July 2025 Summaries

PortfolioMind, a Web3-native AI research copilot, utilized Qdrant to enhance its real-time crypto intelligence capabilities, addressing the dynamic and volatile nature of the crypto market. Traditional crypto platforms focused on static insights, but PortfolioMind recognized the need for a more responsive approach that adapts to users' rapidly changing research interests, such as L2 scaling and DeFi yield fluctuations. By leveraging Qdrant's multivector user-intent modeling, PortfolioMind translated user interactions into actionable insights, using diverse data sources and rich metadata to create personalized intelligence. The adoption of Qdrant resulted in a 70% reduction in query latency, a 58% increase in interaction relevance, and a 22% rise in user retention. Looking ahead, PortfolioMind aims to further enhance its capabilities by exploring cross-user curiosity mapping, temporal drift tracking, and improved cold-start onboarding, thereby transforming real-time crypto research into more personalized and actionable intelligence.

Qdrant Edge is a lightweight, embedded vector search engine designed to meet the unique requirements of AI systems operating on edge devices with limited resources and without reliable network access. It represents a shift from traditional vector search used in cloud environments to a more localized approach, suitable for scenarios like robotics, mobile devices, and IoT systems. This new tool retains Qdrant's core capabilities, such as real-time ingestion and multimodal indexing, while being re-architected to function as a minimal library integrated directly into AI workflows. As vector-based reasoning becomes essential for embedded AI applications, Qdrant Edge aims to facilitate fast, local vector search, offering benefits like low-latency retrieval and privacy-preserving search, and is currently available in a private beta for select partners developing edge-native systems.

Researchers from ETH Zurich and Stanford have developed MIRIAD, an extensive open-source dataset consisting of 5.8 million medical question-answer pairs, each grounded in peer-reviewed literature, to address the lack of structured, high-quality data in medical AI. This dataset, built on the Semantic Scholar Open Research Corpus, aims to mitigate hallucinations in medical AI applications by providing a rich, context-driven knowledge base for Retrieval Augmented Generation (RAG) and enhancing embedding models. Qdrant, chosen for its simplicity, speed, scalability, and open-source nature, plays a crucial role in powering MIRIAD's storage and retrieval experiments. The dataset has demonstrated improvements in medical QA benchmarks and hallucination detection capabilities, and it is openly available for replication and benchmarking on HuggingFace. The researchers aim to keep MIRIAD updated annually, with plans for further integration with Qdrant and potential applications in medical AI, such as medical QA agents and discipline explorers.

Qdrant 1.15 introduces several enhancements aimed at improving vector search and text filtering capabilities. The update includes advanced quantization techniques, such as asymmetric quantization and 1.5 and 2-bit quantization, which optimize memory usage and accuracy for high-dimensional vectors. The release also features significant upgrades to the text index, including a new multilingual tokenizer supporting languages like Japanese and Chinese, stopwords filtering, stemming for better query matching, and phrase matching for exact searches. Additionally, Maximal Marginal Relevance (MMR) reranking is introduced to balance result relevance and diversity, improving search output in dense datasets. The transition from RocksDB to Gridstore as the default storage backend enhances ingestion speeds and storage management, while optimizations such as HNSW healing and connectivity estimation improve indexing efficiency. The release also includes an updated Web UI for easier collection configuration and encourages best practices through an intuitive setup flow.

Qdrant has joined the AI Agents and Tools category on AWS Marketplace, allowing customers to discover, purchase, and deploy its vector search engine for AI agents using their AWS accounts, thereby speeding up the development of agentic workflows. Qdrant offers a vector-native search engine designed to handle complex, unstructured data with accuracy and responsiveness, enabling long-term memory and real-time context retrieval for AI agents. Available in both Qdrant Cloud and Qdrant Hybrid Cloud for private deployments, it supports real-time metadata-aware retrieval, hybrid search, and multivector embeddings, which are crucial for AI agents to deliver context-rich and step-aware reasoning. The availability of Qdrant in AWS Marketplace simplifies procurement, offering centralized control over licensing and payments and ensuring seamless integration into AWS environments, thus facilitating AI innovation in industries such as healthcare, e-commerce, and media.

Choosing the right embedding model for vector search involves balancing multiple factors such as search quality, resource usage, language support, and cost constraints. While public benchmarks like MTEB can guide model selection, they may not reflect the nuances of domain-specific data, making it essential to evaluate models based on specific needs. Evaluating search quality is crucial and should be done before finalizing a model, taking into account the languages it supports and the effectiveness of its tokenizer. Embedding models serve different purposes—some excel in semantic similarity, others in retrieval or question answering—requiring precise task definition and a well-curated ground truth dataset for evaluation. Practical considerations like model size, sequence length, and infrastructure costs also play significant roles. Operational factors such as throughput, latency, and cost further influence model selection, with trade-offs necessary depending on the specific use case. As projects evolve, the need to revisit model choices may arise, and tools like Qdrant offer flexibility in managing multiple models and hosting options, including solutions like Cloud Inference for reducing latency and cost.

Qdrant Cloud Inference is a newly launched service that enables users to generate, store, and index embeddings for text and images in a seamless manner using a single API call, thereby simplifying workflows and accelerating application development for various use cases like RAG, multimodal, and hybrid search. By integrating model inference directly into Qdrant Cloud, it eliminates the need for separate infrastructure and manual data pipelines, reducing complexity, latency, and network costs. The service supports several curated models for different search applications and uniquely accommodates OpenAI CLIP-style models for multimodal tasks. Paid users of Qdrant Cloud can benefit from a monthly allocation of free tokens to ease onboarding and development processes, while inference capabilities are automatically enabled for all paid clusters with the appropriate software version. This integration allows for a more efficient and streamlined approach to developing AI applications without the need for additional tools or APIs.

&AI, co-founded by Herbie Turner and Caleb Harris, is revolutionizing patent litigation through a retrieval-first approach using modern AI tools and the Qdrant vector database. The platform focuses on invalidating patents by utilizing semantic retrieval to search vast, multilingual document corpora for prior art, minimizing reliance on human intuition and traditional methods. By embedding patent data into dense vectors with OpenAI's models and indexing them in Qdrant, &AI handles billions of tokens efficiently without increasing DevOps complexity. Qdrant’s architecture, offering features like multi-tenancy, payload filtering, and quantization, supports &AI in delivering accurate, fast searches while ensuring data privacy. This approach has enabled the development of new products and services, such as real-time invalidity checks and interactive claim visualization, while expanding into new jurisdictions. As &AI continues to enhance its AI patent agent Andy, the integration with Qdrant is set to grow, potentially utilizing upcoming features like hybrid and multimodal search to further streamline and automate patent workflows in the legal tech domain.

Vector Space Day 2025, hosted by Qdrant, is scheduled for September 26, 2025, at the Colosseum in Berlin, offering a full-day event dedicated to exploring advancements in retrieval, vector search infrastructure, and agentic AI. The event will feature deep-dives, lightning talks, and panels, focusing on AI infrastructure, search relevance, semantic retrieval, and the future of AI agents. Attendees can expect to engage in discussions on topics such as vector databases, AI memory, RAG pipelines, and hybrid search systems while also having the opportunity to network with developers reshaping AI systems. The event includes an after-party for further networking and discussions, and a global hackathon, "Think Outside the Bot," which challenged participants to innovate with vector search, concluded prior to the event. Key partners like AWS, Microsoft, and Neo4j underscore the growing momentum in vector search and retrieval-augmented generation systems, emphasizing the event's significance in the AI community.

Pento has developed an innovative recommendation system using Qdrant, aiming to connect people based on shared aesthetic tastes in art rather than popularity. By transforming user-art interactions into a semantic vector space, the system maps these preferences into clusters through Qdrant's recommendation API. Unlike traditional recommenders, which often rely on popularity and collaborative filtering, Pento's approach captures the dynamic and multifaceted nature of aesthetic taste by treating each user's interactions as evolving clusters. These clusters are scored based on recency and interaction frequency to maintain a dynamic representation of a user's current preferences. The system leverages these clusters, represented by medoids, to form a multivector profile that captures both the user's positive and negative aesthetic affinities. Using Qdrant's capabilities, the system matches users with artists whose styles align with their current tastes while filtering out those that do not. This method also addresses the cold start problem by quickly adapting to new users through early interactions with a diverse range of artworks, enabling immediate and relevant recommendations across various domains.

Alhena AI transformed its ecommerce operations by migrating to Qdrant Cloud, which unified its vector search infrastructure and improved its AI agents' performance and scalability. Initially, Alhena faced challenges with fragmented vector search systems, including FAISS and Pinecone, which caused complexities and impacted system reliability and customer onboarding. The shift to Qdrant allowed Alhena to consolidate its backend, enabling consistent performance across both small and large indexes, supporting hybrid search, and facilitating metadata filtering without additional infrastructure. This transition improved product recommendations, enhanced conversion rates, and allowed Alhena to deploy AI agents more swiftly and efficiently. Qdrant's capabilities, such as reduced latency and support for dense and sparse embeddings, empowered Alhena to meet internal SLAs and offer rapid, relevant responses to customers. The adoption of Qdrant's multitenancy and global deployment features further streamlined operations and bolstered Alhena's ability to scale while maintaining data residency compliance. Overall, this strategic move not only enhanced technical performance but also unlocked new business opportunities, allowing Alhena to focus on innovation and deliver superior ecommerce experiences.

In "The Hitchhiker's Guide to Vector Search," Clelia Astra Bertelli shares insights and practical advice from her extensive experience in the AI space, particularly focusing on vector search and its applications in Retrieval Augmented Generation (RAG). The blog post covers key aspects of vector search, including the importance of text extraction, chunking strategies, and embedding techniques, emphasizing the significance of clean data and meaningful chunks for effective RAG pipelines. Bertelli discusses hybrid searches that combine dense and sparse embeddings to enhance semantic understanding while maintaining keyword accuracy, and highlights the value of semantic caching and binary quantization for boosting search efficiency. The post also underscores the critical role of query optimization and evaluation metrics in building reliable vector search systems, encouraging a cycle of iteration and improvement. Throughout, Bertelli advocates for practical experimentation and continuous learning as essential components of advancing in the field of AI and vector search technology.

GoodData has transformed its analytics platform into an AI-powered solution by incorporating Qdrant's scalable vector database to overcome the limitations of traditional BI tools and large language models. Initially, GoodData faced challenges in scaling its prototype that leveraged OpenAI due to high compute costs and slow response times when attempting to load its entire semantic model into AI contexts. To address these issues, GoodData adopted a Retrieval-Augmented Generation strategy with Qdrant, which provided high performance and real-time embedding updates necessary for multilingual semantic layers. This transition enabled GoodData to deliver quick, intelligent responses and insights to its over 140,000 end customers, enhancing the platform's capability to provide real-time decision-making and advanced AI applications. Additionally, Qdrant's architecture supports GoodData's ambitions for advanced AI growth, including document-based semantic search and AI-driven search capabilities in a Kubernetes environment, marking a significant evolution in GoodData's offerings from static dashboards to dynamic, AI-driven analytics solutions.

Frankfurter Allgemeine Zeitung (FAZ) has leveraged Qdrant to develop a sophisticated search engine that unlocks its extensive 75-year archive of journalistic content. This initiative, led by a cross-functional team, addressed the limitations of traditional keyword-based searches by implementing a semantic search platform that utilizes Azure OpenAI's text-embedding model to create high-dimensional vector representations of content. Qdrant's ability to manage complex metadata and support real-time updates was critical, enabling FAZ to handle over 60 metadata fields and ensure rapid search performance across millions of articles. The system facilitates advanced filtering and context-rich search results, enhancing the user experience. As FAZ continues to refine its search capabilities, the next phase involves developing a hybrid search architecture that combines semantic and symbolic retrieval methods to offer both broad semantic understanding and precise control, thereby setting new standards in archival search and AI-driven journalism.