How conscRAG processes a run, and whether we save it

Run sends your title, optional DOI, and abstract to /api/generate. If you left title or abstract empty but provided a DOI, we resolve that DOI on OpenAlex first. Then the pipeline profiles the paper (purpose, mechanism, method, keywords), writes search queries, retrieves and ranks OpenAlex works, hops the top three references of each base paper, extracts triples, builds graph-derived contexts, draws the target-centered graph, and writes findings.

Free use is limited to three generate runs per hour per IP. Pro, when offered, lifts that cap. Find a paper and opening an old link do not use the cap. A run can take several minutes. Status text on the page walks profile → queries → retrieve → graph → contexts → generate. An optional checkbox can read OpenAlex open-access PDFs; it is off by default.

If you continued with Google, the run is saved to your account (Postgres on Neon) and listed under Your runs. You can open it later. We store the title, DOI, abstract, graph, and findings, and we attach your Clerk user id. You can ask us to delete account data at veeconsc@gmail.com.

If you are a guest, we still try to keep a shareable copy so you can copy a link. That copy lives in a short-id key-value store and expires after 30 days. Guest links can disappear earlier if we have to. Signed-in saves are the durable path.

Your browser also keeps the last run in localStorage under a conscRAG key so a refresh does not wipe the page. That copy never leaves your machine unless you clear the site data. We do not treat Find-a-paper queries as runs; typing in Find does not write Your runs or a share link.

The saved payload includes your target title, DOI, and abstract, plus the generated result. For share size we clear abstracts on retrieved papers and keep the target abstract. Figure 2 on a reopened link may therefore show titles without the neighbor abstracts you saw live. The graph and findings remain.

We also see your IP for the hourly cap, and Clerk sees your Google name and email if you sign in. Processors are listed in the Privacy Policy: Vercel, Clerk, Neon, Stripe (if you ever use billing), DeepSeek or OpenAI for the model, OpenAlex for metadata, and Upstash Redis for guest links and rate limits. We do not ask you for a PDF upload.

We do not keep a private copy of the entire OpenAlex index. We do not store every Find query as a document. We do not train a public model on your pasted abstracts as part of this product loop. Rate-limit counters expire in about an hour.

If a generate call fails, there may be nothing to save. If guest storage is down, you can still see the result in the tab; you just will not get a /r/… link. For the legal wording, read Terms and Privacy. For what to do with a successful run, start with where findings can help.