The phrase on the homepage is a technical claim
conscRAG’s tagline is “Graph-grounded research ideas.” Grounded, here, is not a marketing synonym for “smart.” It means the generator is only allowed to write ideas that cite triples from a graph built in the same run. The graph is extracted from your target plus OpenAlex retrieved works. No triple, no legal citation.
That is different from asking a chat model “give me four research ideas about this abstract.” An unconstrained model will invent citations, blend two papers into a third that does not exist, or propose a method that never appeared in the context. Graph grounding is a restriction: recombine evidence we already extracted.
What the graph contains
Nodes are short scientific phrases (a problem, a method, a dataset, a limitation). Edges are one of five relations: addresses, uses, improves, evaluated_on, limited_by. Each edge remembers the source title.
Checked twice
The drawing you see is Figure 1, target-centered. The same structure is sent back to the model as JSON when it plans directions and writes findings. Grounding is checked again after generation. If the model returns an evidence triple that is not in the graph, conscRAG drops it. Ideas that lose too much evidence are backfilled from real triples that overlap the idea text. You still have to read; the software only refuses fake edges.
What graph-grounded does not mean
It does not mean the idea is true, new, or safe to run. It does not mean we retrieved every relevant paper. It does not mean the PDF was parsed. Abstracts omit the exact architecture, the hyperparameter that matters, and the negative result in section 5. Peer review and your own reading still sit above the graph.
It also does not mean a citation graph (who cited whom). conscRAG’s graph is a knowledge graph of extracted claims. Citation links appear only as the one-hop reference expansion that adds more abstracts to the corpus, not as edges in Figure 1.
Why we bother with a graph at all
Retrieval-augmented generation that pastes raw abstracts into a prompt still lets the model ignore them. A graph forces a smaller, typed language: this method uses that mechanism; this approach is limited_by that benchmark. Four strategies (exploratory recombination, mechanism transfer, guided improvement, component reconfiguration) are just different ways to walk those edges.
If you drop a paper on the homepage and read a finding, ask one question: can I click the evidence and land on a real retrieved work? If yes, the idea is graph-grounded in the sense we mean. If you cannot, something failed, and you should not trust the paragraph.
Questions
How is this different from chatting with an unconstrained model?
An unconstrained model can invent citations. conscRAG only allows ideas that cite triples already extracted from your target plus OpenAlex retrieved works.
How many relations can an edge use?
Exactly the five allowed relations: addresses, uses, improves, evaluated_on, limited_by.