Qdrant

In Qdrant, you attach a payload dict with source URL, title, and indexed at to every point.

Since the beginning of this year the number of Qdrant Python SDK downloads grew to over 1,000,000 per day. BTW. I’m in Milan at the AI Week conference. If anybody is around and wants to meet, DM me. 👋

vector stores like Qdrant are used in production

We have chosen Qdrant as a vector DB to give a context.

Now let's chat about TurboQuant implementation in the recent Qdrant engine release and why it actually matters. Quantization compresses vectors to reduce memory, speed up search, and cut infrastructure costs, right? But every existing method forces a tradeoff: - Binary quantization (BQ) is blazing fast but falls apart with non-centered distributions and small vectors. - Scalar quantization (SQ) is reliable but only compresses 4x. - Product quantization (PQ) goes further but tanks accuracy and speed. TurboQuant (TQ), originally developed by Google Research, takes a different approach: apply a fast Hadamard rotation to the vector before compression. This redistributes values evenly across coordinates, so the data distribution no longer becomes a liability. The result: 8x compression (4-bit) with recall comparable to scalar quantization. However, Qdrant implementation goes beyond the original algorithm: - Length renormalization (borrowed from RaBitQ) corrects the recall-degrading bias introduced by quantization error - Per-coordinate calibration fits the data to precomputed codebooks, recovering accuracy lost to distribution mismatch - Full support for cosine, dot product, and L2 as first-class distance metrics - SIMD-accelerated implementation for maximum throughput Benchmark results across different datasets: 4-bit TQ delivers recall on par with scalar quantization at double the compression ratio. On some datasets, it even outperforms SQ. 1-bit TQ consistently outperforms binary quantization at equivalent storage budgets — significant, given how widely BQ is used for extreme compression scenarios. But... TQ is not a silver bullet. BQ can be faster because it is just super simple. It really depends on the use case and data. TQ is another great compression approach that can optimize heavy vector search workloads. TurboQuant is available in Qdrant >=1.18. Try it out and share your feedback. Details on Blog https://lnkd.in/d4JH3tze

👋 Qdrant 1.18.0 release is here, with a new shiny metal SOTA #quantization , aka #TurboQuant [and other optimizations] TL;DR of TQ: same recall as Scalar Quantization(SQ), half the memory. Migration from SQ or BQ is a config change and a reindex. A webinar with the core team & us (me & you) asking them stupid questions regarding benchmarks will be on the 19th of May. 🔗 https://lnkd.in/gsW9Zvvu