Embedding Model Evolution

6 posts · Nov 2023 – Oct 2026

What this trail covers

Qdrant continues to enhance its support for advanced embedding models, now integrating with Google's Gemini Embedding 2. This enables native multimodal search by allowing text, images, video, audio, and PDF documents to be mapped into a single vector space within Qdrant. The integration leverages Gemini Embedding 2's flexible output dimensions (via Matryoshka Representation Learning) and Qdrant's features like named vectors and multi-stage retrieval for efficient, high-precision multimodal search. This post specifically explores the memory optimization techniques for EmbeddingGemma 2, demonstrating how quantization and vector dimension reduction can significantly reduce RAM usage while maintaining high retrieval quality.

All Qdrant trails

Oct 2026How Small Can Google's New EmbeddingGemma 2 Get? - QdrantThis post details experiments with Google's EmbeddingGemma 2 model, focusing on memory optimization techniques within Qdrant. It quantifies the memory savings achievable through vector quantization (1-bit TurboQuant) and dimension reduction (Matryoshka Representation Learning), comparing different configurations with and without rescoring. Key findings include achieving 30x less vector RAM with 99% quality retention using 1-bit TurboQuant without rescoring, and 77x less vector RAM with 94.5% quality retention using 256 dimensions with rescoring. The post provides concrete metrics on retained nDCG@10 and vector RAM usage for various configurations, offering practical guidance on balancing memory footprint and retrieval quality.
Sep 2026Constella Preview: Swap Query Models Without Re-Embedding - QdrantIntroduces Constella, a family of embedding models (Stella, Zero, Nano) designed to allow swapping query models without re-embedding documents. Stella encodes documents, while Zero, Nano, or Stella can encode queries, all searching the same Qdrant collection. Zero uses a learned bag of tokens for minimal query compute. Nano uses a smaller transformer (34.5M parameters) to match Stella's embeddings with context. Benchmarks across 15 BEIR datasets show Nano retaining ~91% of full Stella's average score, with significantly lower query latency (0.081ms for Zero, 3.13ms for Nano vs. 38.95ms for full Stella on an M5 Pro CPU). Provides guidance on using Constella with FastEmbed and Qdrant for offline search, high-volume APIs, and hybrid search with BM25.
Mar 2026Qdrant Meets Google Gemini Embedding 2 - QdrantThis post details the integration of Google's Gemini Embedding 2 with Qdrant, enabling native multimodal search. It highlights how Gemini Embedding 2 maps text, images, video, audio, and PDF documents into a unified vector space, and how Qdrant's features like named vectors and multi-stage retrieval facilitate efficient multimodal retrieval. The post also discusses the benefits of Matryoshka Representation Learning (MRL) for flexible dimensionality and cost optimization, and Qdrant's production-ready features like quantization and payload filtering for handling large, multimodal datasets.
Nov 2025How Cosmos delivered editorial-grade visual search with Qdrant - QdrantThis post details how Cosmos implemented editorial-grade visual search using Qdrant Cloud. Key technical contributions include: storing color data as CIELAB vectors for exact color search, utilizing Qdrant's named vectors to store multiple embedding types (CLIP, CNN, pHash, color) in a single collection for multimodal retrieval, and implementing an application-side fusion model for hybrid search to balance relevance, engagement, and aesthetics. The post also highlights sub-second search latency, reduced element processing time, and predictable operations through quantization, batching, and CPU budgeting.
Jul 2025Introducing Qdrant Cloud Inference - QdrantThis post announces the launch of Qdrant Cloud Inference, a new capability that unifies embedding generation and vector storage within a single API call. It eliminates the need for separate inference infrastructure, reducing complexity, latency, and network costs. The feature supports six curated models at launch, including dense, sparse (BM25), and multimodal (CLIP-style) options. It offers free token allowances for onboarding and is automatically enabled for paid accounts. The post also highlights how this simplifies building RAG, multimodal, and hybrid search applications.
Advertisement
Dec 2024Building a Facial Recognition System with Qdrant - QdrantThis post details the architecture and practical implementation of a facial recognition system using Qdrant. It covers the offline pipeline for generating and storing celebrity embeddings (using MTCNN and FaceNet) in Qdrant, and the online phase where user selfies are embedded and queried against Qdrant for similarity search. It also discusses deployment strategies (Docker, Google Cloud Run, CI/CD) and lessons learned regarding scalability, such as quantization and precomputing average embeddings. The post highlights Qdrant's role in enabling embedding-based applications and provides a practical example of its use case.
Nov 2024Advanced Retrieval with ColPali & Qdrant Vector Database - QdrantThis post introduces ColPali, a novel multimodal retrieval approach that leverages Vision Language Models (VLMs) to process document images directly. It creates multi-vector embeddings from both visual and textual content, capturing document structure and context more effectively than traditional OCR-based methods. The post details how ColPali works under the hood, including image preprocessing, contextual transformation, query processing with MaxSim similarity, and the benefits of Binary Quantization for reducing storage and computational load. It provides a practical example of using ColPali with Qdrant, demonstrating a 2x faster search time compared to Scalar Quantization while maintaining accuracy. Future directions and community engagement are also highlighted.
Jan 2024FastEmbed: Fast & Lightweight Embedding Generation - Nirant Kasliwal | Vector Space Talks - QdrantThis post introduces FastEmbed, a new Python library for fast and lightweight embedding generation, created and maintained by Nirant Kasliwal. It highlights the challenges of traditional embedding creation in production, where training-time capabilities are often wasted. FastEmbed focuses on speed, efficiency, and accuracy, offering quantized embedding models for improved CPU performance and reduced Docker image sizes. The post benchmarks FastEmbed against other solutions, demonstrating significant speed improvements. It also touches on future plans for GPU-friendly quantized models and the importance of selecting and calibrating default models based on benchmarks like MTEB.
Jan 2024Binary Quantization - Andrey Vasnetsov | Vector Space Talks - QdrantThis post details the implementation and benefits of binary quantization for vector embeddings in Qdrant. It explains how binary quantization approximates dot product/cosine similarity using Hamming distance, leading to significant speedups (over 30x) and storage reduction by converting float vectors to binary representations (0 or 1). The post highlights its effectiveness with OpenAI models and introduces oversampling as a technique to maintain search precision in real-time. It also touches upon the challenges of storing large vectors and the necessity of quantization for efficient HNSW index operation.
Nov 2023From Content Quality to Compression: The Evolution of Embedding Models at Cohere with Nils Reimers - QdrantThis post details Cohere's advancements in embedding models, focusing on content quality estimation, compression-aware training to reduce memory footprint, and the application of reinforcement learning from human feedback. It also highlights the importance of evaluating embedding quality in relative terms and previews new features like input type support for Langchain and improved compression techniques. The discussion frames these advancements within the context of using vector databases like Qdrant.