What problem does it solve? Finding relevant papers and web sources for a research topic requires manually searching multiple academic databases, deduplicating results against what you already have, and judging relevance — a slow, error-prone process. This Skill automates candidate discovery, deduplication, and ranking, then stops at a human-reviewable shortlist before anything is ingested. ## Core Features & Use Cases - Multi-source discovery: Fetches candidates from arXiv, Semantic Scholar, OpenAlex, Wikipedia, and DeepXiv using topic, anchor-paper, or wiki-context seed modes. - Deduplication and ranking: Excludes already-ingested sources via external identifiers (DOI, arXiv, S2, URL), scores candidates with discover.py, and adds human-readable rationales and risk notes. - Purpose-aligned shortlist: Judges each candidate against the project's stated purpose, classifies items as ready/maybe/skip, and checkpoints the shortlist for later ingestion via /lumi-ingest. - Use Case: A researcher starting a literature review on retrieval-augmented generation runs the skill with a topic query, receives 5-10 ranked, deduplicated candidates with provenance estimates, and picks which ones to ingest. ## Quick Start Ask the assistant to discover candidate sources for your research topic, for example: "Find and rank recent papers about retrieval-augmented generation for my wiki."
