Reduce note context selection

[luisherrmann]

Problem

• Currently, all notes are being included in the note context.
• In a densely annotated paper, this could become a problem, as the full note context will approximate (or exceed) the full text context.

Example

/

Solution

• Since notes are essentially highlighted text with optional annotations, they should get the same treatment as paragraphs.
• Retrieve $top_k$ notes by cosine similarity of the query embedding with the highlighted text + note embedding (full score through RRF).
• For free-floating sticker notes, retrieve the closest paragraph(s) by location in the paper.
• Either provide the top notes in a separate context as is done now, or integrate them directly into the paragraph context, and modify the system prompt to put special attention on sentences highlighted (for example in the following way):

<this is a sentence>[[with highlighted, this is an example note text]] content.
...

Open questions:

1. For retrieval of $top_k$ notes, use the highlighted text embeddings, or map highlighted text back to paragraphs in the text and use these embeddings' similarities? Mapping back to text paragraphs can be done without computing new embeddings through a search for exact text matches in the text search index, reducing cost and almost certainly increasing retrieval quality.