TurboQuant on Qdrant Edge: 8× Less Disk? We Tested It.

Can TurboQuant really deliver 8× less disk usage on Qdrant Edge? In this video, we put it to the test with a real benchmark using 20,000 vectors, 384 dimensions, six configurations, and 200 queries per configuration. But the benchmark revealed something more interesting: there are two different TurboQuant approaches in Qdrant Edge, and they have very different tradeoffs between storage and recall. In this video, we: Benchmark TurboQuant on Qdrant Edge Compare float32, TurboQuant bit-depths, and Turbo4 datatype Inspect the actual Qdrant Edge shard files Explain why some TurboQuant configurations don't significantly reduce disk usage Break down how rescoring helps preserve recall Uncover a sparse-file disk measurement issue Compare recall, disk usage, and latency Validate results against an independent NumPy ground truth The key takeaway: the “8× smaller” claim depends on which TurboQuant knob you use. Everything is based on an actual local benchmark, with the code and results available so you can reproduce the experiment yourself. Tools & Technologies: Qdrant Edge · TurboQuant · Vector Search · Quantization · NumPy · Python #Qdrant #TurboQuant #VectorDatabase #AI #MachineLearning #RAG #VectorSearch #GenerativeAI