How to fix vector search for long documents

A common approach when building vector search for long documents is to divide documents into small text chunks of around 200 words and embed these chunks into vectors. This leads to a problem. Each vector contains information about its local context, but does not take global context of the full document into account.

We at midpage are building a genAI research and drafting platform for US lawyers and are working on semantic search for long legal documents. We would like to share our ideas and findings.