segments.
Deletion of 14,530 duplicates → accuracy increased from 0.57 to 0.76. Same model, same prompt.
Full write-up, including the one failure none of these checks can see:
qdrant.tech/blog/clean-vecto…
Default hybrid search settings usually aren't tuned for your data.
We benchmarked two fusion methods, RRF and DBSF, across five datasets. Tuning just the k parameter changed the top result for 42% of WANDS queries, a +0.036 nDCG@10 gain.
qdrant.tech/articles/how-to-…
Next week is all about search—at least in Berlin! :-) Numerous events centered on search and information retrieval are taking place, ranging from the two-day Haystack conference (We still ❤️ you) to various workshops and community meetups. Naturally, @qdrant_engine will be part of the action.
On Wednesday, the 16th of September, we are organizing together with the beloved @cognee_ team a Bring Your Own Demo Night, a community meetup for Berlin's AI builders: good 🍕, cold 🍺, and people who want to see what you've built with vector search and AI memory.
▶️ Register: luma.com/berlin-sept-meetup
And for those who are not in Berlin, we have an even bigger global event on Thursday, 09/17. First ever: Vector Space Stream, a >4-hour virtual stream of the latest research topics in (vector) search.
▶️ Agenda: luma.com/vector-space-stream
See the full Search Week schedule at berlinsearchweek.com
Most vector databases give you one knob: RAM or disk.
We built Qdrant @qdrant_engine to give you five, because a collection doesn't outgrow its RAM budget all at once.
New benchmark-backed breakdown of when to use which 🧵
Caching full-precision vectors alone is a fine default. Until it isn't.
We watched one configuration go from a top performer to nearly the worst within a few million points, on the same cluster, just because the RAM-to-dataset ratio shifted.
Takeaway: stop sizing against point count.
Size against how much of your cluster's RAM the working set occupies right now. Re-check that ratio at every scaling step, not just at setup.
Full benchmarks + code: qdrant.tech/articles/memory-…
@ActianCorp launched a vector DB. Their comparison post says it's 22x faster than Qdrant.
I was curious what they'd built, so I opened the docs...
Core concepts: Collections, points, vectors, payload... Wait. I know it from somewhere. 🤔
Compatibility is a door, and doors swing both ways.
Anyone on VectorAI DB can move to Qdrant with near-zero code changes. And when they do: no 5,000 vector cap, real clustering instead of three inert fields, full and up-to-date feature set, and open-source.
Let's see how quickly they obfuscate it: docs.vectoraidb.actian.com/a…
