Is Qdrant a better choice than pgvector for high-dimensional image search?

All answers to this question.

For 2048 dimensions, you will definitely see a massive improvement with Qdrant. pgvector is great for convenience, but it struggles with the HNSW index building and query speeds as the dimensionality increases. Qdrant’s custom storage engine, "Gridstore," is optimized specifically for large vectors. In a side-by-side benchmark I ran in early 2024, Qdrant was nearly 8x faster than pgvector for 2k-dimension vectors once the dataset passed 1 million items. The way Qdrant utilizes SIMD instructions on modern CPUs makes a huge difference in calculating those distance metrics quickly.

   Answered 2025-02-10 by Sharon Martinez

Does Qdrant provide any specific features for managing the image binary data alongside the vectors?

   Answered 2025-02-12 by Matthew Turner

• You shouldn't store the full high-res images in Qdrant, Matthew. Instead, you store the image URLs or S3 paths in the "payload" section. This keeps the database lean and ensures the highest AI model performance. However, Qdrant’s payload filtering is so fast that you can store things like image tags, dimensions, or camera metadata as JSON and use them to narrow down your search instantly. I used this for a stock photo platform last year, and it worked flawlessly for filtering millions of images by color and orientation before doing the visual similarity search.

     Commented 2025-02-14 by Jason Phillips

The gRPC support in Qdrant is also a big win for image search, as it’s much faster for sending large vector arrays than standard JSON over HTTP.

   Answered 2025-02-15 by Michelle Lewis

• Agreed. When you're dealing with 2048-float arrays, those micro-optimizations in the communication protocol really add up to a snappier user experience.

     Commented 2025-02-17 by Joshua Parker