Which repository is this?

When you send repoherald a post, something has to work out which repository it is about. That used to be a person reading the caption and searching GitHub by hand, which is why the catalogue could never grow faster than someone’s evening. This is that step, running on NVIDIA Nemotron via Nebius Token Factory.

The model only proposes names. GitHub decides. Every candidate is checked against the real API, and one that doesn’t resolve is discarded however confident the model sounded — so this can tell you it doesn’t know, but it can’t invent a repository that was never there.

24.6%

identified automatically before — a regex over literal github.com links, and nothing at all from OCR or speech

89.9%

with Nemotron, over the same 228 real cases — one wrong answer, traceable to a stale name in our own catalogue

1.03

model calls per post on average, because a quarter of posts carry a literal link and cost nothing at all

The five rungs, cheapest first — and which model does what
  1. A literal link. A regex. No model, no cost, and the only path that earns full confidence.
  2. Nemotron 3 Super (120B), fast tier. Reads the caption, the OCR and the transcript together. Most posts stop here.
  3. Nemotron 3 Ultra (550B), deep tier. Runs only on candidates GitHub has already rejected, and is told which ones, so it repairs rather than repeating itself. This is the rung that undoes spoken punctuation.
  4. Tavily web search. Only when the post named nothing at all. Two searches, then Nemotron ranks the results — taking the first link would be picking by result order.
  5. Nemotron verifies.The one rung that can say no: it compares the post against the repository’s real GitHub description and vetoes a near-miss. We added it because search alone was wrong a third of the time and never once abstained.

A post with a link answers instantly. A spoken name takes about two seconds. A post that names nothing can take twenty — it is four model calls and two searches, and we would rather it be slow than wrong.

Four cases the real pipeline has to survive. Load one, or paste your own text into the boxes below.

Highest fidelity — often a literal link.

Near-verbatim, with mangled characters.

Lossiest — punctuation is spoken, hyphens are not.

Text you submit here is sent to Nebius for inference and is not stored by repoherald. The demo is rate-limited so a finite credit balance survives being public.