Qdrant Vector Search Engine
Running a Vector Database Comparison? Benchmark Qdrant on Your Own Data
Measure recall, latency, and throughput on your own vectors and filters. Start a free Qdrant Cloud cluster with no credit card and run your first test today.
TRUSTED BY ENTERPRISE TEAMS
Vector database benchmark
Run an Open Source Vector Database Benchmark on Your Vectors
A published vector db comparison describes someone else’s machine and dataset. The numbers that decide your project come from your vectors, your filters, and your latency budget.

Build the benchmark with Supernova.
Supernova is Qdrant’s open source toolkit for vector search benchmarks. It embeds your dataset, computes exact ground truth, loads the vectors into Qdrant, and measures latency and recall under load.
Compare at the same precision.
Approximate search trades recall for speed, so read each speed figure at a matching recall level.
Test many index settings in one run.
Supernova sweeps loading and load tests across a matrix of index and search settings and combines them into one report.
Filtered search
Keep Recall When Filters Tighten
Most production queries carry a filter. Qdrant adds filter-aware links to its graph index and plans each query, so filtered search keeps its recall as the filter narrows.
99.9% to 100% recall across four filters.
In Qdrant’s published test on one million 96-dimension vectors, the default query planner held 99.9% to 100% recall@10 on every single-keyword filter, down to 0.012% of points.
Filter during the graph search.
Index a payload field, and Qdrant links points that share its values, so the search walks a connected path to the nearest neighbors that match.
Vector database performance comparison
Read Latency and Throughput Separately
A vector database performance comparison needs two numbers, because one user waiting and many users at once are different loads.
Latency for the user who waits.
Single-threaded runs show the response time one user feels, including p95.
Throughput for traffic at scale.
Parallel runs show requests per second under load.
Tune per query.
Raise hnsw_ef for higher recall or lower it for speed, and record both numbers at each setting.
Results
What Teams Measured After Their Own Comparison
Bayer
benchmarked several providers when its first prototype ran out of room, chose Qdrant on price-performance, latency, and openness, and now runs about 135 million points for 116,000 employees.
Read the Bayer case studyLyzr
20 to 50 ms P99 query latency with 100 agents running, more than 250 queries per second sustained, and 30% lower infrastructure cost.
Read the Lyzr case studyHimalayas
about 30 ms median inside Qdrant across 25M+ requests a month.
Read the Himalayas case studyGetting started
From Sign-Up to Your First Benchmark Run
Create a free cluster.
Sign up for Qdrant Cloud with no credit card, then copy the cluster URL and an API key.
Load a sample of your data.
Follow the Supernova quickstart to embed a sample of your data and load it into your cluster.
Run your queries.
Use your real filters and record recall, p95 latency, and requests per second.
