Qdrant Feature Trails

Qdrant enhances its filtered vector search capabilities with ACORN algorithm for improved search quality with low-selectivity filters, Inline Storage for HNSW index for disk efficiency, and full-text search enhancements. OpenTable leverages Qdrant for its Concierge AI Dining Assistant, utilizing sparse embeddings and high-precision filtering to narrow down restaurant searches from over 60,000 options to a single restaurant with predictable performance. Flipkart leverages Qdrant for real-time mul. Cosmos uses Qdrant Cloud for visual search, powering text, color, and hybrid search with named vectors and application-side fusion for precise color matching and sub-second latency.

Qdrant continues to optimize performance for vector search, demonstrating significant speed improvements for disk-based retrieval. This post details benchmarks showing Qdrant achieving 2x higher throughput and 50% lower latency than Elastic's DiskBBQ at the same recall target, using significantly less CPU and RAM per node. This is achieved through advanced quantization (TurboQuant 4-bit), async disk scoring, and a two-stage retrieval process. The results highlight Qdrant's efficiency and performance. Version 1.15 introduces new quantization modes including 1.5-bit and 2-bit quantization for improved compression and accuracy, as well as asymmetric quantization for combining binary storage with scalar queries. Text indexing is enhanced with multilingual tokenization, stop words, stemming, and phrase matching. MMR reranking is introduced for diversifying search results. Optimizations include HNSW healing and migration to Gridstore for faster ingestion.

Qdrant now supports branch-aware semantic code search, enabling users to scope vector search queries to specific Git branches. This addresses the 'wrong-version problem' where a static vector index might return outdated information for a particular branch. The system indexes code by structural units (functions, classes, etc.) and uses Git commit information to track versions. Queries are filtered at runtime to ensure results are relevant to the user's current branch context. This capability is demonstrated with Qdrant's own codebase and is available for use with documentation or codebases.

Qdrant Edge enables on-device vector search for real-time anomaly detection in video streams. This system leverages a split architecture with an immutable HNSW shard for baseline normal activity and a mutable shard for live writes, allowing for sub-millisecond kNN lookups with no network dependency and full offline resilience. Video embeddings are provided by Twelve Labs Marengo 3.0, and NVIDIA Metropolis VSS orchestrates GPU-accelerated ingestion on Vultr Cloud GPUs. The system escalates anomalous clips to the cloud for higher-fidelity analysis, reducing cloud processing volume and costs while maintaining high anomaly detection rates. This approach extends beyond surveillance to manufacturing safety, retail analytics, and traffic monitoring.

Qdrant is being used to build agentic retrieval systems that go beyond simple semantic search. These systems leverage multi-vector representations, hybrid search, and LLM orchestration to enable granular retrieval and complex, multi-criteria searches. PortfolioMind uses Qdrant to model dynamic user curiosity in real-time crypto intelligence by transforming user interactions into multivector user-intent models. This involves ingesting diverse data, embedding it with rich metadata, and using clustering to identify micro-interests. Qdrant's fast, filterable searches, native multivector support, low-latency retrieval, and managed services were key to their success, resulting in significant latency reduction and improved user retention. Future plans include cross-user curiosity mapping, temporal drift tracking, and improved cold-start onboarding.

Qdrant Hybrid Cloud is launched, enabling deployment in any environment (cloud, on-premise, edge). This offering integrates with various infrastructure providers (OCI, Red Hat OpenShift, Vultr, DigitalOcean, OVHcloud, Scaleway, Civo, STACKIT) and AI frameworks (LlamaIndex, LangChain, Haystack, Airbyte, JinaAI, Aleph Alpha). Tutorials demonstrate building AI applications like RAG systems, chatbots, and recommendation engines on partner platforms, emphasizing flexibility, data sovereignty, and con. This post specifically details the integration with Jina AI's embedding models for enhanced RAG capabilities, emphasizing seamless deployment, scalable vector search, and cost efficiency.

Qdrant Cloud now offers GPU-accelerated indexing for up to 4x faster HNSW index builds, utilizing NVIDIA T4 GPUs. This is beneficial for high-write workloads. Multi-Availability-Zone (Multi-AZ) replication is introduced for enhanced availability, offering a 99.95% Uptime SLA on the Premium tier. Audit logging is now available on all paid clusters, capturing API operations for compliance and security, with logs accessible via an endpoint for SIEM integration.

Qdrant is being used in a wide variety of innovative projects showcased through hackathons, demonstrating its versatility beyond traditional chatbot applications. These projects span areas like robotics, 3D e-commerce, spatial-temporal NPC behavior, route-aware event discovery, personality-driven social platforms, AI choreography for dance, and multi-modal classroom assistants. The hackathons highlight the use of Qdrant with various AI frameworks and technologies, including Neo4j, CrewAI, Mistral-7B, and Apify. Sentinel, a hackathon winner, uses Qdrant for semantic search to detect misinformation by comparing news articles.

This post analyzes the limitations of using pgvector for vector search, highlighting scenarios where dedicated vector databases like Qdrant offer superior performance and features. It details six conditions under which pgvector is sufficient and explains why most applications quickly outgrow it due to limitations in dataset size, metadata filtering accuracy, hybrid search capabilities, and architectural flexibility. The post also addresses the common concern of data synchronization between relational databases and dedicated vector stores, offering solutions and guidance.

Qdrant is being developed as fundamental infrastructure for AI, emphasizing composable vector search primitives that can be combined at query time. This approach aims to address architectural limitations encountered in production AI workloads, such as RAG and agents, by providing flexible and performant retrieval across diverse environments from edge devices to supercomputers. The focus is on building a robust, long-lasting engine that controls the stack down to assembly, written in Rust for predictability and performance. Key aspects include supporting dense vectors, sparse vectors, metadata filters, multi-vector representations, and custom scoring. The goal is to provide one retrieval architecture from the data center to the device, enabling agentic AI and on-device applications.

Qdrant has introduced FastEmbed, a Python library for fast and lightweight embedding generation. This post details its focus on production-ready embeddings, emphasizing speed, efficiency, and accuracy. It highlights techniques for improving embedding models, including quantized models for enhanced CPU performance and future plans for GPU-friendly quantized models. The post also discusses selecting default models based on MTEB benchmarks and calibrating them for domain-specific tasks, and contras. This post details the integration and benefits of Google's Gemini Embedding 2, a natively multimodal embedding model capable of processing text, images, video, and audio into a unified vector space. It highlights Qdrant's ability to store all modalities in a single collection, use named vectors for hybrid strategies, and leverage MRL-friendly architecture for multi-stage retrieval. The post also discusses production-ready features like quantization and payload filtering for scaling multimodal applications.

Qdrant enhances its relevance scoring capabilities by detailing the implementation and usage of decay functions (Linear, Exponential, Gaussian) within its score boosting functionality. These functions allow numeric payload fields or similarity scores to influence search relevance by transforming values into a 0.0-1.0 scale based on a target value and decay parameters (scale, midpoint). The post provides examples for use cases like time-based relevance, distance-based relevance, and price sensitivity. Version 1.17 introduces Relevance Feedback Query, a scalable, vector-native approach to incorporating relevance feedback by using lightweight feedback on a few top results to create context pairs of more- and less-relevant examples. These pairs adjust the scoring function during the next retrieval pass by modifying how similarity is computed, leading to improved recall without retraining models. Weighted Reciprocal Rank Fusion (RRF) is also introduced to allow assigning weights to different rankers.

Qdrant introduces a beta Vector Data Migration Tool designed to simplify moving data between different Qdrant instances and from other vector database providers. This tool streams vectors in live batches, supporting migrations between open source and Qdrant Cloud, or between cloud regions. It allows for changes in collection configuration, such as replication factor or quantization method, during migration. The tool can be run as a container on any machine with connectivity to both source and target. Bazaarvoice successfully migrated billions of vectors from PostgreSQL to Qdrant, achieving a ~100x reduction in storage footprint and sub-100ms query latency with ~98% recall, demonstrating the effectiveness of Qdrant's quantization and multitenancy features for large-scale data migration and operational efficiency.

Dragonfruit AI has successfully scaled its real-time computer vision platform using Qdrant. The solution leverages a Split AI architecture, processing inference on edge devices and utilizing Qdrant in the cloud for vector search. Key aspects include multi-camera person re-identification, self-checkout product verification, and full-frame semantic search. Qdrant's per-collection configurability, efficient numeric formats (float16), and open-source nature were crucial for meeting strict latency targets and managing large-scale vector data across thousands of locations. This enabled faster development and deployment of new AI agents for various industries.

Qdrant introduces granular access control via JSON Web Tokens (JWT) to implement Role-Based Access Control (RBAC), allowing fine-grained permissions for users and restricting access to sensitive endpoints. This enhances security for enterprise customers and those productionizing large-scale Generative AI applications. The release also includes optimized shard transfers using the `wal_delta` method for faster node recovery and native support for `uint8` embeddings, offering memory savings and sea. Qdrant Cloud now offers a Cloud API for simplified management via API-driven control and Terraform support. It also introduces granular Database API Keys with TTL expiration policies and payload-based filters for fine-grained access control to data within clusters, collections, or even at the vector level, with read-only or read/write permissions. Single Sign-On (SSO) is now available for Premium Tier customers, integrating with Okta, Google Workspace, Azure AD, SAML, and PingFederate for simplified authentication and MFA enforcement. Advanced monitoring and observability features are enhanced with Prometheus/OpenMetrics support, enabling real-time performance insights directly in the Qdrant Cloud Console and integration with third-party tools like Datadog and Grafana for deeper analytics and customizable alerts.

Qdrant continues to explore and integrate with AI frameworks to enhance LLM application development. This post discusses the challenges and strategies for using LLMs as judges for evaluating chatbot responses, focusing on cost-effectiveness and real-time assessment. It highlights the use of smaller models for evaluation, the importance of tailored metrics, and techniques for mitigating issues like hallucinations and jailbreaking. The discussion touches upon integrating evaluation into the development process, recommending frameworks like Ragas, Quotient AI, and Arize Phoenix for assessing RAG system performance. It also details common RAG underperformance causes related to data ingestion (chunking strategies), embedding model selection (using MTEB benchmarks and FastEmbed), and retrieval optimization (Precision@k, MRR, NDCG).

Qdrant has addressed a critical arbitrary file upload vulnerability (CVE-2024-2221) in versions prior to v1.9.0. The fix restricts file uploads to a dedicated folder. Qdrant cloud deployments were not materially affected due to read-only filesystems and default authentication. Users are advised to upgrade to v1.9.0 or above. This update also incorporates a fix for CVE-2024-3829, further enhancing security.

Qdrant Hybrid Cloud is launched, enabling deployment in any environment (cloud, on-premise, edge). This offering integrates with various infrastructure providers (OCI, Red Hat OpenShift, Vultr, DigitalOcean, OVHcloud, Scaleway, Civo, STACKIT) and AI frameworks (LlamaIndex, LangChain, Haystack, Airbyte, JinaAI, Aleph Alpha). Tutorials demonstrate building AI applications like RAG systems, chatbots, and recommendation engines on partner platforms, emphasizing flexibility, data sovereignty, and control. This post specifically details the integration with Scaleway, highlighting its AI-focused resources, scalability, European focus, and sustainability commitment, and provides a tutorial for building RAG applications on this combined platform.

Qdrant now supports the ARM architecture, providing Docker images compiled specifically for ARM users. Performance benchmarks show that ARM64 is approximately 10% slower on average and 20% slower on the median for vector search compared to x86, but offers more consistent performance and cost savings on cloud providers like AWS.