# How Discover finds papers worth reading today

- Published: 2026-07-30
- Authors: CORTEXA
- Category: Engineering
- HTML: https://researchhub-vert.vercel.app/blog/how-discover-finds-trending-papers

A transparent look at the daily signal behind CORTEXA’s default paper feed—and why it is refreshed ahead of time instead of on every page load.

A blank research search box assumes you already know what to ask. Discover now
starts somewhere more useful: a short list of papers receiving attention today.

## The source

The primary signal is [Hugging Face Daily Papers](https://huggingface.co/papers),
whose official API exposes a trending order alongside publication dates, authors,
and arXiv identifiers. The paper itself remains the source of truth: every item
links to its arXiv abstract and PDF.

“Trending” is not the same as “best.” It means a paper is receiving unusual
attention now. That makes the feed a starting point for discovery, not a claim
about scientific quality.

## Why the feed is stored

CORTEXA refreshes the list on a schedule and stores the latest verified
snapshot in Supabase. Discover reads that snapshot immediately, so opening the
page does not wait for another service. If a scheduled refresh fails, the last
verified snapshot remains available; if storage has not been configured yet, a
cached live request keeps the page useful.

The result is intentionally simple: open Discover and see current papers before
typing anything, then search when you have a specific topic, method, or author in
mind.

## Why precomputed rather than live

Fetching a trending list on every page load would be simpler to write and worse in three ways.

It puts an external dependency in the critical path, so an upstream outage becomes your outage. It multiplies identical requests — every visitor asking the same question within the same hour. And it makes the page as slow as the slowest upstream response.

Precomputing moves all three problems to a scheduled job where failure is recoverable and latency does not matter.

## The staleness trade

The cost is that the feed can be out of date. The mitigation is a freshness threshold: serve the stored snapshot while it is within one refresh interval, and fall back to a live fetch past that.

Getting that threshold wrong is a subtle bug. Set it shorter than the refresh interval and you guarantee a window every day where every request falls through to the live path — the cache stops being a cache exactly when traffic is highest.

The correct value is derived from the schedule, not chosen independently.

## What the ranking is and is not

The signal is community attention, not quality. A paper trends because people are discussing it, which correlates with importance and also with controversy, timing and the author's reach.

Treat it as a prior over what to look at first, never as a judgement. The most useful paper for your specific problem is frequently one nobody is discussing.

## Failure behaviour

If the upstream source is unavailable and the snapshot is stale, the honest response is to say so rather than silently serve old data. A feed that quietly shows last week's papers as today's is worse than one that admits it is degraded.