A list of papers is not a map of ideas
Related-work search usually returns a ranked list: titles, years, venues. That is useful for citation, and it is what conscRAG shows as Figure 2, Retrieved works. A list does not show how two papers share a mechanism, or that three of them are limited by the same dataset. Those links live in the text. A target-centered graph is a way to draw them.
conscRAG’s Figure 1 puts the target paper at the center. Neighbor concepts sit on an inner ring. More distant concepts sit outside. Edges are typed: addresses, uses, improves, evaluated_on, limited_by. Color is the relation, not the venue. You are looking at claims extracted from titles and abstracts, not at a citation network of who cited whom.
The center is the paper you dropped
“Target-centered” means every other node is placed relative to your paper. The target is not just another hit in a search ranking. Retrieval, triple extraction, and idea generation all start from a profile of that paper (purpose, mechanism, method, keywords). The graph is the visual form of that profile plus the neighbors OpenAlex returned.
This is different from a force-directed blob of the whole field. In a field-wide graph your paper can disappear into a cluster. In a target-centered view the question is always: what is next to this work, and by which relation? That is the question you ask when you are trying to invent a follow-up, not when you are surveying a discipline.
Rings separate close neighbors from distant analogies
The inner ring is meant for concepts that share language and role with the target: the same problem class, a method the target already uses, a limitation named in nearby abstracts. The outer ring is for bridges and farther nodes.
Why two rings
conscRAG later uses farther nodes for exploratory recombination: take a problem/method pair from a more distant work and map it onto the target. Without rings, every edge looks equally important. With rings, you can see that “uses transformer encoder” is a close neighbor while “evaluated_on remote-sensing benchmark” might sit farther out. That geometry is a reading aid. It does not prove that the outer node is less true; it only marks distance in the extracted graph.
Typed edges beat a word cloud
A word cloud of keywords collapses “uses attention” and “limited_by attention span” into the same token. A typed edge keeps the verb. When you click evidence on a finding, conscRAG highlights that head–relation–tail in Figure 1. You can see whether the idea borrowed a method, attacked a limitation, or transferred a mechanism.
That is why a target-centered graph is useful for related concepts: it shows concepts as nodes and roles as edges, anchored on the paper you care about. It will not replace reading. Abstracts miss methods that only appear in the PDF. Extraction can drop or merge phrases. Use the graph to decide what to read next, not as a substitute for the papers themselves.
Questions
Is this a citation network?
No. Figure 1 is a knowledge graph of extracted claims from titles and abstracts. Who-cited-whom appears only as a one-hop expansion that adds more abstracts to Figure 2.
How many relation types are there?
Five: addresses, uses, improves, evaluated_on, and limited_by. Color in Figure 1 is the relation, not the venue.