March 2026 Summaries

Qdrant Skills for AI Agents introduces an innovative approach to enhancing the capabilities of AI agents in managing production vector search by providing them with encoded solutions architect knowledge. Unlike traditional documentation, which focuses on feature-based instructions, Qdrant Skills offers a problem-oriented, situational guidance system that helps agents navigate engineering decisions like memory optimization, search strategy, and scaling. This system enables AI agents to make informed decisions beyond API calls, addressing complex issues such as memory usage, search quality, and scaling through a decision tree format that links to relevant documentation. The skills are designed to bridge the gap between basic usage and expert-level production management, providing AI agents with the ability to diagnose and solve problems effectively. By integrating these skills with the qcloud-cli tool, agents can apply their diagnostic insights directly to Qdrant Cloud management, enhancing both the knowledge and operational layers for vector search solutions.

Qdrant's Multi-Vector Search Course, created by Kacper Łukawski, addresses the need for a structured, hands-on resource for implementing advanced multi-vector retrieval systems in production environments. This free and advanced course is specifically designed for machine learning, backend, and search engineers who are already familiar with vector search fundamentals and seek to master multi-vector search techniques. The course is divided into four comprehensive modules, covering setup, text multi-vectors, multi-modal search, and optimization and evaluation, each with video lessons and practical exercises using Google Colab notebooks. Participants who complete the course and pass the certification exam can earn a Qdrant Multi-Vector Search certificate, demonstrating their ability to design and optimize multi-vector retrieval pipelines. Additionally, the first 20 individuals to achieve certification and share their accomplishment on LinkedIn with the hashtag #QdrantCertified will receive free Qdrant swag. The course is self-paced and encourages community interaction via a dedicated Discord channel.

The advice to "start with pgvector, graduate later" for vector databases is prevalent, but an analysis of over 110 community threads reveals that this approach is more nuanced and often limited by specific conditions. While pgvector offers seamless integration with Postgres, allowing for vector search without additional infrastructure, its effectiveness is contingent upon six conditions: managing a vector dataset under 1M vectors, not requiring accurate metadata filtering, having embeddings tightly coupled to relational data, not needing hybrid search, relying heavily on Postgres for business logic, and having a small team with manageable search logic in SQL. Most applications quickly outgrow pgvector in terms of features rather than scale, necessitating a shift to dedicated vector stores like Qdrant, which offer advantages such as efficient metadata filtering, native hybrid search, and scalability beyond 10M vectors. Despite the challenges of synchronizing data between dedicated vector stores and Postgres, these issues are manageable with well-established patterns, and opting for dedicated solutions can alleviate the limitations of pgvector when not all conditions are met.

The blog post discusses a novel approach to video anomaly detection, utilizing Qdrant for edge-to-cloud integration. Unlike traditional classifiers requiring specific anomaly labels, this system reframes detection as a nearest-neighbor search problem, calculating vector distances to identify deviations from normal activity. Leveraging Qdrant's capabilities, video embeddings are indexed as a baseline, allowing new clips to be quickly evaluated for anomalies without retraining. The deployment involves NVIDIA Jetson devices for local processing, with cloud infrastructure provided by Vultr for comprehensive analysis and semantic search. This architecture, which includes components like Twelve Labs Marengo 3.0 for video embeddings and NVIDIA Metropolis VSS for GPU-accelerated processing, efficiently balances local and cloud resources. It significantly reduces cloud costs and ensures scalability across large camera networks by escalating only significant anomalies for detailed cloud analysis. The system's utility extends beyond surveillance to various industries needing real-time detection of unusual activities without exhaustive configuration of potential anomalies.

Qdrant has announced a $50 million Series B funding round, led by AVP and including investors like Bosch Ventures and Spark Capital, to advance its development of composable vector search infrastructure, which is crucial for AI systems. Unlike traditional fixed-pipeline search systems, Qdrant's architecture, built in Rust, offers customizable retrieval methods that adapt to various workloads, enhancing performance and efficiency from edge devices to supercomputers. This flexibility is particularly beneficial for agentic AI workflows that require fast and adaptable retrieval strategies without latency issues. Qdrant's approach has already been adopted by enterprises such as Canva and OpenTable, and the company has achieved over 250 million downloads and 29,000 GitHub stars, reflecting its widespread use and community engagement. The company emphasizes the importance of retrieval as core infrastructure for AI, and its community-driven development is continually improving the system under real-world operational pressures.

Google's Gemini Embedding 2 is a groundbreaking fully multimodal embedding model that integrates text, images, video, audio, and PDF documents into a unified vector space, enhancing semantic information preservation by processing each modality directly. Available through the Gemini API and supporting over 100 languages, it features flexible output dimensions enabled by Matryoshka Representation Learning (MRL) and demonstrates strong benchmark performance, ranking highly on the MTEB Multilingual leaderboard. Qdrant, a high-performance vector database, supports Gemini Embedding 2 from its public preview, offering a unified collection for all modalities and named vectors for hybrid strategies, facilitating efficient and precise search pipelines. This combination allows for innovative use cases like multimodal retrieval-augmented generation systems, cross-modal semantic search, and multilingual document intelligence, significantly simplifying the development of advanced search systems. Qdrant's robust architecture and managed cloud scaling ensure the system is production-ready, making it accessible for developers to implement sophisticated multimodal search capabilities with ease.

In Part 4 of a series on fine-tuning sparse embeddings for e-commerce search, the focus is on the trade-off between specialization and generalization of a SPLADE model trained on Amazon ESCI data. The model significantly outperforms BM25 in in-domain tests, but its performance varies in cross-domain evaluations, with improvements observed in other e-commerce datasets like Wayfair and Home Depot due to shared structural elements, while performance drops in general web search (MS MARCO) due to overfitting to e-commerce-specific patterns. To address generalization issues, a multi-domain model trained on combined datasets from Amazon, Wayfair, and Home Depot shows balanced improvements across domains, demonstrating the potential of diverse training data in retaining general language understanding while offering better cross-domain transfer. The document also outlines scenarios for choosing between domain-specific fine-tuning and multi-domain training, emphasizing the benefits of fine-tuning for single retailers with extensive data and multi-domain training for platforms serving multiple retailers. The text concludes by suggesting further enhancements, such as cross-encoder reranking and using larger base models, to improve e-commerce search capabilities.

In Part 3 of a series on fine-tuning sparse embeddings for e-commerce search, a SPLADE model, trained previously, is evaluated using Qdrant for indexing products, running retrieval benchmarks, and implementing hard negative mining, showing notable improvements over BM25 and off-the-shelf SPLADE models. The evaluation employs standard information retrieval metrics such as nDCG@10, with the fine-tuned model achieving a 28% improvement over BM25 and a 19% improvement over the off-the-shelf SPLADE, highlighting the importance of domain-specific training. Hybrid search with sparse and dense vectors is explored, revealing that while it provides moderate improvement with the off-the-shelf model, it can degrade performance when the sparse model is finely tuned. Hard negative mining through ANCE shows potential for additional performance gains by challenging the model with difficult negative examples, although it adds complexity. The study emphasizes the benefits of fine-tuning for domain adaptation, enhancing query expansion, term weighting, and domain-specific vocabulary, while managing latency effectively in a production setting.

This article, part of a series on fine-tuning sparse embeddings for e-commerce search, focuses on training the SPLADE model using Amazon's ESCI dataset on Modal's serverless GPUs. The ESCI dataset, notable for its graded relevance labels, allows the model to learn nuanced product matches by treating both exact and substitute products as relevant during training. The SPLADE model's performance depends on careful product text formatting, using specific tokens to maintain lexical signals. The training process involves a SparseEncoder built from a DistilBERT base model, utilizing a contrastive loss combined with sparsity regularization to optimize query and product embeddings. The article emphasizes the use of Modal's infrastructure for efficient training, highlighting persistent storage to manage checkpoints and the advantages of detached runs to prevent data loss. It also warns against the pitfalls of replacing transformers with static embeddings, which led to poor results due to the loss of contextual understanding essential for e-commerce queries.

Part 5 of the series on fine-tuning sparse embeddings for e-commerce search focuses on transforming the research-oriented SPLADE fine-tuning pipeline into an accessible tool for practical application. Previously, users had to navigate through multiple complex steps, including data formatting, query labeling, configuring environments, and manually publishing results, which were detailed across the first four parts of the series. The newly introduced tool, qdrant-sparse-finetune, simplifies this process by offering a streamlined, open-source command-line interface (CLI) and a web dashboard that automate the entire pipeline—from synthetic query generation and SPLADE training with ANCE, to evaluation and publishing on HuggingFace—using just a few commands. By eliminating the need for manual intervention and detailed technical knowledge, the toolkit enables users to achieve the 28% performance improvement over BM25 on Amazon ESCI demonstrated in the series, benefiting from automatic data handling, synthetic query generation, multi-backend GPU support, and interactive publishing. This evolution from research to product aims to make the powerful search model improvements readily accessible to users with a product catalog, without requiring them to delve into the intricacies of the underlying code.

In the first installment of a five-part series on fine-tuning sparse embeddings for e-commerce search, the article highlights the advantages of sparse embeddings over dense ones, particularly in addressing the shortcomings of BM25 in e-commerce applications. Dense embeddings, while effective at capturing semantic meaning, often fail in e-commerce contexts where exact matches are crucial, leading to issues like the misrepresentation of specific product attributes. Sparse embeddings, such as those generated by the SPLADE model, maintain individual term signals by projecting data onto a large vocabulary space, allowing for more precise and interpretable results. These embeddings outperform BM25 by 29% on Amazon's ESCI dataset, a significant benchmark in e-commerce search, by expanding queries with related terms learned from extensive training data. The series will explore the entire process of building a production-ready system, including data loading, GPU training, evaluation, and the integration of sparse vectors in Qdrant, a vector database optimized for such tasks, promising to enhance search accuracy without the need for traditional query rewriting or synonym expansion.

GlassDollar, a platform that assists enterprises in discovering and partnering with innovative startups, improved its high-recall sourcing capabilities by migrating from Elasticsearch to Qdrant. Initially, GlassDollar used Elasticsearch with OpenAI embeddings for its vector search needs, but as the platform grew to encompass millions of company documents, retrieval became a bottleneck, necessitating a shift to a system that prioritized recall over speed. The migration to Qdrant allowed for faster retrieval, better handling of query expansion, and a 40% reduction in infrastructure costs. This transition not only improved the quality of search results by aligning user intent with company descriptions through contextual chunking, but also increased user engagement, as evidenced by a threefold increase in the bookmarking of companies. GlassDollar's focus remains on enhancing accuracy through continuous improvements in query expansion and reranking models to ensure that the platform effectively matches corporate needs with innovative startup solutions.