# qdrant Published articles for qdrant. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## Faster and Cheaper Search With Qdrant 1.20 DevFeed: [Faster and Cheaper Search With Qdrant 1.20]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-11-02T01:00:00Z Content type: article Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [Qdrant](), [quantization](), [Embeddings](), [cosine similarity search]() Tags: [2026](), [advanced](), [algorithm](), [approximate-nearest-neighbor-search](), [array](), [arxiv](), [benchmark](), [bert](), [bits](), [embeddings](), [faster](), [fasttext](), [hnsw](), [image-search](), [indexing](), [knn-algorithm](), [matching](), [metadata](), [native](), [nested](), [neural-network](), [performance](), [qdrant](), [quantization](), [recall](), [recommender-system](), [saas](), [search](), [semantic](), [semantic-search](), [simaes-networks](), [similarity](), [transformer](), [vector-search](), [vector-search-engine](), [vectors](), [word2vec]() ### AI overview Qdrant 1.20 adds `turbo8` and `turbo16` vector datatypes plus per-element indexes for nested metadata. TurboQuant stores vectors at 8 or 16 bits per dimension, using one-quarter or one-half the space of `float32`. Queries remain `float32` and are only rotated for asymmetric scoring; the article reports server-side latency within 10% of `float32` and near-`float32` recall without rescoring or disk reads. In tests across 14 embedding datasets, `turbo4` saved more space but lost a few recall points on some datasets. For nested filters, indexing each array element separately avoids reading candidate payloads: on one million arXiv papers, this served up to four times the queries per second of regular nested indexes with the same correct results. Flat filters were close in speed but missed 20-33% of the right papers because conditions could match different array elements. Set the datatype when creating a collection; for nested indexes, set `nested` to the array key on each filtered field. ### Source excerpt With Qdrant 1.20, we shipped two main features that can make your search cheaper and faster: two new quantization-native datatypes (turbo8 and turbo16) and advanced metadata indexing for nested fields. We measured the performance and cost improvements so you don't have to, and in this blog post we go through our findings. Reduce Your Disk Usage, Not Your Recall Storing vectors is always the main expense when it comes to semantic search: they cost disk when cold, and memory when warm. ## Scalable PDF Retrieval with Cohere Embed 5 and Qdrant Asymmetric Quantization DevFeed: [Scalable PDF Retrieval with Cohere Embed 5 and Qdrant Asymmetric Quantization]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-10-23T00:00:00Z Content type: article Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [information retrieval](), [Cohere API for embeddings](), [Qdrant](), [quantization](), [Embeddings](), [search-engines](), [cohere]() Tags: [approximate-nearest-neighbor-search](), [bert](), [cohere](), [embedding](), [embeddings](), [family](), [fast](), [fasttext](), [hnsw](), [image-search](), [images](), [knn-algorithm](), [languages](), [lighter](), [matching](), [model](), [models](), [neural-network](), [pdf](), [qdrant](), [quality](), [quantization](), [recommender-system](), [retrieval](), [saas](), [share](), [simaes-networks](), [similarity](), [space](), [support](), [text](), [transformer](), [v5](), [vector-search](), [vector-search-engine](), [vectors](), [with](), [word2vec]() ### AI overview On five public English ViDoRe V3 corpora, Embed 5 Pro's single dense vector for each PDF page image scored above several tested late-interaction models, according to the article's standard nDCG comparison. The authors' Qdrant recipe uses 512-dimensional, 1-bit page vectors and 8-bit query encoding to retrieve 20 candidates, then re-scores them with full vectors. It retained 99.2% of full-vector ranking quality while using 64 bytes per page for search vectors in RAM; full vectors remain on disk and candidate reads add latency and may populate the OS page cache. These figures exclude indexes and payloads, and the authors note possible training-data overlap. If OCR already works in a pipeline, Cohere Rerank 5 can rerank OCR text (it does not take images); on the same corpora, reranking the top 20 improved mean nDCG@10 from 66.8 to 73.2 with full-vector text search and from 66.2 to 71.7 with the compressed recipe. ### Source excerpt 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 - Full-Text Search Improvements DevFeed: [Qdrant 1.20 - Full-Text Search Improvements]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-10-22T01:00:00Z Content type: release Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [Qdrant](), [search-engines](), [code search](), [Embeddings](), [quantization](), [AI search](), [hybrid-search]() Tags: [approximate-nearest-neighbor-search](), [bert](), [bm25](), [code](), [code-search](), [embeddings](), [faster](), [fasttext](), [full-text-search](), [hnsw](), [image-search](), [knn-algorithm](), [matching](), [nested](), [neural-network](), [payload](), [qdrant](), [quantization](), [recommender-system](), [saas](), [search](), [simaes-networks](), [similarity](), [storage](), [transformer](), [vector](), [vector-search](), [vector-search-engine](), [vectors](), [version](), [word2vec]() ### AI overview Qdrant 1.20 adds BM25 ranking directly on full-text indexes, deriving scoring statistics from payload text and allowing query-level tuning without generating embeddings; the new `code` tokenizer splits identifiers on underscores and case changes while retaining full identifiers. Per-element payload indexes can answer nested filters without payload reads when all conditions use indexed fields on the same array; a benchmark on 100,000 points found them 7-38 times faster than regular indexes with payload reads. New `turbo16` and `turbo8` datatypes reduce vector storage: reported tests found `turbo16` matched `float32` recall@10 within 0.001, while `turbo8` approached `float16` recall at half its size. The release also adds `bits8` quantization, memory-tier controls for the ID tracker, and internal cluster authentication by default when API keys are configured. Upgrades require attention to read-only-key clusters and memory tiers for inline storage, whose vectors now follow the HNSW index tier. ### Source excerpt 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. ## Evaluating Medical RAG with Qdrant, Jev and MLflow -- A Production Ready Setup DevFeed: [Evaluating Medical RAG with Qdrant, Jev and MLflow -- A Production Ready Setup]() Original publisher: [Read original article]() Author: M K Pavan Kumar Published: 2026-10-20T13:25:31Z Content type: tutorial Language: en Sources: [DevOps.dev]() Topics: [ai observability](), [Qdrant](), [hybrid-search](), [llamaindex](), [Ollama](), [Monitoring](), [Embeddings](), [reproducibility](), [Data Quality](), [hugging face]() Tags: [2](), [3](), [analysis](), [architecture](), [automated](), [blog](), [bm25](), [build](), [but](), [code](), [combines](), [compare](), [jev](), [jevevals](), [medical](), [mlflow](), [production](), [productionrag](), [qdrant](), [rag](), [setup](), [tracedetection](), [with]() ### AI overview The article builds a medical RAG evaluation pipeline with LlamaIndex, Qdrant, Ollama and MLflow, comparing Jev decision judges with a text-generating Gemma judge on MedQuAD answers. Hybrid dense and BM25 retrieval uses reciprocal rank fusion; evaluators separately score contradictions, coverage and whether the expected context was retrieved. On this run, Jev agreed with the text judge on 17 of 19 contradiction verdicts and 16 of 18 coverage verdicts; routing uncertain Jev scores to the text judge raised agreement to 18/19 and 18/18. Jev returned scores for all 28 completed traces, while the text judge timed out on some. Of 50 questions, 22 failed before scoring, and the gold context ranked first for 22 of 28 scored questions. The authors therefore present Jev as a fast first-pass judge with text-judge review for uncertain cases, while noting that a confident Jev error can fall outside the routing band. MLflow traces and issue detection helped expose failures and latency. ### Source excerpt We often focus on retrieval accuracy and response quality, but what happens when we need to evaluate every answer against ground truth, detect contradictions, and understand where things went wrong? In this hands-on blog, we will build a complete medical RAG pipeline from scratch using LlamaIndex, Qdrant, and Ollama, then bring in MLflow and Jev to evaluate responses, compare them against ground-truth answers, inspect traces, and generate meaningful evaluation reports. 💡 The interesting part? We will put a traditional LLM judge and Jev side by side to explore how they evaluate the same RAG responses. No theory alone. Real datasets, real code, real evaluations, and measurable results. 🔬 created by author M K Pavan KumarThe Architecture Our architecture brings together LlamaIndex, Qdrant, Ollama, MLflow, and Jev to build a complete medical RAG pipeline with hybrid retrieval, automated evaluation, trace monitoring, and result analysis. The workflow is divided into four major components. 1. Data Preparation and Ingestion We start with the MedQuAD Context GroupedQA dataset from Hugging Face, which contains medical contexts and their corresponding question-answer pairs. A Python script (create_testset.py) randomly selects 50 question-answer pairs and creates a JSONL test set containing questions, ground-truth answers, and context IDs. A separate ingestion script (ingest_contexts.py) processes the medical contexts without chunking. Ollama's EmbeddingGemma generates dense embeddings, while FastEmbed BM25 generates sparse vectors. Both representations are stored in Qdrant, enabling hybrid retrieval with context-level traceability. 2. RAG Pipeline The RAG pipeline is built using LlamaIndex, which handles retrieval, response generation, and integration with Qdrant. When a question arrives, hybrid retrieval combines dense semantic search and sparse keyword matching using Reciprocal Rank Fusion (RRF). The retrieved contexts are passed to Ollama's Gemma 3, which generates the fin ## UltraBERT Embeddings: Whole Documents Versus Chunks for Retrieval DevFeed: [UltraBERT Embeddings: Whole Documents Versus Chunks for Retrieval]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-10-19T00:00:00Z Content type: article Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [Vectors & LLMs](), [Qdrant](), [vector embeddings in AI](), [tokenization](), [elasticsearch]() Tags: [1-5](), [approximate-nearest-neighbor-search](), [benchmark](), [bert](), [compression](), [embedding](), [embeddings](), [fasttext](), [hnsw](), [image-search](), [knn-algorithm](), [latency](), [matching](), [neural-network](), [qdrant](), [quantization](), [recommender-system](), [retrieval](), [saas](), [simaes-networks](), [similarity](), [transformer](), [vector-search](), [vector-search-engine](), [vectors](), [word2vec]() ### AI overview Tests of UltraBERT-Embed-150M in Qdrant found that 512-token chunks generally retrieved document details better than whole-document vectors, though whole-document vectors performed better for queries about an entire episode and for finding passkeys. Combining both representations helped only slightly on two tasks and performed worse on others. The article also reports that 1-bit TurboQuant compression reduced chunk-vector storage for 200,000 articles to about 140 MB, with quality and rescoring latency tradeoffs. ### Source excerpt 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. ## How Poliscope Made All of Germany's Local Politics Searchable on One Server with Qdrant DevFeed: [How Poliscope Made All of Germany's Local Politics Searchable on One Server with Qdrant]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-10-15T00:00:00Z Content type: article Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [Qdrant](), [self-hosted databases vs managed services](), [AI Infrastructure](), [Open Source](), [Local AI]() Tags: [approximate-nearest-neighbor-search](), [bert](), [cross-encoder](), [data-sovereignty](), [database](), [embeddings](), [fasttext](), [hnsw](), [image-search](), [knn-algorithm](), [matching](), [neural-network](), [qdrant](), [recommender-system](), [saas](), [server](), [simaes-networks](), [similarity](), [transformer](), [vector-search](), [vector-search-engine](), [vectors](), [word2vec]() ### AI overview Poliscope consolidated German municipal records from thousands of council portals into a searchable corpus using Qdrant. The case study describes its hybrid retrieval design, regional filtering, vector compression, and self-hosted single-server setup, along with reported search performance and plans for future growth. ### Source excerpt 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 Remembers: Giving a Robot a Local Semantic Memory with Qdrant Edge DevFeed: [Reachy Remembers: Giving a Robot a Local Semantic Memory with Qdrant Edge]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-10-14T00:00:00Z Content type: article Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [hybrid-search](), [Retrieval Augmented Generation (RAG)](), [Local AI](), [AI Chat](), [DuckDB]() Tags: [2026](), [approximate-nearest-neighbor-search](), [bert](), [embeddings](), [fasttext](), [hnsw](), [hugging-face](), [image-search](), [knn-algorithm](), [language](), [language-model](), [matching](), [neural-network](), [open-source](), [pi](), [qdrant](), [raspberry](), [raspberry-pi](), [reachy](), [recommender-system](), [robotics](), [run](), [saas](), [simaes-networks](), [similarity](), [source](), [transformer](), [vector-search](), [vector-search-engine](), [vectors](), [word2vec]() ### AI overview The article describes adding local semantic memory to Reachy Mini with Qdrant Edge. It explains how retrieval combines image and text vectors, uses score gates and an empty-frame baseline to avoid irrelevant matches, and stores memories on the robot. In the demo, model inference runs on a nearby laptop because the robot lacks enough free RAM, while memory search remains on-device. ### Source excerpt 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. ## I Built an End-to-End Enterprise Intelligence Simulation to Answer One Question DevFeed: [I Built an End-to-End Enterprise Intelligence Simulation to Answer One Question]() Original publisher: [Read original article]() Author: Rick Bahague Published: 2026-10-13T06:14:04Z Content type: article Language: en Sources: [HackerNoon]() Topics: [AI Chat](), [data-governance](), [Neo4j](), [Qdrant](), [DuckDB](), [AWS Glue](), [text-classification]() Tags: [agentic-ai](), [ai](), [ai-agent-architecture](), [ai-agents](), [ai-governance](), [audit-trail](), [docker](), [docker-compose](), [duckdb](), [enterprise-ai-architecture](), [fastapi](), [governance](), [governed-ai](), [knowledge-graphs](), [mlflow](), [neo4j](), [qdrant]() ### AI overview The article describes a simulated enterprise AI system that answers a question about transformer loading and customer impact risk. It combines ontology based request interpretation, certified data and model outputs, a Neo4j topology graph, Qdrant policy search, and tools that record evidence lineage and audit calls. The author reports that the agent produced a correct answer in the initial run, then added data side grouping after observing that the model struggled with grouping. The simulation also uses role based authorization, schema validation, parameterized queries, and a degraded mode when Qdrant is unavailable. ### Source excerpt Explore how I built a governed enterprise AI agent using Neo4j, Qdrant, DuckDB, and MLflow to deliver traceable, auditable answers. ## Compressing Perplexity's Visual PDF Embeddings with TurboQuant DevFeed: [Compressing Perplexity's Visual PDF Embeddings with TurboQuant]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-10-13T00:00:00Z Content type: article Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [document ai](), [Compression](), [Data visualization](), [self-hosted runner]() Tags: [approximate-nearest-neighbor-search](), [benchmark](), [bert](), [document-retrieval](), [embedding](), [embeddings](), [fasttext](), [hnsw](), [image-search](), [knn-algorithm](), [matching](), [neural-network](), [ocr](), [pdfs](), [perplexity](), [qdrant](), [recommender-system](), [report](), [saas](), [simaes-networks](), [similarity](), [transformer](), [vector-search](), [vector-search-engine](), [vectors](), [word2vec]() ### AI overview Qdrant evaluates TurboQuant compression for Perplexity's pplx-embed-v2-late models, which search PDF pages as images. At 1 bit per vector, TurboQuant reduces vector storage by 21 times while retaining 96.5% of float32 retrieval quality; rescoring raises retained quality to 99.9%. The article compares image based retrieval with OCR and Markdown approaches, and reports tests across four ViDoRe v3 datasets. It recommends testing page image retrieval when PDF layout matters, while retaining OCR where applications need extracted text or structured fields. ### Source excerpt 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. ## Jevjitsu: Or, How We Tried Generalized Classifiers on Everything DevFeed: [Jevjitsu: Or, How We Tried Generalized Classifiers on Everything]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-10-08T00:00:00Z Content type: article Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [Machine learning](), [Artificial Intelligence](), [Large Language Model](), [LLM observability](), [tgi]() Tags: [approximate-nearest-neighbor-search](), [benchmark](), [bert](), [classification](), [classifiers](), [dataset](), [embeddings](), [fasttext](), [hnsw](), [hugging-face](), [image-search](), [knn-algorithm](), [llm](), [machine-learning](), [matching](), [ml](), [model](), [neural-network](), [openrouter](), [prompting](), [qdrant](), [recommender-system](), [saas](), [simaes-networks](), [similarity](), [training](), [transformer](), [transformers](), [type-safety](), [vector-search](), [vector-search-engine](), [vectors](), [word2vec]() ### AI overview The article reports experiments using Jev, a decision model for classification, in search reranking, product result diversity, query understanding, and semantic chunking. It describes improved ranking and chunking metrics in the tested datasets, while noting trade-offs such as latency, API cost, and the risks of filtering ambiguous queries. ### Source excerpt 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: ## How Small Can Google's New EmbeddingGemma 2 Get? DevFeed: [How Small Can Google's New EmbeddingGemma 2 Get?]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-10-06T00:00:00Z Content type: article Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [Qdrant](), [information retrieval](), [gemma]() Tags: [approximate-nearest-neighbor-search](), [bert](), [embeddings](), [fasttext](), [gemma](), [gemma-4](), [google-deepmind](), [hnsw](), [image-search](), [knn-algorithm](), [latency](), [matching](), [neural-network](), [qdrant](), [quantization](), [recommender-system](), [retrieval](), [saas](), [simaes-networks](), [similarity](), [transformer](), [vector-search](), [vector-search-engine](), [vectors](), [word2vec]() ### AI overview Qdrant's tests of Google's EmbeddingGemma 2 compare vector quantization and reduced dimensions for text retrieval. Full-size 1-bit TurboQuant vectors retained 99% of baseline retrieval quality with 30 times less vector RAM; 256-dimensional vectors with rescoring used 77 times less RAM and retained 94.5%. The article recommends testing relevance and latency on an application's own queries. ### Source excerpt 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%. ## Constella Preview: Swap Query Models Without Re-Embedding DevFeed: [Constella Preview: Swap Query Models Without Re-Embedding]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-09-29T00:00:00Z Content type: article Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [sentence-transformers](), [cosine similarity search](), [hugging face](), [lora]() Tags: [approximate-nearest-neighbor-search](), [bert](), [cost](), [embedding](), [embeddings](), [evaluation](), [fasttext](), [features](), [hnsw](), [image-search](), [knn-algorithm](), [low-power](), [matching](), [neural-network](), [qdrant](), [recommender-system](), [saas](), [simaes-networks](), [similarity](), [transformer](), [vector-search](), [vector-search-engine](), [vectors](), [word2vec]() ### AI overview Constella is a family of query embedding models built around Stella, a 400M-parameter English embedding model. Stella encodes documents, while Zero, Nano, or Stella can encode queries against the same Qdrant collection, allowing query models to be switched without re-embedding stored documents. The research preview evaluates the models across 15 BEIR datasets. ### Source excerpt 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. ## How to Build a Retriever From Scratch for a Support System DevFeed: [How to Build a Retriever From Scratch for a Support System]() Original publisher: [Read original article]() Author: Pavel Yakovlev Published: 2026-09-27T13:13:49Z Content type: tutorial Language: en Sources: [HackerNoon]() Topics: [Embeddings](), [AI Chat](), [Support]() Tags: [ai-agents-for-banking](), [bm25](), [bm25-and-dense-retrieval](), [cross-encoder-reranking](), [hybrid-rag-architecture](), [qdrant](), [rag](), [rag-architecture](), [rag-customer-support-chatbot](), [rag-system-in-banking](), [rrf-retrieval-fusion]() ### AI overview The tutorial develops a customer-support retrieval system by separating document search from LLM response generation. It starts with BM25 keyword search, explains its limitations in matching meaning, and introduces sentence embeddings and vector search. It also distinguishes shared reference documents from private customer data. ### Source excerpt Build a customer-support retriever from scratch: BM25 + dense (E5) + RRF + a cross-encoder, plus an identity-scoped agent path, added step by step. ## Oxidizing Cross-Encoders DevFeed: [Oxidizing Cross-Encoders]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-09-25T00:00:00Z Content type: article Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [Python](), [onnx runtime](), [tokenizers](), [LLM optimization](), [Transformers]() Tags: [approximate-nearest-neighbor-search](), [bert](), [embeddings](), [fasttext](), [hnsw](), [image-search](), [knn-algorithm](), [matching](), [neural-network](), [onnx](), [onnx-runtime](), [python](), [qdrant](), [recommender-system](), [rust](), [saas](), [sentence-transformers](), [simaes-networks](), [similarity](), [transformer](), [vector-search](), [vector-search-engine](), [vectors](), [word2vec]() ### AI overview The article describes rewriting cross-encoder inference in Rust as the cross-encode-rs library and comparing it with Python implementations that use the same ONNX Runtime C++ engine. The Rust version was 1.1x faster at the median on a small reranking model and about 1.4x faster on a larger one, while model load time did not improve. ### Source excerpt 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 library, cross-encode-rs, how it works, where it does pull ahead (about 1.4x on a larger model), and where it does not (model load time). ## How Lucy Built a Source-of-Truth Retrieval Layer for 2.8M+ SEC and DART Filings with Qdrant DevFeed: [How Lucy Built a Source-of-Truth Retrieval Layer for 2.8M+ SEC and DART Filings with Qdrant]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-09-21T00:00:00Z Content type: article Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [Qdrant](), [managing large-scale text data in AI](), [AI Integration Strategies](), [hybrid-search](), [knowledge-engineering]() Tags: [approximate-nearest-neighbor-search](), [bert](), [data](), [data-infrastructure](), [embedding](), [embeddings](), [fasttext](), [hnsw](), [image-search](), [knn-algorithm](), [latency](), [matching](), [neural-network](), [provenance](), [qdrant](), [recommender-system](), [regulatory](), [retrieval](), [saas](), [securities-and-exchange-commission](), [simaes-networks](), [similarity](), [source-of-truth](), [structured](), [transformer](), [vector-search](), [vector-search-engine](), [vectors](), [word2vec]() ### AI overview Lucy builds a retrieval layer for more than 2.8 million SEC and DART filings, preparing them as structured data that AI systems can search. The article describes preserving table structure, standardized XBRL facts, and links back to source documents, while narrowing retrieval to rank the right chunk first. It also identifies cost and latency as concerns with frontier models. ### Source excerpt 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. ## How Himalayas Built Semantic Job Matching for 100,000+ Remote Listings with Qdrant DevFeed: [How Himalayas Built Semantic Job Matching for 100,000+ Remote Listings with Qdrant]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-09-18T00:00:00Z Content type: article Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [hybrid-search](), [Qdrant](), [voyage-ai]() Tags: [ai-job-search](), [algolia](), [approximate-nearest-neighbor-search](), [bert](), [elasticsearch](), [embedding](), [embeddings](), [fasttext](), [hnsw](), [hybrid](), [hybrid-search](), [image-search](), [job](), [knn-algorithm](), [matching](), [neural-network](), [qdrant](), [recommender-system](), [remote](), [saas](), [semantic](), [simaes-networks](), [similarity](), [transformer](), [vector-search](), [vector-search-engine](), [vectors](), [word2vec]() ### AI overview Himalayas uses Qdrant Cloud to power hybrid semantic and keyword search, structured filters, personalized job matches, and related-job recommendations across its remote job platform. The article describes the move from Algolia and Elasticsearch, a production retrieval system handling more than 25 million monthly requests, and offline search-quality improvements of about 19 percentage points in recall and roughly four times fewer irrelevant results. ### Source excerpt 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. ## SHIFTing Languages in Multilingual RAG DevFeed: [SHIFTing Languages in Multilingual RAG]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-09-16T07:00:00Z Content type: article Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [Retrieval Augmented Generation (RAG)](), [information retrieval](), [Embeddings](), [AI search](), [qwen]() Tags: [approximate-nearest-neighbor-search](), [bert](), [embeddings](), [fasttext](), [hnsw](), [image-search](), [information-retrieval](), [knn-algorithm](), [matching](), [neural-network](), [qdrant](), [qwen](), [qwen3](), [rag](), [recommender-system](), [saas](), [search](), [semantic-search](), [simaes-networks](), [similarity](), [transformer](), [vector-search](), [vector-search-engine](), [vectors](), [word2vec]() ### AI overview This article examines language bias in multilingual retrieval-augmented generation systems. It explains how multilingual embedding models can favor documents written in the query language, making relevant answers in other languages harder to retrieve, and discusses translation, per-language search, and larger embedding models as possible responses. ### Source excerpt 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. ## Qdrant Releases Large-Scale, Reproducible Datasets and Tooling for Vector Search Benchmarking DevFeed: [Qdrant Releases Large-Scale, Reproducible Datasets and Tooling for Vector Search Benchmarking]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-09-01T00:00:00Z Content type: article Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [Qdrant](), [Benchmark](), [dataset](), [Ground truth / benchmark quality](), [Embeddings](), [Developer Tools](), [reproducibility](), [GPU]() Tags: [approximate-nearest-neighbor-search](), [benchmarks](), [bert](), [distributed](), [embeddings](), [fasttext](), [gpu](), [hnsw](), [image-search](), [knn-algorithm](), [matching](), [neural-network](), [production](), [qdrant](), [recommender-system](), [reproducibility](), [research](), [saas](), [simaes-networks](), [similarity](), [tooling](), [transformer](), [vector-search](), [vector-search-engine](), [vectors](), [word2vec]() ### AI overview Qdrant describes Qdrant-FineWeb-10B, a large vector-search benchmarking dataset with exact top-1000 ground truth for 120,000 queries across a 10-billion-document corpus. The article also announces additional dense, sparse, and multimodal datasets and open-sources Supernova, a distributed framework for large-scale dataset generation and ground-truth computation. ### Source excerpt 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. ## Qdrant x LangChain: Endgame Performance DevFeed: [Qdrant x LangChain: Endgame Performance]() Original publisher: [Read original article]() Author: LangChain Accounts Published: 2026-08-26T15:02:51Z Content type: article Language: en Sources: [LangChain Blog]() Topics: [Qdrant](), [LangChain](), [LLM observability](), [Enhancing AI model performance](), [gRPC]() Tags: [api](), [async](), [langchain](), [qdrant](), [rag]() ### AI overview The article presents Qdrant and LangChain as a vector database and framework combination for production RAG applications. It explains how Qdrant stores and retrieves vectors for LLM prompts, and describes async operations, quantization, and io_uring as ways to improve resource use and performance. It also recommends evaluating vector stores against an application's needs. ### Source excerpt Qdrant and LangChain deliver production-ready RAG performance with async support, optimized resource usage, and scalable vector search for LLM apps. ## How to Tune Vector Search Without Guessing DevFeed: [How to Tune Vector Search Without Guessing]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-08-24T00:00:00Z Content type: article Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [AI search](), [Qdrant](), [data]() Tags: [approximate-nearest-neighbor-search](), [bert](), [bm25](), [embeddings](), [fasttext](), [hnsw](), [how-to](), [image-search](), [knn-algorithm](), [matching](), [neural-network](), [qdrant](), [queries](), [rank](), [recommender-system](), [retrieval](), [saas](), [search](), [simaes-networks](), [similarity](), [transformer](), [vector](), [vector-search](), [vector-search-engine](), [vectors](), [word2vec]() ### AI overview This article explains how to tune vector search by measuring collection settings and separating retrieval failures from ranking failures. It highlights issues involving sparse-vector weighting, BM25 configuration, candidate depth, fusion, and reranking, based on tests across five public datasets. ### Source excerpt 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? ## How Bayer Built an Enterprise-Scale Search Engine with Qdrant DevFeed: [How Bayer Built an Enterprise-Scale Search Engine with Qdrant]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-08-13T00:00:00Z Content type: article Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [Qdrant](), [AI Platform](), [Artificial Intelligence](), [AI search](), [Generative AI](), [Retrieval Augmented Generation (RAG)](), [Large Language Model](), [Redis]() Tags: [ai](), [ai-platform](), [approximate-nearest-neighbor-search](), [bert](), [embeddings](), [fasttext](), [generative-ai](), [hnsw](), [image-search](), [infrastructure](), [knn-algorithm](), [large-language-models](), [large-language-models-llms](), [matching](), [neural-network](), [production](), [qdrant](), [recommender-system](), [saas](), [search](), [simaes-networks](), [similarity](), [transformer](), [vector-search](), [vector-search-engine](), [vectors](), [word2vec]() ### AI overview A case study of Bayer's enterprise generative AI platform, myGenAssist, and its use of Qdrant for vector search. The article describes the platform's growth from an initial drug-discovery application to company-wide use, with retrieval designed to support large language models, handle sustained workloads, and reduce hallucinations. ### Source excerpt 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. ## Qdrant and Minima Deliver 2.92x More Agentic RAG Tasks per GPU-Hour DevFeed: [Qdrant and Minima Deliver 2.92x More Agentic RAG Tasks per GPU-Hour]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-08-13T00:00:00Z Content type: article Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [Retrieval Augmented Generation (RAG)](), [Qdrant](), [Latency](), [Inference](), [hybrid-search](), [GPU](), [NVIDIA Blackwell](), [payload]() Tags: [approximate-nearest-neighbor-search](), [bert](), [blackwell-gpu](), [embeddings](), [fasttext](), [gpu](), [hnsw](), [hybrid](), [image-search](), [inference](), [knn-algorithm](), [latency](), [matching](), [neural-network](), [payload](), [qdrant](), [rag](), [recommender-system](), [saas](), [simaes-networks](), [similarity](), [transformer](), [vector-search](), [vector-search-engine](), [vectors](), [word2vec]() ### AI overview This case study describes how Minima built a bounded retrieval agent with Qdrant hybrid search, payload filtering, late-interaction reranking, and Qwen3.6-27B generation. Across evaluated tasks and agent episodes, the stack improved throughput, latency, and grounded task success compared with dense retrieval using BF16 inference. ### Source excerpt 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. ## How to Clean Up a Qdrant Collection DevFeed: [How to Clean Up a Qdrant Collection]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-08-10T00:00:00Z Content type: tutorial Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [Qdrant](), [data]() Tags: [agent](), [approximate-nearest-neighbor-search](), [bert](), [data](), [embedding](), [embeddings](), [fasttext](), [hnsw](), [how-to](), [image-search](), [knn-algorithm](), [matching](), [neural-network](), [pipeline](), [points](), [qdrant](), [query](), [recommender-system](), [results](), [retry](), [saas](), [search](), [simaes-networks](), [similarity](), [transformer](), [vector](), [vector-search](), [vector-search-engine](), [vectors](), [word2vec]() ### AI overview A tutorial on cleaning Qdrant vector collections whose repeated ingestion creates duplicate or stale points that displace useful search results. It explains how to inspect collection contents, deduplicate records, and use stable point IDs to make retries idempotent. ### Source excerpt 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. ## Pre-Filtering vs Post-Filtering (and Why Qdrant Does Neither) DevFeed: [Pre-Filtering vs Post-Filtering (and Why Qdrant Does Neither)]() Original publisher: [Read original article]() Author: info@qdrant.tech (Andrey Vasnetsov) Published: 2026-08-07T00:00:00Z Content type: article Language: en Sources: [Qdrant Blog on Qdrant - Vector Search Engine]() Topics: [Qdrant](), [AI search](), [Benchmark](), [Graphs]() Tags: [approximate-nearest-neighbor-search](), [benchmark](), [bert](), [embeddings](), [fasttext](), [filter](), [hnsw](), [image-search](), [knn-algorithm](), [matching](), [neural-network](), [qdrant](), [recommender-system](), [saas](), [simaes-networks](), [similarity](), [transformer](), [vector](), [vector-search](), [vector-search-engine](), [vectors](), [word2vec]() ### AI overview This article compares pre-filtering and post-filtering for vector search and explains Qdrant's in-place filtering approach. It describes how metadata filters can reduce recall and how Qdrant uses filtered HNSW traversal and graph repairs to preserve search quality. ### Source excerpt 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%. [Next page]()