How conscRAG’s target-centered graph and retrieved works connect

After you press Run, conscRAG shows a target-centered graph (Figure 1) and a numbered list of retrieved works (Figure 2). They are not two independent searches. The list is the corpus. The graph is built from triples extracted from that corpus. If a paper never enters Figure 2, its abstract cannot contribute an edge in Figure 1.

The pipeline order is fixed: profile the target, write several search queries, retrieve and rank OpenAlex works, expand the top three references of each base paper, extract triples, build up to four graph-derived contexts, then draw nodes and edges. Findings come last, and they may only cite triples that already exist in that graph.

The first paper in the corpus is your target (the title and abstract you dropped, or resolved from a DOI). Next come OpenAlex hits for several profile-derived queries, ranked by an LLM against that profile. Then come the top three referenced works of each base paper: one hop, not a crawl of the whole citation graph.

Search hits are marked via search and show which query found them. Hop papers are marked via reference of the base title. Duplicate titles are dropped. Abstracts shorter than a usable threshold never become generate-time papers, because there is nothing to extract. An optional Run checkbox can read OpenAlex open-access PDFs for the target and base papers; it is off by default.

For each paper in the corpus, a language model is asked for a few short triples: head, allowed relation, tail. Allowed relations are only addresses, uses, improves, evaluated_on, and limited_by. Heads and tails are meant to be scientific components (problem, method, mechanism, finding, dataset, limitation), not author names. conscRAG then builds up to four compact graph-derived contexts from those triples, scored against the target profile.

toGraph turns those triples into nodes and edges. The target sits at the center; other concepts are placed on rings. So Figure 1 is a projection of Figure 2. Click a finding’s evidence and you highlight an edge whose sourceTitle is one of those retrieved works. That is the mechanical link between the two panels.

It is not a full citation map of every paper that cited your target. We do not walk the OpenAlex cited_by graph. We do not score PageRank. A highly cited neighbor can be missing if it did not appear in the profile queries or in the top-three reference hop we fetched.

It is also not a guarantee that every retrieved abstract produced a triple. Extraction can fail on a chunk, or return too little, in which case we retry smaller batches. Empty or noisy abstracts yield thin graphs. If Figure 1 looks sparse, look at Figure 2: the usual cause is a thin corpus, not a drawing bug. For why we start from OpenAlex at all, see why conscRAG uses OpenAlex.