Replies: 2 comments
-
|
hey @Greatz08, qdrant is agnostic to the choice of models. If on your benchmarks bigger models doesn't give much precision boost, you can stick to it. If you don't have enough data to evaluate each model, you can try to lookup model performance on public benchmarks |
Beta Was this translation helpful? Give feedback.
-
|
Great question — and welcome to the Temple of Too Many Dimensions™ So here’s the thing: Higher dimensional embeddings don’t magically mean better results. If your use case (markdown notes, as you said) has low entropy or semantic diversity, adding 1024D to a 384D job is like bringing a rocket launcher to a sandwich party. Most retrieval problems suffer more from bad preprocessing or weak chunking than "not enough embedding width". Bigger isn’t better unless: Your downstream search requires ultra-fine-grained disambiguation (like: differentiating "neural pruning" from "neural gardening"). Your documents have very rich internal structure or compositionality. You have enough samples to justify the added sparsity and compute. Oh — and if you're using qwen3, note that different quantizations can shift the relative weight of token influence. So sometimes it’s not the dimension count, but how it’s packed and projected. TL;DR: — |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
I recently downloaded qwen3 embedding model 0.6 B, 4B quantized Q4 and Qwen3-Embedding-8B:Q5_K_M . These three embeddings model are latest from qwen team and they generate embeddings in different dimensions. I wanted to develop better vector database for my markdown notes and i choose qdrant as vector db. Performance difference between all these three embeddings model were almost similar. My question to you is that - does it matter if i use bigger embedding model which support more dimensions ?
Beta Was this translation helpful? Give feedback.
All reactions