September 2026 Summaries

Constella is a research-preview family of retrieval models that lets users encode documents once with the 400M-parameter Stella embedding model and then switch among Zero, Nano, or full Stella query encoders without re-embedding a Qdrant collection. Zero uses token lookup and pooling for extremely low-cost queries but lacks word-order awareness, while Nano is a 34.5M-parameter transformer distilled from Stella that adds contextual understanding; across 15 BEIR datasets, Nano retained about 91% of Stella’s average retrieval score, while Zero performed lower overall but was substantially faster and sometimes outperformed Nano. On an Apple M5 Pro CPU, reported warm query encoding times were 38.95 milliseconds for Stella, 3.13 milliseconds for Nano, and 0.081 milliseconds for Zero, excluding vector search and network costs. The approach is intended for offline and low-power search, high-volume APIs, search-as-you-type interfaces, repeated agent retrieval, and hybrid Zero-plus-BM25 search, with models available through Hugging Face and preview FastEmbed support as the project undergoes internal review ahead of release.

Cross-encoders improve search reranking by jointly evaluating a query and document for relevance, but their computational cost makes efficient batching and inference important. The post describes cross-encode-rs, an open-source Rust library built on ONNX Runtime that tokenizes query-document pairs, batches documents by similar token lengths to reduce padding, and applies model-specific sigmoid or softmax scoring. Benchmarks on a MacBook M4 Max against Python libraries fastembed and sentence-transformers found only modest gains for Rust on the small MiniLM model, where all tools rely on the same underlying C++ ONNX Runtime engine, but a more substantial 1.3–1.5x latency advantage for cross-encode-rs on a larger multilingual Jina reranker. Model session creation times were nearly identical when Python startup costs were excluded, indicating that language choice has limited influence on loading performance, while Rust can still offer deployment benefits through a standalone binary, absence of Python runtime and GIL overhead, and greater control over concurrent batching.

Lucy has developed a retrieval layer for more than 2.8 million SEC EDGAR and Korean DART filings, converting complex regulatory documents into searchable text, preserved tables, XBRL facts, and source-linked chunks intended to provide verifiable evidence for financial AI systems. Using self-hosted Qdrant, the company combines metadata filtering with dense semantic and sparse BM25 search, fusing results through reciprocal rank fusion to narrow searches by company, filing type, period, section, and document identifier before ranking. Its multi-lane query process includes narrow, broad, XBRL-focused, and companion-document searches, with the final context linked back to original filing passages for analyst verification. Lucy reports that Qdrant quantization reduced memory use substantially while largely preserving retrieval quality, and that its system achieved 94.7% Recall@10 on FinanceBench; in one evaluation, replacing general web-search context with Lucy RAG context increased answer accuracy from 0.792 to 0.952 using the same answer model. The production deployment contains roughly 52.25 million dense and sparse vectors across 12 collections, and Lucy says separating collections by filing type and moving metadata into indexed fields improved both latency and recall as the corpus grew. Planned additions include expanded filing coverage, a Neo4j relationship graph for amendments and cross-referenced filings, and tools for creating evidence-based financial benchmarks.

Himalayas, a remote job marketplace with more than 100,000 listings and 300,000 public talent profiles, adopted Qdrant Cloud in 2023 to improve matching across varied job descriptions, candidate backgrounds, and practical employment constraints. After finding Algolia and Elasticsearch costly at scale and limited by keyword-based retrieval, the company sought a system that could combine semantic understanding with exact keyword matches and filters for factors such as location, time-zone overlap, work authorization, and employment type. Its production pipeline uses Qdrant for hybrid retrieval and payload filtering, while Himalayas ranks and presents results for search, personalized recommendations, and related-job suggestions. Qdrant now supports hundreds of thousands of daily requests across roughly 3 million data points, with median retrieval latency near 30 milliseconds, and recent search improvements reportedly raised recall by about 19 percentage points while reducing irrelevant results by roughly fourfold. Himalayas plans to further use AI-driven evaluation to continuously measure and improve how well job results and recommendations fit users’ queries and profiles.

SHIFT is a no-training approach for multilingual retrieval-augmented generation that reduces language bias in embedding models by estimating and subtracting a language-specific vector offset, bringing documents and queries from different languages into a shared semantic space. Testing with the 384-dimensional multilingual-e5-small model on the XRAG news collection and additional benchmarks found that the method substantially improved cross-language Recall@10 and overall retrieval when relevant answers were frequently in languages other than the query, while reducing same-language recall, particularly for German. The offsets stabilized using only hundreds to thousands of translation pairs but varied somewhat by corpus, suggesting that domain-matched parallel data is preferable to large generic datasets. The shifted vectors performed nearly identically under Qdrant’s HNSW approximate index and exact search, and a full shift strength of alpha 1.0 worked best in the reported small-model experiments. SHIFT is therefore most suitable for systems limited to a single shared index with genuinely multilingual answer distributions, whereas collections dominated by same-language questions and answers may benefit more from language-specific retrieval or translation-based methods.

Hyperbolic embeddings use negatively curved geometry to represent branching hierarchies such as product catalogs, taxonomies, and part-whole relationships more naturally than conventional Euclidean or spherical embeddings, because hyperbolic space expands exponentially with distance from its center. The discussion reports that low-dimensional Poincaré embeddings substantially outperformed higher-dimensional Euclidean embeddings when reconstructing WordNet and Google Product Taxonomy relationships, with radial position also offering an interpretable indication of specificity from broad concepts near the center to detailed descendants near the boundary. It also examines deployment challenges in Qdrant: although converting hyperbolic distance to an inner product worked in brute-force search, HNSW indexing performed poorly because vector norms varied widely. A more effective approach stores original Poincaré coordinates, uses Euclidean HNSW to prefetch candidates, and applies Qdrant Formula Query rescoring with true hyperbolic distance, reaching recall@10 of 0.920 with a prefetch of 1,000 taxonomy points. The results suggest hyperbolic embeddings can be compact and effective for genuinely hierarchical data, but require careful retrieval tuning because stronger embeddings may place more vectors near the Poincaré ball’s edge, making Euclidean candidate selection less reliable.

Qdrant Labs, in partnership with Vultr, has released Qdrant-FineWeb-10B, an open-source vector-search benchmark containing more than 10 billion dense and sparse vectors, roughly 25 TiB of vector data, and exact top-1,000 nearest-neighbor ground truth generated through more than a quadrillion distance calculations. The release aims to address limitations of smaller, synthetic, proprietary benchmarks by supporting evaluation of high-recall, high-throughput, low-latency production workloads, including filtering and deep retrieval. Qdrant also introduced PubMed-Multi-Vector for comparing dense, sparse, and ColBERT-style hybrid retrieval over a shared corpus, and Coyo-Vector-Embeddings for multimodal image-text retrieval. Accompanying the datasets is Supernova, a fully open-source distributed framework that automates embedding generation, brute-force ground-truth computation, database ingestion, and benchmark evaluation across systems including Qdrant, Milvus, and Elasticsearch. Supernova uses GPU-native processing and SkyPilot-based infrastructure orchestration to scale workloads across major cloud providers, Kubernetes, and HPC clusters through shared YAML configurations.