Qdrant Blog

Cohere released Embed 5, a family of embedding models. The family has a quality-focused model, Embed 5 Pro (embed-v5.0-pro), and a lighter one, Embed 5 Fast (embed-v5.0-fast). Both accept text and images, support 100+ languages, and share an embedding space.

Qdrant 1.20.0 is out! Let’s look at the main features for this version: BM25 on Full-Text Indexes and Code Search: Rank by BM25 straight from your payload text, with no embeddings to generate. Plus, better code search with a new code tokenizer. Faster Nested Filters: A new per-element payload index answers nested filters without reading payloads, making them faster. 16-Bit and 8-Bit TurboQuant: A 16-bit TurboQuant datatype that halves vector storage without accuracy loss, and an 8-bit TurboQuant datatype and quantization option with recall close to float16.

Every retrieval pipeline starts with the same call: how big is a chunk? Too small, and a chunk loses its context. Too big, and the embedding model cuts it off. DMIS Lab’s UltraBERT-Embed-150M moves that cutoff to 32K tokens. It’s a 151M-parameter dense embedding model built on UltraBERT, an English encoder trained for 32K-token inputs. Most documents fit in one pass, and you get back one 768-dimension vector. So is chunking dead? We tested it in Qdrant: one vector per document against 512-token chunks, the two together, and both compressed.

Poliscope makes Germany’s local political data searchable in one place. Germany has about 11,000 municipalities, and almost every one of them publishes through a council information system (in German, a Ratsinformationssystem). That is where agendas, minutes, proposals, and planning documents get published. Many portals serve several municipalities at once, so Poliscope’s roughly 4,800 active portals cover about 90% of German municipalities and 98% of the population. They run on roughly a dozen software products, each with its own quirks, and none of them talk to each other.

Reachy Mini is an open-source robot from Pollen Robotics and Hugging Face, run by a Raspberry Pi with 4 GB of RAM. We wanted it to do what a good host does: greet you by name the second time you walk in, remember what you talked about, and tell you what it saw while you were gone. A robot that talks through a language model can’t do any of that on its own. The model only knows what’s in its context: restart the robot and it knows nobody, talk long enough and the start of the conversation falls out, and nothing in it can be searched.

A number in a financial report needs its row, column, and unit to mean anything. A chart needs its labels. When you build search over PDFs, that context is part of what you want to retrieve. Perplexity released pplx-embed-v2-late, two late interaction models that search document pages as images. The model embeds the page image, including its tables, charts, and layout. You can skip the optical character recognition (OCR) step and search the page as it appears.

Classification is one of the oldest problems in machine learning. Unlike modern decoding transformers, most classification models were non-autoregressive by nature. Before Jev, the common approaches were training your own classifier, which needs labels, using a zero-shot classifier (compared in this Hugging Face benchmark), or prompting an LLM with some hacky type safety on top. There were also meta-ML methods, but none of them reached the mainstream. With the introduction of Jev, a decision model from TypeSafe served through OpenRouter, the taxonomy of classifiers has changed yet again:

Google’s new EmbeddingGemma 2 can run on a phone. The model travels light. The vectors need their own suitcase. At 10 million documents, full-size float32 vectors alone take 30.7 GB of RAM. We got early access ahead of today’s Google DeepMind release and tested how much of that memory we could save in Qdrant. Quantized full-size vectors kept 99% of their retrieval quality with 30x less vector RAM. Shorter vectors with rescoring reached 77x less vector RAM and retained 94.5%.

As query traffic grows, so does the compute bill for embedding it. On a low-power device, a large model may not fit in memory. A smaller query model could reduce that cost, but switching usually means re-embedding the collection. Constella lets you make that switch. It’s a family of models on Hugging Face built around Stella, a 400M-parameter English embedding model. Stella encodes your documents. Zero, Nano, or Stella itself can then encode your queries, all searching the same Qdrant collection.

We rewrote cross-encoder inference in Rust, benchmarked it against Python, and on a small reranking model it came out only 1.1x faster at the median. That is not a sign of a slow Rust implementation. The Python libraries we compared against, fastembed and sentence-transformers, hand the same work to the same C++ engine, onnxruntime, so all three spend almost all of their time in the same compiled code. In the Rust community, rewriting code in Rust is called “oxidizing” it, since rust is also what iron turns into when it oxidizes. This post covers how we oxidized cross-encoder inference into a

Lucy builds data infrastructure for financial AI. The company turns large-scale regulatory filings from the U.S. Securities and Exchange Commission (SEC) EDGAR system and Korea’s DART disclosure system into structured data that AI systems can search, retrieve, and reason over.

Himalayas is a remote job board and AI job search platform. It brings together more than 100,000 live remote job listings and over 300,000 public talent profiles, and it works both sides of the market. Job seekers explore roles by skills, salary, location, and time zone, research remote companies, and receive personalized job matches. Employers advertise remote roles and discover candidates. Around that core, Himalayas offers tools to tailor resumes and cover letters, practice interviews, and track applications.

If you speak more than one language, you know the feeling when the mental switch in your head starts up with the rattling sound of a struggling engine, mixing every word you have to produce into some Denglish, Frenglish, or Spanglish. Work-related thoughts come back from your inner search engine of a brain in English, life wisdom – in the mother tongue, and the mix is unpredictable, a little weird, but it works.

Real world vector search workloads are increasingly large and complex. Enterprises are not using vector search to occasionally search through a couple of PDF files. They are indexing and searching billions of vectors at thousands of requests per second (RPS) and sub 50ms tail latency. Large enterprises also can’t tolerate faulty assumptions. Too many benchmarks use gated, proprietary managed services. And even worse, the data is synthetic, the queries are hidden, and the engines are locked behind paywalls.

Your collection works. Queries return in a few milliseconds, results are mostly right, and product keeps forwarding you the ones that aren’t. You open the search API reference and get exact definitions for hnsw_ef, reciprocal rank fusion k, and quantization oversampling. The definitions are correct. They still don’t tell you which setting is failing on your data. So you change one setting, rerun the queries, and the score moves by 0.01. Did relevance improve, or did the same queries land differently?

Reducing Retrieval and Calls When a retrieval-augmented generation (RAG) agent runs, it often has to plan a search, check the evidence it gets back, and try again when that evidence falls short. Those inefficiencies compound. Every extra retrieval and every extra model call adds latency, context, and inference cost.

Bayer is a global life sciences company operating at the intersection of two of the most consequential fields in human life: health and nutrition. Its pharmaceutical work supports drug discovery and patient care, while its crop science work supports food production at planetary scale. The company’s guiding ambition, “Health for all, hunger for none,” frames how it thinks about technology: AI is not a side project, but a lever applied across the entire organization, from improving the productivity of colleagues to accelerating yield prediction and drug discovery.

Every crawl, retried job, and embedding pipeline change writes points into a vector collection. The stored data keeps moving even when the query code never changes, and the top results move with it. At first, little looks wrong. Search returns results, and latency stays normal. In our baseline run, the context-relevance score sat at 0.92 out of 1.00 while four answers in ten came back wrong, because duplicate chunks and one outdated record were filling the five results the agent could read.

Adding a metadata filter to vector search can make good results disappear without making the query look broken. It still runs fast, returns something, and keeps the dashboards quiet, while some of the true nearest matches drop out. In the benchmark behind this post, a broad-value filter lowers recall to 90.8% and an AND filter over two broad values lowers it to 39.7%, while every other filter shape stays above 97%.

Qdrant 1.19.0 is out! Let’s look at the main features for this version: TurboQuant Datatype: A new storage format that compresses vectors to four bits without keeping their original full-precision representation, reducing storage by up to nine times compared to TurboQuant quantization. Memory Tiers: A single memory parameter unifies per-component memory tier placement, with three tiers: pinned, cached, and cold. Per-Tenant IDF Statistics: Narrow the IDF corpus to a specific tenant so term rarity reflects that tenant’s vocabulary rather than the whole dataset, improving BM25 scoring in multi-te