Ask HN: Aggregate chunk embeddings to create intermediate-level representations

TL;DR What are the current best practices for semantic search over a set of chunked documents? And, in particular, what is the most economic approach today (best "bang for the buck," if you will), taking into account the commercial or open-source tools that are currently available? Or... am I thinking about this all wrong and there's a better way to do it?

Details:

I have a set of documents against which I would like to run a semantic search. I'm thinking of using langchain to chunk the documents, generate embeddings for each chunk, and then go from there.

The difficulty, of course, is that concepts in the documents may span chunks, including non-contiguous chunks, and so we need more than just the chunk embeddings for this search to work.

From research, I'm aware of the following methods for aggregating the chunk embeddings to create a document-level representation: average pooling, max-pooling and weighted average. I'm aware that there are advantages and drawbacks to each of these approaches.

I'm also aware, from a quick scan of the literature, of techniques such as Vector of Locally-Aggregated Word Embeddings (VLAWE)[1].

I'm also open to building a solution, so it's fine if the answer is "Look at what they do in [some paper], it shouldn't be too hard to implement that."

[1]: https://aclanthology.org/N19-1033/