Representing academic literature in a multilayer graph

[soelderer]

This is highly related to #2000 and #2430.

LightRAG has a high potential for scholarly work, e.g., comparing and contrasting definitions, identifying jingle-jangle-fallacies and research gaps. Here, the content is not the source of truth, but represents a discourse between authors and documents.
The documents reference each other, which means that right now, I end up with nodes representing the publications, authors, journals, as well as the content. Documents might contradict each other frequently, and these contradictions are important. Moreover, we have a lot of metadata, such as bibliographic information and co-citations.

In this very interesting preprint they leveraged on author similarity, reference similarity and semantic similarity to map research landscapes. A possible translation to LightRAG could be a multilayer graph with interlayer links:

• We keep the content KG as-is, it becomes the semantic layer
• We add a document layer. Nodes represent indexed documents and mentioned documents that are not (yet) indexed. Documents can hold metadata, like bibliographic information. They have a (directed?) relation if they mention/cite other documents. This way, we can find co-citation networks and related publications. Nodes have interlayer links to all chunks on the content KG
• We add an author layer. Nodes represent the authors of documents. They have a relation if they are co-authors of a document. This way, we can find people collaborating frequently. They have interlayer links to all the documents they co-authored.

I think this has several benefits:

• You could ask queries like "What authors to approach for collaboration on XY? What are the most relevant publications that I haven't yet indexed?" This is quite powerful.
• It might be backwards compatible if we allow for a single-layer graph
• It's scalable and flexible: for example, not everybody might need an author layer, but someone else might want an organization layer

I thought I'd use this thread for discussing ideas. Totally get it that there are more pressing issues, and this seems like a special case. But I might end up implementing parts of it, so I want it to be potentially useful for upstream.