Qwen3 late chunking

[Jacqkues]

What are the missing components for implementing late chunking with Qwen3 embedding models?

[sonam-pankaj95]

Hi, thank you for reaching out.

Can you please post any screenshots? I don't think I understood the question.,

[Jacqkues]

In the file /rust/src/embeddings/local/qwen3.rs, within the Qwen3Embedder, we only have the embed() function, but not embed_late_chunking like in the Jina or ModernBERT models.
So I think that even if I use late chunking with Qwen3, it won’t actually be real late chunking.
How can we fix this?
Is it possible to copy and paste the embed_late_chunking function from another model?

[akshayballal95]

Hey @Jacqkues. We can add it easily. Please let me know if you'd like to raise a PR for this. I can then assign this issue to you.

[Jacqkues]

I have added it in a fork on my profile , can I open a PR for this ?

[akshayballal95]

Yes sure. Go ahead

[Jacqkues]

After some testing, I found that late chunking on Qwen3 does not improve embedding performance. I’m wondering whether my implementation is correct, or if the fact that Qwen3 uses the last-token pooling technique makes late chunking irrelevant, what do you think about it ?

[akshayballal95]

Ohh yes. Now I remember, that was the reason I hadn't added late chunking to Qwen3 model. Late chunking works only with mean pooling based models. I wonder if there is as way to make it work well with CLS token and last token embedding models.

[Jacqkues]

Thank you , yes I think it's not possible to do it right now . I have learned something at least :) , I will close my PR