Filtered Vector Search

8/7/2026

This post introduces and benchmarks ACORN-1, a query-time algorithm for improving filtered vector search performance in Qdrant. It details how ACORN-1 addresses the issue of graph connectivity degradation caused by metadata filters in HNSW. The post provides a comparative analysis of ACORN-1 against Filterable HNSW (an index-time optimization) and the default query planner, using specific benchmarks with single and double filters. It identifies scenarios where ACORN-1 excels (e.g., large value groups in indexed fields) and where other strategies are superior (e.g., Filterable HNSW for dense intersections, payload index for very sparse results). The contribution includes performance metrics, latency figures, and recall scores under various filtering conditions.

7/24/2026

This post details building an e-commerce search system on Qdrant, focusing on practical implementation strategies. Key contributions include: demonstrating hybrid retrieval with dense and BM25 sparse vectors fused in a single query; advocating for filtering within the query for efficiency and accurate facet counts; emphasizing embedding relevant product fields over using larger models; proposing quantization in RAM without rescoring for performance gains; shifting personalization from retrieval to ranking for better control; implementing merchandising as weighted formulas in a single rescore pass; and introducing a multi-faceted evaluation approach using an LLM judge, recall@10, intrusion, and persona overlap metrics. It also identifies a specific gotcha related to query fusion and sharding.

5/21/2026

Sunny Health migrated their core retrieval workload from Postgres to Qdrant, focusing on provider matching. They leveraged Qdrant's hybrid search capabilities, combining hard filters (network adequacy, licensing, language) with semantic similarity ranking. This resulted in up to a 57% reduction in retrieval latency for complex queries. The post highlights the importance of Qdrant's flexible JSON payload model for deeply nested provider data and its first-class geo re-ranking. Future plans include using Qdrant for insurance carrier matching, semantic specialty mapping, and as an agent memory layer.

1/9/2026

Flipkart implemented a real-time multimodal similarity search system using Qdrant for their Trust & Safety team. This system addresses limitations of prior batch-based methods (HBase with LSH) that took up to 9 hours for similarity searches. Qdrant was chosen for its deployment flexibility, efficient HNSW indexing, and support for high-dimensional embeddings. The new system reduced detection time to under 1 minute, integrating with the Java gRPC SDK and Prometheus metrics.

11/20/2025

Cosmos implemented a visual search system using Qdrant Cloud, leveraging named vectors to store multiple embedding types (CLIP, CNN, pHash, and color vectors) in a single collection. They developed an exact color search using CIELAB vectors and an application-side fusion model for hybrid search, balancing relevance, engagement, and aesthetics. This resulted in sub-second search latency and improved element processing pipelines.

11/19/2025

Introduced ACORN, a new search algorithm for filtered vector search that improves accuracy with multiple low-selectivity filters by exploring second-hop neighbors in the HNSW graph. Implemented Inline Storage for HNSW index, enabling disk-efficient vector search by storing quantized vector data directly within HNSW nodes. Enhanced full-text search with `text_any` condition and ASCII folding. Introduced Tiered Multitenancy, combining payload-based and shard-based multitenancy with tenant promotion from a fallback shard to dedicated shards.

9/2/2025

This post details how OpenTable integrated Qdrant into their Concierge AI Dining Assistant. Key technical aspects include the use of sparse embeddings for keyword expansions and fine-grained filtering to achieve high precision in narrowing down search results from a large dataset (over 60,000 restaurants). Qdrant's ability to handle sparse embeddings without HNSW graph quality degradation and its predictable performance for complex, multi-criteria filtering (reviews, metadata, structured data) were crucial for meeting latency targets. Qdrant Cloud simplified deployment.

7/10/2025

Alhena AI migrated from FAISS and Pinecone to Qdrant Cloud, unifying their vector search infrastructure. They leveraged Qdrant's hybrid search (dense and sparse embeddings), metadata filtering, and boosting capabilities to improve e-commerce conversion rates and agent response times. The post highlights the benefits of Qdrant's multitenancy for scaling and Qdrant Cloud for simplified global deployment and feature adoption.

7/9/2025

This post details GoodData's successful migration to Qdrant for their AI analytics platform, moving from an unsustainable LLM context loading approach to a RAG strategy. It highlights Qdrant's role in enabling near-real-time embedding updates and fast semantic search, contributing to AI assistant response times of 5-10 seconds. The technical details include the use of Qdrant's Helm chart for Kubernetes deployment and the performance metrics achieved (embedding updates in seconds, search results in 100ms).

6/11/2025

This case study details Lawme's migration from PGVector to Qdrant to scale their AI legal assistants. Key technical contributions include the successful implementation of Qdrant's binary quantization for efficient retrieval from tens of millions of legal vectors, and the utilization of metadata filtering with Qdrant's HNSW index to quickly filter queries by jurisdiction or case type. The post highlights the benefits of Qdrant's flexible deployment options, enabling Lawme to meet strict data residency and compliance requirements for legal clients.

6/10/2025

This post details how ConvoSearch, an AI-powered recommendation engine for D2C e-commerce brands, migrated from Pinecone to Qdrant. The migration was driven by Qdrant's significantly lower query latency (10ms vs 50-100ms) and superior metadata handling capabilities, which were crucial for ConvoSearch's real-time re-ranking and personalization needs. The implementation involved hosting Qdrant on dedicated servers with NVIDIA GPUs, enabling substantial revenue increases for ConvoSearch's clients, including a 60% uplift for The Closet Lover. The post highlights the direct business impact of Qdrant's performance and flexibility in enabling advanced e-commerce search functionalities.

6/10/2025

This post details the application of Qdrant's filtered vector search capabilities within the LegalTech domain. It specifically highlights the use of Filterable HNSW for pre-filtering, hybrid search combining semantic and keyword retrieval, late-interaction models for rich documents, ColBERT for token-level similarity, and score boosting for prioritizing legal logic in search rankings. It also mentions GPU acceleration, vector quantization (e.g., Binary Quantization), and enterprise-grade features like RBAC and compliance.

5/14/2025

This post details Tripadvisor's use of Qdrant to power their AI Trip Planner, which has resulted in a 2-3x revenue increase from engaged users. It highlights the creation of a user graph using Qdrant for personalized travel recommendations and conversational search, emphasizing Qdrant's suitability for real-time, unstructured data in AI applications.

5/13/2025

This post details Aracor's implementation of Qdrant for legal due diligence, highlighting the use of hybrid search (semantic + metadata filtering) for precise document retrieval. It emphasizes Qdrant's scalability and ability to handle large document volumes, leading to significant improvements in workflow speed and accuracy for legal professionals.

5/9/2025

This post details how Garden scaled its patent intelligence platform using Qdrant Cloud, specifically highlighting the impact of Qdrant's filterable HNSW capabilities. Garden's use case involves processing over 200 million patents, each with extensive metadata, requiring efficient vector search combined with surgical-grade filtering. The post emphasizes Garden's migration to Qdrant Cloud due to its managed Rust backbone, which offloaded operational burdens while maintaining source-level transparency. Key technical aspects highlighted include the use of 8-bit scalar quantization for efficient memory management (hot vectors in RAM, colder on disk), achieving sub-100ms p95 query latency, and achieving a 10x lower cost per stored GB. The migration process itself was streamlined, involving a weekend of scripted ETL to push embeddings into Qdrant Cloud with minimal code alterations. The business impact includes addressing a 10x larger patent corpus, managing hundreds of millions of vectors, and unlocking a new revenue line through high-confidence infringement detection.

5/1/2025

This post details how Pariti integrated Qdrant into their talent marketplace to improve candidate matching. They used Qdrant Cloud for zero-ops deployment and an intuitive Python SDK. The integration involved encoding job requirements, embedding candidate experiences, and comparing vectors. This resulted in millisecond-level search results, reducing candidate vetting time from 4 minutes to 1 minute and increasing the fill rate from 20% to 48%. Pariti is considering Qdrant's GPU-assisted indexing and vector quantization for future scaling.

4/24/2025

This post introduces a novel application of Qdrant for hotel search by integrating it with the Superlinked framework. It details the concept of 'Spaces' within Superlinked for creating specialized vector embeddings for different data types (text, numerical with logarithmic/linear scaling, categorical). The post explains how these diverse embeddings are indexed in Qdrant and queried using natural language, which Superlinked translates into weighted nearest neighbor searches and hard filters. It showcases a multimodal hybrid search approach that combines text, numerical, and categorical filtering for nuanced results, moving beyond traditional keyword/vector hybrid search.

4/22/2025

Introduced a Score-Boosting Reranker that allows combining vector similarity with custom logic based on payload data (e.g., prioritizing titles, boosting recent results via datetime, factoring in geographical proximity using gauss_decay). Implemented Incremental HNSW Indexing to extend existing graphs instead of rebuilding them. Optimized batch query processing by splitting large batches into chunks for concurrent execution across threads. Improved resource utilization during segment optimization by saturating CPU and disk IO more effectively. Reduced memory usage in immutable segments.

4/22/2025

This post details how Pathwork uses Qdrant for life insurance underwriting, achieving significant accuracy improvements (MSE from 3.5 to 1.8) through scalar quantization, hybrid search, and advanced filter utilization. Performance was enhanced by storing vectors in RAM and improved scaling methods (replicas, shards, nodes), reducing latency from 9s to 2s. Usage grew 50% MoM, processed thousands of cases, and scaling was managed with additional nodes, sharding, and replicas.

4/15/2025

This post details how Lyzr migrated from Weaviate and Pinecone to Qdrant to improve the performance and scalability of their AI agents. It quantifies the performance gains, including a >90% reduction in query latency (from 300-500ms to 20-50ms P99), 2x faster indexing, and a 30% reduction in infrastructure costs. The benchmarks show Qdrant handling over 1,000 queries per minute and sustaining throughput of more than 250 queries per second with over 100 concurrent agents, demonstrating significant improvements in query throughput, indexing performance, and resource efficiency compared to their previous solutions. The post also includes case studies from NTT Data and NPD, highlighting Qdrant's role in improving retrieval accuracy and low-latency retrieval for AI applications.

3/24/2025

This post details how HubSpot uses Qdrant to scale its intelligent AI assistant, Breeze AI. It highlights the challenges of scaling AI, why HubSpot chose Qdrant for its performance and developer-friendliness, and the resulting improvements in customer satisfaction and engagement. HubSpot is also exploring advanced features like multi-vector search and sparse vectors.

2/24/2025

This post introduces the concept of metadata automation and optimization as a critical component for enhancing vector search performance. It highlights how metadata can be used for segmentation, enrichment, and embedding into sparse vectors for hybrid search. The post also discusses LLM-powered metadata extraction and classification, and its application in role-based access control.

12/3/2024

This post details the implementation of a facial recognition system (Twin Celebrity app) using Qdrant. It describes the architecture involving FaceNet for embeddings, MTCNN for alignment, ZenML for orchestration, and Streamlit for the UI. The process involves an offline phase for generating and storing celebrity embeddings in Qdrant, and an online phase for querying user selfies against these embeddings. It also covers deployment strategies and the use of Qdrant's visualization tools for quality testing. Scalability considerations and potential future enhancements are discussed.

11/27/2024

Introduced a two-stage retrieval process for ColPali to address scaling challenges with visually rich PDFs. This involved pooling (mean and max) to reduce the number of vectors per page from 1030 to 38 for initial retrieval, followed by reranking with original high-resolution embeddings. Achieved a 13x speed improvement with mean pooling maintaining high NDCG@20 (0.952) and Recall@20 (0.917).

11/21/2024

This post details how QA.tech uses Qdrant for their AI testing agents, migrating from pgvector due to scalability issues. They leverage Qdrant's real-time analysis capabilities, batch operations to reduce network overhead, and efficient CPU utilization for embedding processing. The ability to store multiple embeddings per data point is highlighted as a key feature for their diverse use cases, enabling high-velocity lookups and supporting complex agent reasoning.

10/17/2024

This post details how Sprinklr integrated Qdrant into their AI-driven customer experience solutions, focusing on RAG and GenAI applications. It highlights Qdrant's benefits for Sprinklr, including developer-friendly documentation, high customizability (quantized vectors, memory mapping), and cost-efficiency. The post presents benchmark results showing Qdrant's superior write performance, low latency (20ms P99 for 1M vectors), and high throughput (250 RPS) compared to Elasticsearch and Milvus. Sprinklr also achieved a 30% cost reduction in their retrieval infrastructure.

9/10/2024

This post details how Nyris, a visual search solutions provider, adopted Qdrant for their visual product search engine. It covers Nyris's journey from early SIFT features to CNN embeddings, their evaluation process for vector search engines (focusing on accuracy, speed, insert speed, TCO, data sovereignty, and Qdrant's specialization), and the specific benefits they derive from Qdrant in production (JWT security, scalability, flexible search, data handling, Rust implementation, cost-effectiveness). It also touches on Qdrant's quantization and integer8 support as key differentiators.

8/28/2024

This post details Kern AI's use of Qdrant for financial and insurance AI solutions, specifically highlighting the application of multi-vector storage and hybrid search/filters for RAG chatbots. It emphasizes achieving low hallucination rates (<1%) and reducing customer service response times. The post also mentions Qdrant's interactive Discord community and its open-source nature as key decision factors.

7/10/2024

This post details Kairoswealth's successful implementation of Qdrant for wealth management use cases including internal RAG, financial regulatory reports RAG, and recommendations. It highlights Qdrant's ability to handle 2.4 million vectors with consistent high performance (queries under 0.3 seconds) and significant memory efficiency (8x smaller storage size compared to previous solutions). The post also emphasizes Qdrant's advanced features like multi-tenancy and support for recommendation engines, and the value of Qdrant's community and support.

5/1/2024

VISUA implemented Qdrant for computer vision quality control and anomaly detection, achieving 40x faster query processing and 10x scalability. They utilized Qdrant's hybrid query capabilities, support for multiple vectors per record, and quantization. The post highlights Qdrant's advantages over competitors like Weaviate and Pinecone in handling hybrid queries and performance for VISUA's specific use cases.

4/12/2024

This post features a discussion about the application of RAG systems, including visual RAGs, and highlights Qdrant's seamless performance in handling complex multimedia data retrieval, emphasizing its accuracy and ease of use in such advanced applications. The speaker specifically mentions using Qdrant for a family tree app that uses video interactions with deceased loved ones, and praises Qdrant's search capabilities for this visual RAG use case.

4/12/2024

This post demonstrates building search and RAG over an OpenAPI specification using Qdrant and Trieve. It details how Qdrant's group-based system can be used to organize API endpoints by tags, enabling more structured search results. The technical implementation involves parsing the OpenAPI JSON, creating groups in Qdrant for each tag, and then inserting individual API paths as points within their respective groups. The post also briefly mentions Tantivy as a complementary search engine.

3/27/2024

This post highlights VirtualBrain's use of Qdrant as a core component of their AI assistant for knowledge workers, emphasizing RAG capabilities. It details how Qdrant's scalability, speed, low latency, and open-source nature were critical factors in VirtualBrain's decision to use it for handling large datasets (terabytes) and performing complex tasks like audits and RFPs. The post also touches on the importance of fast ingestion and semantic search to meet user expectations for quick responses, and the use of Qdrant for document retrieval and AI-assisted tasks.

3/25/2024

This post details the application of Qdrant in a production-scale RAG system for real-time news distillation. It highlights Qdrant's role in enabling real-time indexing and semantic similarity searches, crucial for processing over a million news articles daily. The author emphasizes Qdrant's contribution to modularity and performance in AI pipelines, contrasting it with less flexible all-in-one solutions. The post also touches upon the importance of embedding quality for effective semantic search and the benefits of Qdrant's evolving features, such as sparse vector support, in enabling such products.

3/19/2024

This post details Dailymotion's implementation of Qdrant for content-based video recommendations. It covers their approach to generating video embeddings from metadata, audio, and transcripts, and how these embeddings are used with Qdrant to address the cold-start problem for new and niche videos. The post highlights the use of a streaming pipeline for real-time updates and the positive impact on click-through rates for low-signal content.

2/27/2024

Dailymotion implemented a content-based video recommendation engine using Qdrant. They process video metadata (title, tags, description, transcript) into MUSE embeddings and use Qdrant for fast and accurate similarity search. This resulted in a significant reduction in recommendation processing time (from hours to minutes) and a 3x increase in interaction and CTR for recommended videos, particularly for fresh and low-signal content. Qdrant's capabilities in handling high-dimensional data, scalability, and metadata filtering were crucial for this implementation.

2/25/2024

This post provides a comparative analysis of Qdrant and Pinecone, focusing on their vector database capabilities. For Qdrant, it details its advanced similarity search metrics, payload indexing and filtering features (including keyword matching, full-text filtering, numerical ranges, nested object filters, and geo filtering), and its support for hybrid search with sparse vectors. It also touches upon Qdrant's Rust implementation, scaling, multitenancy, quantization, and deployment flexibility.

2/24/2024

This post provides a foundational explanation of vector similarity, its calculation methods (Dot Product, Cosine Similarity, Euclidean Distance, Manhattan Distance), and its significance in AI applications. It also outlines the challenges of scalability, efficiency, and security in vector similarity search and briefly mentions Qdrant's role in addressing these challenges.

2/6/2024

This post details how Dust integrated Qdrant into their AI assistant platform, focusing on retrieval augmented generation (RAG). Key technical aspects include Dust's two-step approach: starting with Qdrant Cloud for quick setup and then scaling and optimizing by leveraging Qdrant's features like quantization (Scalar Quantization) and MMAP payload threshold configuration. The post highlights the resulting cost reduction (2x) and ability to maintain low latency and high accuracy at scale. Observability features of Qdrant were also noted as beneficial for scaling.

1/29/2024

This post discusses the 'bitter lesson' in AI, emphasizing the power of data and compute over handcrafted methods. It delves into retrieval-augmented generation (RAG) workflows, specifically how the Yokot AI platform handles diverse data inputs and the importance of re-ranking for improving LLM output quality. The discussion touches on data flexibility, document uploading, web scraping, and the evolution of information retrieval methods.

1/15/2024

This post introduces updated benchmarks for Qdrant, comparing its performance against other vector search engines. Key changes include applying suggestions for more efficient engine operation, resulting in up to four times improvement in certain cases. A new dataset of 1 million OpenAI embeddings for RAG applications has been added. The benchmark methodology now separates latency and requests-per-second (RPS) scenarios, simulating 1 or 100 parallel readers. The post reiterates Qdrant's commitment to open-source, accessible benchmarks, and focuses on production-ready vector databases over indexing libraries.

1/10/2024

This post details Qdrant's strategy for in-place filtering during graph traversal in vector search, which optimizes precision and exactness when scaling to billions of vectors. It introduces techniques like using subgraphs with overlapping intervals and geo hash regions for enhanced precision and connectivity. The post also discusses controlling search precision and speed by adjusting the beam size in HNSW indices and highlights the challenges of maintaining graph connectivity during filtering. Additionally, it touches upon vector compression methods.

1/9/2024

This post discusses the application of vector search in image search and RAG projects. It highlights the effectiveness of the Dino V2 model for image representation, noting its ability to understand objects and patterns without fine-tuning. The post also touches on challenges in data deduplication for real estate listings and the use of multi-language embedding models for document retrieval in RAG. It mentions the importance of efficient storage and retrieval, including using memory for CI integration and the complexities of metadata filtering in vector databases.

1/9/2024

Malt successfully transitioned to Qdrant for their freelancer matching application, significantly reducing latency from 10 seconds to 1 second. This involved adopting a retriever-ranker architecture, utilizing multilingual transformer-based encoder models for high-fidelity embeddings, and leveraging Qdrant's capabilities for geospatial filtering and scaling.

1/9/2024

This post discusses building LLM-powered applications in production, focusing on complex search, conversational search, and personalization. It highlights challenges in LLM service downtime and strategies for mitigation, as well as the unique approach of Traversal AI in creating unified platforms for LLM integration and domain-specific search. The importance of user experience (UX) in GenAI product development is also emphasized.

1/9/2024

This post details the use of Qdrant for high-performance entity matching, specifically for hotel data. It highlights the challenges of data inconsistency, duplication, and real-time processing across multiple languages. The solution leverages Qdrant's HNSW algorithm for efficient vector indexing, demonstrating significant improvements in speed and recall compared to initial PostgreSQL experiments with Pgvector. The post also touches upon the importance of geofiltering for accurate matching and the experimentation with different embedding models (Mini LM) and HNSW parameters (M and efConstruct) to optimize performance.

1/9/2024

This post details the application of Qdrant's vector search capabilities for a music recommendation system. It describes a novel approach to generating song embeddings using natural language descriptions queried from LLMs, and then encoding these descriptions with sentence transformers. The core of the recommendation system relies on Qdrant for performing cosine similarity search on these embeddings to find songs matching user-defined moods or vibes. The post highlights the use of Qdrant's vector indexing and search for this specific use case.

1/5/2024

This post demonstrates the integration of Qdrant with n8n to build a "chat with a codebase" service. It details two n8n workflows: one for ingesting GitHub repository data into Qdrant, and another for querying this data to provide chat functionality. The ingestion workflow utilizes Qdrant's Vector Store Insert node, GitHub Document Loader, OpenAI Embeddings, and Recursive Character Text Splitter. The chat workflow uses Qdrant's Vector Store Retrieve node, Retrieval Q&A Chain, OpenAI Embeddings, and OpenAI Chat Model. It provides specific node configurations and prerequisites for setting up the system.

2/28/2023

This post details how Qdrant's vector computation capabilities, specifically its efficient nearest neighbor search and distributed deployment mode, enhance Large Language Models (LLMs) for enterprise customers like Pienso. It highlights Qdrant's ability to provide dynamic long-term storage, preserve extensive context windows, and enable scalable, cost-efficient high vector computation. Technical details include storage efficiency benchmarks (128 million documents in 20.4GB storage), the use of Rust for performance, memmap storage for fast access, and write-ahead logging for data safety.

2/28/2023

This post details how bloop uses Qdrant for its semantic code search engine. It explains the problem of semantic search for code, the benefits of using embeddings and vector search, and how Qdrant provides fast, accurate, and reliable vector similarity search for large codebases. The post emphasizes Qdrant's performance and ease of integration via its Rust client.