When a research field has few papers, is it better to match by meaning or to link several ideas at once by hand?

This explores whether, when a literature is sparse, it works better to rely on what the papers say (content embeddings) or on hand-built structures that connect several items at once (hypergraphs). The corpus has no head-to-head test of these two on sparse literature, so it can't name a winner. What it does show is where each approach breaks, and that turns out to matter more than an overall score.

Start with the limits of embeddings. One result proves that any fixed embedding size can only return a limited number of distinct 'top-k' combinations of documents. Even embeddings tuned directly on the test data hit this ceiling on simple tasks Do embedding dimensions fundamentally limit retrievable document combinations?. Embeddings are also good at broad similarity and weak at exact details. A search agent that runs plain text commands over the raw corpus beat dense embeddings on queries that hinge on specific entities, because embeddings tend to blur similar-sounding names together Can direct corpus search beat embedding-based retrieval?. Embeddings do carry real structure, though. Their main dimensions sort concepts from broad categories down to fine ones, in a way that mirrors a human-built taxonomy (WordNet) Do embedding eigenvectors organize taxonomy from coarse to fine?. So embeddings partly rediscover hierarchies people would otherwise draw by hand. But that structure comes from word co-occurrence statistics, so it may be weaker for niche vocabulary.

Now the case for explicit structure. Graph databases beat vector search when a question is about relationships, such as 'which X connects to Y through Z?' or 'list all of them'. Following a graph is exact and complete, while similarity search is only probabilistic When do graph databases outperform vector embeddings for retrieval?. Hypergraphs go a step further. Ordinary graph edges link two things, but a hyperedge can bind three or more at once, so joint constraints survive across several retrieval steps instead of being broken into pairs Can hypergraphs capture multi-hop reasoning better than graphs?. Hierarchical knowledge graphs also answer big-picture, cross-document questions that flat chunk retrieval cannot reach Can multimodal knowledge graphs answer questions that flat retrieval cannot?. One caveat: these papers build their graphs automatically, not by hand.

What the corpus suggests is that 'which performs better' is the wrong frame. Each method fails in a different way. Embeddings find papers that use similar words, but they miss exact relationships and blur entities together, which hurts most in a small field where one specific connection may be what matters. Human hypergraphs capture that kind of connection exactly, but only where someone has already drawn it, and in a sparse literature most links haven't been drawn yet. The practical takeaway is to use them together: embeddings to propose possible neighbors, and explicit structure to confirm and connect them. The text-search result hints that plain keyword matching may also deserve a place Can direct corpus search beat embedding-based retrieval?.

One caution about what's missing. None of these papers studies sparse literatures directly, and none studies hypergraphs that people curate themselves. If that is your actual question, this collection only gives you neighboring evidence.