What conscRAG added to Run

The old MVP searched OpenAlex once by title, hopped a handful of references, kept one bag of triples, and wrote four ideas. That is still the same product job — start from a paper, get cited drafts — but the machinery underneath is new.

What landed: several retrieval queries from the target profile; an LLM ranker that picks twelve base papers; a top-three reference hop per base; up to four graph-derived contexts; two-stage idea generation (plan a direction, then write the card); richer findings (motivation and an experiment sketch); live step status instead of a fake timer; context chips on Figure 1; query and reference provenance on Figure 2; and an optional open-PDF checkbox that stays off unless you turn it on.

After the profile (purpose, mechanism, method, keywords), the model writes four to six short search queries. OpenAlex is called per query. Duplicates drop. If OpenAlex returns 429, we wait on Retry-After and try again. A ranker then keeps twelve bases that match the profile, not the first twelve keyword hits.

Each base paper contributes up to three references, scored by overlap, year, abstract present, and citation count. Figure 2 now says which query found a work, or “reference of” which base. For the older one-search story, see how the graph and retrieved works connect.

Triples still use the five relations: addresses, uses, improves, evaluated_on, limited_by. Instead of one 40-triple bag, conscRAG builds up to four compact contexts. Edges are scored against the profile (OpenAI embeddings when that key is set; otherwise word overlap). Figure 1 can show All or one context.

Each context is paired with four strategies. The model plans two directions per strategy, then writes one idea per direction. Invented evidence is still dropped after generation. The default four prefer one idea per strategy. Show all is for the rest.

We did not add an automatic novelty judge that searches the literature again for every idea. We did not add a second index, user PDF upload, or paywall scraping. Findings stay labeled model-generated, not peer-reviewed. What we removed is the other half of this changelog.