Hi, it seems to me the idea in the optional project is different from what Kacper said in the videos.
The project asked us to implement Scalar + HNSW without mentioning something like MUVERA to reduce the embeddings to one vector per document, but Kacper reiterated in the videos that HNSW doesn’t work with multi-vector and that’s why he keeps hnsw_config=models.HnswConfigDiff(m=0) in the code.
Maybe I miss something here?
Hi Pang.luo, Qdrant team member here.
No contradiction, you’re not missing anything.
Kacper’s right: HNSW can’t build a search graph directly on multi-vectors because MaxSim isn’t symmetric. That’s why m=0 is used. With m=0, it’s just an exact brute-force scan.
The project lists optional improvements, not things you have to combine. Scalar quantization simply makes that brute-force scan smaller and faster. It doesn’t make HNSW work on multi-vectors.
If you want actual HNSW graph search, that’s what MUVERA is for. It compresses each document into a single vector HNSW can index, then reranks the results with MaxSim.
So m=0 + scalar quantization is a valid setup. It’s just brute-force, not HNSW graph search.
Let me know if you have any other questions. 
That makes sense. Thank you for the reply Dylan.