Vector Space Stream | Latest Research from Qdrant nitter.net/i/broadcasts/1PKqrNwdO…

I just published Creating coding agents with Agno github: github.com/pavanjava/coding_… medium.com/@manthapavankumar…

New course Launch: Building AI Assistants with On-Device Memory We’re excited to announce Building AI Assistants with On-Device Memory, a new course built in partnership with @DeepLearningAI and taught by @DylanCouzon, Developer Relations Lead at Qdrant. In the course, you’ll build a local memory system that can store, retrieve, filter, and forget experiences using vector search. You’ll then extend it to work with text, voice, and images, and teach the assistant to recognize new objects from just a few examples, without retraining the model. The course also covers how to run the assistant locally on Mac, Windows, and Linux, making it a practical introduction to building AI applications that can work offline and keep data on the device. It’s an intermediate course with 8 lessons, 5 code examples, and a hands-on AI coding lab. Enroll here: deeplearning.ai/short-course…

400K+ document vectors, migrated from Milvus to Qdrant. The team at Kaivid Labs recently migrated their vector database from Milvus to Qdrant and documented the entire process. They cover how they handled the migration, verified that the data arrived intact, and then compared both systems across latency, throughput, and memory usage. If you’re considering a similar migration, this is a useful real-world reference to see how they approached it and what they measured along the way. Read the full write-up: kaivid.com/blogs/migrate-mil…

Worth your time if you're interested in SPLADE and Neural Search!

You've uploaded a couple millions points to your @qdrant_engine collection: search latency spikes for a few minutes and then goes back to normal. Many people assume it's a bug, and they change vendor, but it's actually Qdrant doing you a huge favour by running optimizations for you in the background. I explored how to tune optimizers in this article: qdrant.tech/articles/tuning-… But if you want a TL;DR: - optimizers compete for resources with your queries, as both search and optimizations run on CPU - in general, optimized collections can give you up to 100x faster search than unoptimized ones (on a sufficiently large collection), so don't turn optimizers off completely - continuous indexing (even during upload) pays in the long run: it makes the upload slower, but allows for faster recovery of query latency once the upload done. Turn off indexing while uploading if you can tolerate a longer waiting time for you search latency to recover but want faster uploads. - using only one segment per collection speeds up search once everything is optimized, but takes a very long time to reach a steady state. Keep Qdrant's defaults here for optimal lantecy/optimization trade-off - serialise optimizations (allocating only one thread to optimizers) if you want smooth query times: if you can tolerate spikes and want faster recovery, leave Qdrant defaults - be careful with deletions: if you delete more points than the vacuum optimizer threshold, it will kick in and slow down your queries. Set a high vacuum threshold (if your storage has enough room) to avoid it Next time your queries have higher latency than usual, take a look at in-progress optimizations and, if you see room for improvement, make sure to optimize the optimizers!👩‍🚀🤗

I'm talking about Inference Free SPLADE models at @softwaredoug's Retrieval Augmented Gathering today. SPLADE uses a transformer neural network to expand your document's keywords, so keyword search can match synonyms (and other semantically similar keywords). But it runs a model on every query, which adds upto 50ms p50 latency on CPU and needs a GPU to be faster. Inference-free SPLADE skips that step. It does more model work upfront on your documents, so queries cost about as much as BM25. We'll talk about how it works, what you trade off, and how to take it to production :) Join today if you wanna learn more! 📺👇

How many candidates should your vector search retrieve? If you want 5 final results, retrieving 100 candidates first gives a reranker or fusion stage more to work with. We experimented with candidate depth across five public datasets and documented what we found. Increasing limit from 10 → 500 improved the best possible nDCG@10 by up to 0.28, but the current score changed by at most 0.01. Sometimes the documents are already retrieved, they just aren’t ranked highly enough. qdrant.tech/articles/candida…

And that’s a wrap of Vector Space Stream! 🎉 We spent 4+ hours going deep into vector search, from token-native storage and vector compression to robotics, new embedding geometries, dual encoders, adaptive hybrid search, billion-scale evaluation, and compute/storage separation. And we loved having you with us throughout the entire stream. Thank you to everyone who joined, asked questions, shared their thoughts, and stuck around for the whole thing. If you missed it live, the recording is now available: piped.video/live/PxGlBlqTxJI And if you want to continue the conversations, ask questions, or dive deeper into any of the sessions, join our Discord community and let’s keep the discussion going, discord.gg/qdrant We’ll definitely try to do more streams like this. Until then, thanks for spending 4+ hours with us! 🫶

When financial search needs both meaning and precision, hybrid search becomes interesting. Satyam Sahu explores this with Qdrant using dense embeddings, BM25, and SPLADE across 47 SEC filings from 10 tech companies. A practical look at where different retrieval approaches work, and where they fall short. aiadvances.org/rag-hybrid-se…

A research agent that gets better from its own run history. At n8n in the loop 2026, @itsclelia, DevRel Engineer at Qdrant, will show how to build a self-healing research agent that learns from its own run history, without fine-tuning. The setup is simple: → The agent runs research and produces an answer. → A Discord user gives it a human verdict. → That verdict is stored back in the same vector store the agent retrieves from. → On the next run, the agent retrieves both past successes and failures to understand what worked and what didn’t. The same collection becomes the agent’s run history, evaluation set, and training signal, with the human in the loop acting as the writer. Date: October 13–14, 2026 Get your ticket now: n8n.io/intheloop/

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-…

A reminder that we’ve got an amazing series of events lined up for you all next week We’ve got three community events happening across September 16-17: → Sep 16: Discord Office Hours, Virtual An open discussion with the Qdrant team. No fixed agenda, bring your questions, share what you’re building, or just drop in and listen. → Sep 16: Bring Your Own Demo Night, In person in Berlin Six community members, five minutes each, showing what they’ve built around vectors, embeddings, and agents. Pizza, beer, and networking with cognee. → Sep 17: Vector Space Stream, Virtual 4+ hours of technical talks covering token-native storage, vector compression, robotics, new embedding geometries, hybrid search, and extreme-scale evaluation. All three are free to join. Register here: qdrant.tech/events/

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.

We’ve just launched the Qdrant Beginner Course, built for people who want to understand vector search from the ground up, with 𝐧𝐨 𝐩𝐫𝐢𝐨𝐫 𝐞𝐱𝐩𝐞𝐫𝐢𝐞𝐧𝐜𝐞 𝐫𝐞𝐪𝐮𝐢𝐫𝐞𝐝. The course starts with the basics: why traditional search struggles, how embeddings represent meaning, and how similarity works. From there, you’ll build your first Qdrant collection, learn about vectors, payloads and HNSW, and understand the differences between dense, sparse, and hybrid search. You’ll also learn how to design a complete vector search system, including filtering, RAG, and deployment considerations, before putting everything together in a hands-on multimodal supplier risk intelligence capstone project. The core course takes under 2 hours, with the capstone taking around 3 more hours. It’s self-paced, hands-on, and you can earn a 𝐬𝐡𝐚𝐫𝐞𝐚𝐛𝐥𝐞 𝐜𝐞𝐫𝐭𝐢𝐟𝐢𝐜𝐚𝐭𝐞 when you complete it. Start here: qdrant.tech/course/beginners…

Big news: @qdrant_engine has a new beginners course! If you're new to vector search and don't have much extra time to spare for learning, this ~3hrs course is for you: qdrant.tech/course/beginners… It is mostly focused on introductory concepts, so if you're looking for something more advanced and you have a little more time, you should also check out the Qdrant Essentials course :) Huge props to @EASzyszka, @themanasai, @DylanCouzon and @kanungle for driving this! Let's explore vector space together👩‍🚀🚀

@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. 🤔

Your search can be working perfectly and still be badly tuned. Vector search has a lot of tuning knobs: `hnsw_ef`, candidate depth, RRF `k`, quantization, reranking, and more. The problem is, when search quality isn't where you want it to be, which knob should you change first? We tested this across five public datasets and found that the answer depends on where the problem actually is. For example, increasing candidate depth from 10 → 500 improved the best achievable score by up to 0.28. But the final score improved by at most 0.01 because the relevant documents were already being retrieved, they just weren't ranking high enough. So instead of changing settings until the score improves, first identify what's failing, then tune the knob connected to that problem. That's what we explore in our latest article: how to systematically tune retrieval without guessing. qdrant.tech/blog/tuning-retr…