Find answers from the community

Updated 3 months ago

late-chunking/chunked_pooling at main · ...

Hi, just a random feature request. Is it possible to create a an embeddings class for jina.ai to leverage their new "late embeddings" approach via HF? Here is the sample code which still has a bunch of custom methods. https://github.com/jina-ai/late-chunking/tree/main/chunked_pooling

blog: https://jina.ai/news/late-chunking-in-long-context-embedding-models/

4 comments

LLogan M

I think this could be implemented as a node parser/text splitter?

But also, the accuracy improvements don't entirely convince me that its worth the effort to personally contribute this myself 😅

ffarzzy528

#shotsFired

LLogan M

🌶️

LLogan M

But for real, looking at the code here, most of it could be copy-pasted into a node-parser integration package implementing the BaseNodeParser class
https://github.com/jina-ai/late-chunking/blob/main/chunked_pooling/chunking.py

I see it actually uses our semantic splitter too 👀

Add a reply