Qdrant

What if you could change your query embedding model without re-embedding all your documents? That’s the idea behind Constella, a research preview from Qdrant.

Recently, we released an open-source embedding generation tool, Supernova, with a 10B dataset (Qdrant-FineWeb-10B) for free on @huggingface. These resources are excellent for evaluating, testing, and configuring your vector search infra. ICYMI - @BetterStackHQ just posted a high-quality video explainer for these resources: http://youtube.com/watch?si=HA1qmbr3-uJHTv7u&v=6Uj6GXIpX9s&feature=youtu.be

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: https://kaivid.com/blogs/migrate-milvus-to-qdrant-vectordb/

77× less vector RAM. 94.5% of baseline retrieval quality. We got early access to Google EmbeddingGemma 2 and tested how far its Matryoshka-compatible embeddings could be compressed in Qdrant. With shorter vectors + quantization + rescoring, we reduced the vector memory footprint by ~77× while retaining 94.5% of baseline retrieval quality. Google also featured our work in their EmbeddingGemma 2 announcement. Qdrant deep dive: https://qdrant.tech/blog/embeddinggemma-2/ Google’s announcement: https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/

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

Lucy built a source-of-truth retrieval layer over 2.8M+ SEC and Korea DART filings. Four hybrid lanes on Qdrant, payload filters applied inside the query, not after it. 94.7% Recall@10 on FinanceBench. +0.160 answer accuracy from retrieval alone, with the answer model held fixed. An open-weight model on Lucy RAG beat frontier models searching the web. https://qdrant.tech/blog/case-study-lucy/

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: https://www.deeplearning.ai/short-courses/building-ai-assistants-with-on-device-memory/

What happens when engineers are given a controversial topic and told to debate it? We’re bringing something a little different to San Francisco Tech Week. Join Qdrant × @neo4j for Hard Negatives: Engineers Debate Night on October 9. We’ll put engineers, AI researchers, and builders into 1v1 debates on spicy topics like: “Open weights should be banned.” The audience votes live on who wins each round. Plus, your RSVP includes food and one drink. October 9 | 6:30–9:00 PM PDT Spots are limited and registration requires approval. Link: https://luma.com/sf-meetup-oct26

780 ms during background optimization. 4.3 ms after it finished. Same Qdrant collection, very different search latency. We tested 13 optimizer configurations on 1.76M vectors to understand the trade-offs between indexing, recovery time, and query performance. Here’s what we found 👇

21× smaller vectors. 99.9% of baseline retrieval quality retained with rescoring. We tested Perplexity’s pplx-embed-v2-late with Qdrant’s TurboQuant across 4 ViDoRe v3 datasets. 856 KB → 40 KB per page, with a trade-off: rescoring requires the original vectors. ⁠Full benchmarks below 👇

How is AI changing software development? Qdrant's @EASzyszka joined @WeAreDevs at Agent Conf Berlin to discuss token usage, vector search, data quality, and the evolving developer workflow. Watch the conversation: https://www.youtube.com/watch?v=CJa6YRq-tjU

Can one dense vector per PDF page compete with late interaction? @cohere's Embed 5 Pro turns a whole page image into a single 2048-dimensional vector, so we tested it against models like ColPali.