To be fair, documents context with things like semantic search could be the upcoming bottleneck for a lot of industry applications.
Even though a hundred+ QA products doing this (Chat with your files vibes etc) haven't had much success, they are ignoring the fact that trust is key to document storage, and that we haven't mastered correct context/snippet amalgamation yet - for example, one may want to include both statements that have high cosine similarity, and also statements with clashing natural language inference similarity (contractionary sentences). This gives the context window the statements needed to correctly reason and use logic from different parts of the document. Small things like order (e.g. grouping snippets from different documents together in per-document batches, and also ordering by page number the statements within each document) matter greatly.
Even though a hundred+ QA products doing this (Chat with your files vibes etc) haven't had much success, they are ignoring the fact that trust is key to document storage, and that we haven't mastered correct context/snippet amalgamation yet - for example, one may want to include both statements that have high cosine similarity, and also statements with clashing natural language inference similarity (contractionary sentences). This gives the context window the statements needed to correctly reason and use logic from different parts of the document. Small things like order (e.g. grouping snippets from different documents together in per-document batches, and also ordering by page number the statements within each document) matter greatly.