How Research Radar works

Research Radar is a small open-source tool for keeping up with new research. Keyword alerts return many near-misses and still miss relevant work. This tool instead reads each new abstract against a description of your research that you write and can edit. This page describes what runs each morning.

Stages 1 and 2 are retrieval. Each morning the pipeline pulls the newest publications from OpenAlex (which indexes most scholarly journals) and the arXiv, SocArXiv and PsyArXiv preprint servers, searching the research areas derived from users' seed papers. Errata, book reviews and other non-articles are filtered out. Each candidate's title and abstract is converted to an embedding, a numerical representation of its meaning, and compared to the embeddings of your seed papers. The closest candidates form your daily shortlist. This stage is deliberately generous: its job is to make sure relevant papers reach the judge, not to decide.

Stage 3 is where selection happens. A large language model reads each shortlisted title and abstract against your Selection Criteria, the description, topics and exclusions you wrote, and scores the paper's fit from 0 to 10 with a one-sentence rationale shown on every card. Your briefing lists what clears your fit threshold, ranked by fit score, with embedding closeness to your seeds breaking ties. Thumbs-up and thumbs-down votes become boundary examples the judge reads on every following run.

The full methods document, with each design and filtering decision and the exact thresholds and formulas, is docs/how-it-works.md in the GitHub repository. The code is open source under the GNU AGPL-3.0.