discover

Run discovery in anchor mode with one or more arXiv IDs. The tool queries Semantic Scholar's recommendations, references, and citations endpoints per anchor, then ranks and deduplicates candidates into a shortlist with per-paper rationales.

Use wiki mode with the --from-wiki flag. It selects the most recently modified paper pages in wiki/papers/, extracts their arXiv IDs, and uses them as anchors so recommendations track what you have been working on lately.

No. Discovery only proposes a shortlist and writes a checkpoint file plus a log line; it never modifies wiki pages or downloads papers. You choose a candidate and run the ingest step separately.

Candidates are deduplicated against the arxiv or arxiv_id frontmatter of existing wiki/papers/ pages before ranking. A high dedup count means the wiki already covers that neighborhood well, and the report surfaces this count explicitly.

In topic mode the tool falls back to DeepXiv search and notes the degradation in the report. If every channel fails, the run aborts with an error rather than emitting an empty shortlist as if it were a real recommendation.

Use anchor mode whenever you can name specific papers, since Semantic Scholar's recommendation model is a stronger signal than keyword search. Use topic mode only for open-ended exploration of a subject area with no concrete reference paper.