July 2026 Summaries

Filtered vector search can become problematic when metadata filters disrupt a coherent nearest-neighbor graph, creating isolated clusters, which leads to ineffective traversal. Qdrant addresses this issue using two methods: ACORN, which enhances the graph at search time, and Filterable HNSW, which strengthens it at index time by adding additional edges for indexed payload values. These methods were tested on a dataset of one million image vectors using various filtering strategies, with ACORN proving beneficial in scenarios where fields lack extra edges. However, Filterable HNSW generally provided better performance by maintaining high recall rates at lower latencies, especially when filters were tight. ACORN's effectiveness was limited to fields without additional edges, and it required higher latencies to achieve comparable recall. The study concludes that while ACORN is useful for repairing traversal in certain conditions, integrating payload indexing to enable extra edges generally enhances search efficiency.

In the blog post "Lessons From Building E-Commerce Search on Qdrant," Dylan Couzon outlines the development of Qdrant Shopping, an e-commerce search platform that integrates various search and recommendation services into a single, efficient system. The platform was tested on a massive dataset of over 5.8 million Amazon fashion products, achieving rapid search results with a single API request. Qdrant's approach includes hybrid retrieval using dense vectors and BM25 for precise and intent-based matches, with a focus on filtering within queries to optimize search results. The importance of embedding the correct data fields is emphasized over the choice of model, as demonstrated by a precision benchmark that showed diminishing returns from larger models. The system also incorporates personalization by adjusting ranking rather than retrieval and uses a flexible merchandising strategy that allows non-engineers to tweak search result orders without additional services. Evaluation methods are designed to avoid biases from previous models, ensuring relevance and precision through a combination of metrics. Qdrant Shopping's architecture leverages a single collection with diverse vector types and operates on a scalable Qdrant Cloud infrastructure, providing a template for implementing similar solutions in other product catalogs.

Efficiently managing bulk data uploads to Qdrant is crucial for maintaining system stability and performance, especially when handling millions of vectors, payloads, and indexes. Different vector types, such as dense and sparse vectors, require distinct indexing approaches, impacting memory usage and search performance during ingestion. The guide emphasizes the importance of choosing the right bulk upload strategy tailored to specific workload needs, whether optimizing for upload speed, memory usage, or search availability. Best practices include storing dense vectors on-disk to alleviate memory pressure, creating payload indexes before uploading to enhance search efficiency, and using quantization to balance memory and search performance. Additionally, techniques like batching, parallelization, and sharding are universally recommended to enhance throughput and stability. The guide cautions against a one-size-fits-all approach, advising a thoughtful configuration based on workload requirements and system resources to ensure predictable and safe bulk data ingestion.

Qdrant has outperformed Elastic's DiskBBQ in a benchmark that demonstrated Qdrant's ability to deliver roughly twice the throughput and half the latency using only a third of the compute resources compared to DiskBBQ. The benchmark highlighted discrepancies in Elastic's testing methodology, which did not utilize Qdrant's two-stage retrieval and async disk scoring features, leading to an I/O bottleneck. Qdrant employed TurboQuant 4-bit quantization on smaller instances than Elastic, achieving superior performance on the same dataset. The Qdrant engine, available both on open-source and Qdrant Cloud, was tested on AWS m6g instances and outperformed Elastic's setup, which used GCP n4-standard-8 nodes. Elastic's DiskBBQ, positioned as a memory-efficient alternative, requires substantial RAM to maintain JVM stability, while Qdrant's approach leverages advanced quantization and disk scoring to maintain high efficiency even on smaller hardware configurations.

Branch-aware semantic code search using Qdrant enhances traditional code searching by indexing codebases as vectors based on meaning, rather than mere lexical matches, which allows AI agents to retrieve relevant context efficiently. Unlike conventional methods like grep, which only reflect the currently active branch, semantic search faces challenges due to its static nature, leading to potential mismatches between the indexed version and the branch-specific code. Branch-aware search resolves this by scoping queries to the live view of a specific branch—considering its own commits, inherited changes, and excluding replaced elements. This approach involves making strategic decisions on what to index, using stable identities like paths and qualified symbol names to track changes across versions, and maintaining synchronization with Git through a derived index. The process ensures that each query is accurately scoped to the correct branch, thereby providing precise results even as code evolves across different branches. The implementation of branch-aware search is detailed in a tutorial, highlighting the nuances of handling merges, rebases, and scaling as repositories grow, ensuring efficient version tracking and retrieval without cluttering the system.