How conscRAG retrieved works and findings connect

A conscRAG finding is a short, model-generated research idea: a title, a problem, motivation, method, experiment sketch, expected contribution, and evidence. The model is told not to invent papers. Evidence must be triples that already exist in the graph, and those triples were extracted from the retrieved works plus your target abstract.

That is the connection. Figure 2 supplies text. Extraction turns text into triples. Idea generation recombines those triples into ranked findings. If a work never entered the corpus, it cannot appear as sourceTitle on a finding. If extraction never built a triple, the finding cannot cite it.

conscRAG plans directions in each graph context, then synthesizes candidates. The default four prefer one of each strategy: exploratory recombination, mechanism transfer (reuse a neighbor method on the target problem), guided improvement (keep the target approach, fix a weakness), and component reconfiguration (rearrange known pieces).

Each idea is asked to copy two to four evidence triples from the graph. After the model returns, conscRAG grounds those triples against the real graph. If the model invented an edge, it is dropped. If too few edges survive, we backfill from triples that overlap the idea’s wording. The finding you see is therefore constrained twice: once in the prompt, once in code.

On the Findings panel, evidence is shown as a short triple and a source title. Click it and Figure 1 highlights that edge. The source title should match a row in Retrieved works (or the target itself). That click path is how you audit a suggestion: which paper, which relation, which concept.

This is also why findings are labeled model-generated, not peer-reviewed. The citation is to an extracted triple, not to a page number in a PDF, and not to a human reviewer who checked novelty. Read why peer review still matters before you treat a finding as a result.

A missing abstract on a retrieved work means no triples from that paper. A vague target abstract yields a vague profile, so retrieval drifts and findings wander. Rate limits and model errors can stop the run before ideas are written. Guest and signed-in saving stores the same findings you saw; it does not re-fetch OpenAlex when you open a share link.

Used well, the chain is: drop a paper → retrieve works → extract graph contexts → write ranked cited drafts. Used poorly, it is a language model with a bibliography costume. The retrieved-work list is there so you can tell the difference.