[HN Gopher] Late Chunking: Contextual Chunk Embeddings Using Lon...
___________________________________________________________________
Late Chunking: Contextual Chunk Embeddings Using Long-Context
Embedding Models
Author : mfiguiere
Score : 14 points
Date : 2024-09-20 02:59 UTC (20 hours ago)
(HTM) web link (arxiv.org)
(TXT) w3m dump (arxiv.org)
| will-burner wrote:
| This is an interesting development. A couple blog posts exposing
| on late chunking
|
| https://weaviate.io/blog/late-chunking
|
| https://jina.ai/news/late-chunking-in-long-context-embedding...
|
| The late chunking idea originates from ColBERT, an embedding
| technique from 2020 https://arxiv.org/pdf/2004.12832
| leobg wrote:
| Another (simpler?) approach is to also split your chunks into
| sentences. So you'll end up with chunk embeddings and sentence
| embeddings. Now you can do sentence level search. And also
| distill chunks down to their most relevant sentences at query
| time before you dump 'em into your LLM's context window. If you
| use Sentence Transformers, you get your chunk embeddings for
| free, because they are just the np.mean of the embeddings of all
| the sentences in that chunk.
___________________________________________________________________
(page generated 2024-09-20 23:02 UTC)