How a conscRAG run changed

You still start from Find a paper or a pasted title and abstract. SciMON is still the example because it is a well-known idea-generation paper, not because it is our method. After Run, the page is a notebook: Figure 1, Figure 2, Findings.

What changed is everything between the button and those three panels. The old hallway was profile, title search, hop, one extract, four ideas. The new one is profile, queries, retrieve, graph, contexts, generate. How we process and save a run has the storage story; this post is the live path.

The result header shows purpose, mechanism, and method, not only keywords. Figure 2 lists the queries. Each retrieved work says search-plus-query or reference-of-base, plus year, plus abstract or open PDF. Figure 1 has All plus Context 1–4. Clicking a finding’s evidence still lights an edge and can jump you to that context.

A finding is a fuller card: problem, motivation, method, experiment, contribution, and two to four grounded triples. Strategy and context sit in the kicker line. We still refuse to invent a novelty score.

Several distinct queries, not the title repeated six times. About twelve bases and a few dozen hops, not three near-duplicate titles. Two to four contexts you can flip without the graph collapsing to one node. Four default findings that cite real Figure 2 titles. Extra candidates if you ask.

A thin run still happens: missing abstracts, OpenAlex 429s, a vague target. The honest read is the same as before. Look at Figure 2 first. If the corpus is weak, the findings will be too. We did not add a badge that says the graph is complete.