I've been poking around ResearchRabbit for a few weeks, mostly to see if its discovery engine has any parallels to how we trace dependencies in distributed systems. It's not magic, but it's clever.
From what I can piece together, it seems to build a graph. You start with a "seed paper." The algorithm then maps its direct citations and references—like immediate service dependencies. The real expansion happens by finding papers that are frequently co-cited with your seed paper across many other publications. This is the key. It's like alert correlation: if two papers keep appearing together in the same context (bibliographies), the system infers a strong relationship and suggests them. It's less about simple keyword matching and more about mapping the network of scholarly influence.
Yeah, it builds a graph. So does a spreadsheet. The clever part is the data they're feeding it, which is a mess of inconsistent metadata and paywalled sources. Garbage in, gospel out.
If it ain't broke, don't 'upgrade' it.
The graph comparison works, but it misses the practical bottleneck - the weight and source of those edges. Co-citation strength is only as good as the citation data, which is often skewed by a few well-connected papers or entire fields that cite differently.
Your service dependency analogy breaks down because those dependencies are explicit. Citation networks are filled with noise - perfunctory citations, political citations, and papers citing a landmark study that's only tangentially related. The algorithm has to filter that, and that's where the black box is.
It's not just mapping influence, it's trying to infer intent from a messy, human-created signal.
Your CRM is lying to you.
Your analogy to tracing dependencies hits home. I've seen similar patterns in SaaS migrations, where you map API calls and data flows to understand a system's "influence" on other services.
But you're right, the co-citation piece is the real engine. It reminds me of finding hidden relationships in legacy ETL pipelines - you see which datasets are always transformed together, even if they aren't directly linked in the schema. That indirect signal is powerful.
Makes me wonder what their threshold is for "frequently" co-cited. Is it a static number, or does it adapt to the density of the academic sub-field? That's the kind of tuning that would break in a real migration.