Keeping up with a research field is not a reading problem, it is a filtering problem. The volume of new preprints in any active area is far past what a person can skim, and the standard tool for it, a keyword alert from one of the big indexes, is tuned for recall rather than relevance. You end up with a mailbox full of papers that share your vocabulary and none of your problem.
Why Keyword Alerts Miss The Right Papers
A keyword alert matches strings. A research interest is not a string, it is a shape: a method you care about, the problems it applies to, and the neighbouring fields where the same trick shows up under a different name.
That mismatch cuts both ways. A paper from an adjacent field using your exact method with different terminology never fires the alert. A paper that happens to use your three keywords in an unrelated context fires every week. After a month of that most people stop opening the digest, which is the actual failure mode, not the false positives themselves.
The Four Step Pipeline
Literature Radar is a free self-hosted tool that splits the job into four steps.
First you describe your interests in plain language. Not a keyword list, a short paragraph written the way you would explain your work to a colleague. Alongside it you add topic queries that tell the poller what to pull from each source.
Second, a scheduled job polls the open indexes on your schedule and writes new papers into a local SQLite database. Anything already seen is skipped, so the store never duplicates and stays cheap to query.
Third, your model reads each title and abstract and scores it from 0 to 10 against your interest profile, attaching a one line reason. A paper about your exact method in a neighbouring field can score high, while a paper that merely shares your keywords scores low. This is the step a string match cannot do, and it is the reason the ranking layer earns its place.
Fourth, on your chosen schedule you get one email, best first, each paper with its score and the line explaining why it surfaced. Nothing repeats across digests, and everything stays browsable in the admin area afterwards.
Bringing Your Own Model
The scoring step is model agnostic on purpose. You point it at whatever you already run, a hosted API or something local, and the ranking quality tracks the model you picked rather than a vendor default you cannot change.
For this workload that matters less than you might think. Scoring one abstract is a short classification with a written rubric, not a reasoning marathon, so a small local model handles it well and the marginal cost per digest lands at zero. If you already run something on the machine that polls, you are done.
Who This Is Actually For
If you read a couple of papers a month, a plain keyword alert is fine and this is overkill. It earns its keep when the field moves fast enough that missing the right preprint by six weeks has a real cost: graduate students tracking a thesis area, lab leads watching several subfields at once, engineers following a method that is still changing under them.
The setup is a database file, a cron entry and an interest paragraph, so the cost of trying it is an afternoon.
The Pattern Generalizes
Strip the papers out and the shape is reusable: poll a firehose into local storage, have a model rank each item against a written statement of what you actually care about, then deliver a short ranked digest with reasons attached. That works for any feed you are drowning in, and the reasons are what make it trustworthy enough to keep reading. The other free AI science tools on the site are built on the same idea.
Top comments (0)