September 2025 Summaries

Nearly 400 developers, researchers, and engineers convened in Berlin's Colosseum Theater for the inaugural Qdrant Vector Space Day on September 26, 2025, focusing on the burgeoning field of AI-native search. The event highlighted the importance of community engagement, with participants exchanging insights on embeddings, hybrid search, and AI agent technology. Keynotes from Qdrant's leadership, including Co-Founder and CEO André Zayarni and CTO Andrey Vasnetsov, framed the shift towards more advanced AI-native search technologies, emphasizing the need for a retrieval layer tailored to unstructured data and agent workflows. The day featured talks by industry experts and breakout sessions exploring retrieval pipelines and agentic AI use cases, as well as networking opportunities that fostered connections among attendees. The event also included lightning talks and a hackathon, showcasing innovative applications of vector search. The day concluded with an upbeat after-party, cementing the event as a vital platform for knowledge exchange and community building in the AI domain.

The 2025 Hackathon, hosted by Qdrant and held in Berlin, showcased innovative uses of vector search beyond traditional chatbot applications, with participants worldwide presenting projects that integrated vector retrieval into various domains. The competition awarded $10,000 in prizes based on creativity, technical depth, and Qdrant usage. First place went to Vector Vintage for its 3D e-commerce exploration experience using Neo4j; second place to RoboBank, an AI system for robotic safety; and third place to Spatio-Temporal NPCs, which created evolving non-playable characters using spatio-temporal memory. Additional awards recognized projects like ReMap for event discovery, CosmicTwin for a social platform matching users to cosmic twinmates, and Bachata Vibes for an AI choreographer of dance sequences. The hackathon highlighted the versatility of vector retrieval in applications ranging from robotics and gaming to social platforms and educational tools, emphasizing the power and creativity of the Qdrant community in exploring future technologies.

Vector Space Day 2025, scheduled to take place in Berlin on September 26, will feature a comprehensive lineup of speakers and sessions exploring cutting-edge advancements in AI and vector search technologies. The event will include keynotes from industry leaders such as Andre Zayarni and Robert Eichenseer, who will discuss the future of vector search and AI system transformations. Attendees can choose from diverse breakout sessions divided into two tracks: Milky Way, focusing on architectures, infrastructure, and multimodal retrieval, and Andromeda, emphasizing AI workflows, agents, and applications. Highlights include practical insights from workshops on integrating knowledge graphs with retrieval systems, developing scalable AI memory, and deploying advanced multimodal embeddings. The event will conclude with a series of lightning talks from various innovators and an award ceremony for the Think Outside the Bot Hackathon, followed by networking opportunities at an after-party. Registration for the highly anticipated event is nearing capacity, underscoring the community's enthusiasm for these topics.

Tavus, a human-computer research lab, developed the Conversational Video Interface (CVI) to deliver natural, human-like interactions by reading tone, gesture, and on-screen context in real-time. To address the challenge of maintaining subsecond conversational flow, Tavus used Qdrant Edge for fast, local data retrieval, eliminating network latency by implementing per-conversation edge vector stores. This design allowed for immediate data processing, avoiding serialization delays and focusing on retrieval quality and multimodal accuracy. By reducing retrieval time to 20-25ms, Tavus maintained timely, accurate responses and enhanced the user experience, even for complex conversations. The architecture improved operational efficiency, with Tavus indexing millions of data points and providing a seamless launch experience without the need for customers to build their own Retrieval-Augmented Generation (RAG) systems. This approach demonstrated the effectiveness of architecture over micro-optimizations, enabling Tavus to prioritize quality and safety in their AI system.

MUVERA embeddings, developed by Google Research, address the challenge of slow multi-vector searches by transforming multi-vector representations into single vectors for faster initial retrieval and then using the original multi-vectors for reranking the top results. This approach combines the speed of single-vector searches with the accuracy of multi-vector retrieval, significantly improving search efficiency. MUVERA embeddings are created by clustering vector spaces and using Locality-Sensitive Hashing (LSH) techniques like SimHash to transform variable-length sequences into fixed-dimensional representations. The FastEmbed 0.7.2 version supports MUVERA, offering approximately 7x speed improvements while maintaining the quality of search, making it a practical solution for multi-vector retrieval applications. However, the larger size of MUVERA embeddings compared to traditional single-vector embeddings necessitates careful consideration of storage and retrieval efficiency.

The blog post describes how Maximum Marginal Relevance (MMR) can enhance search results by balancing relevance with diversity, particularly in the context of fashion discovery using the DeepFashion dataset. Traditional vector searches often return overly similar results, creating an echo chamber effect, while MMR aims to present a wider variety of options by reranking results based on relevance to a query and diversity from already selected items. This approach is particularly useful for fashion searches, where visual similarity doesn't always align with user intent, allowing users to explore a range of styles, such as bomber jackets, hoodies, and blazers, when searching for a "black jacket." MMR's implementation in the Qdrant vector search engine enables such diverse search capabilities, which can be further refined with metadata filtering for targeted discovery. This methodology can be applied to various domains beyond fashion, including document retrieval and recommendation systems, offering a more nuanced and exploratory search experience.

Fieldy AI, a wearable AI note-taking device that records and transcribes conversations into a searchable memory, faced reliability challenges with its initial vector database, Weaviate, due to persistent operational errors affecting data integrity and user trust. To resolve these issues, Fieldy migrated to Qdrant, a vector database known for its stability and suitability for self-hosted deployment, which eliminated query failures, reduced latency, and cut infrastructure costs by two-thirds, while scaling to handle tens of millions of embeddings. The migration process was efficient, leveraging existing vector schemas and updating backend API calls to enhance performance. With enhanced reliability and cost efficiency, Fieldy's engineering team is now focused on improving retrieval quality by introducing advanced filtering and embedding strategies to refine search results and enhance the device's memory recall capabilities.

OpenTable has revolutionized its restaurant discovery process by introducing Concierge, an AI-powered dining assistant that leverages Qdrant for efficient data management. Recognizing the shift towards conversational and intelligent search tools due to the rise of generative AI, OpenTable aimed to transform how diners find restaurants. Concierge, developed with a focus on accuracy, domain-centric responses, and speed, relies on Qdrant for its ability to handle sparse embeddings and provide high-precision filtering, which is crucial for narrowing down results from over 60,000 restaurants. The deployment of Qdrant Cloud simplified the process, ensuring stable, global operations from the outset. This new system not only enhances user experience but also serves as a testbed for refining OpenTable's AI and retrieval infrastructure, setting the stage for future innovations like conversational search and visual dish discovery, while maintaining operational stability and allowing the team to concentrate on model performance improvements.

The blog post delves into the intricacies of relevance score boosting and decay functions within the Qdrant search engine, aiming to demystify their applications for users. It focuses on three types of decay functions—Linear, Gaussian, and Exponential—each of which adjusts the relevance score based on how a numeric property deviates from an ideal value. These functions transform dataset properties, like size or ratings, into scores between 1.0 (most relevant) and 0.0 (not relevant), allowing them to influence the final relevance score meaningfully. Key parameters such as x, target, scale, and midpoint are explained to help users tailor decay functions to match their definitions of relevance, using examples like video lengths and promo code freshness. The article provides practical advice on setting these parameters and discusses the challenges of dynamically normalizing scores without pre-known parameters, emphasizing the importance of understanding the context of input data for effective score boosting.