Papers: semantic search over 725K ML papers and their code
A papers-with-code search engine rebuilt in Go: gobed vector search over every abstract, method, and dataset, a live arXiv crawler, an embedded PDF reader, and figure-aware social cards rendered natively — all in one fast binary at papers.app.nz.
- papers with abstracts
- 725K+
- paper ↔ code links
- 613K+
- datasets
- 15.5K
- methods
- 9.5K
- eval results
- 59K+
the product
Search, read, and trace papers to code
figure-aware cards
Every paper ships with a generated social card
Share a paper link and the preview is built from the paper itself: the engine renders the PDF with in-process MuPDF, extracts and scores embedded figures against detected page crops, falls back to an abstract screenshot, then composes a 1200×630 card — palette, typography, and layout derived from the winning figure. The whole pipeline is native Go inside the server; the original Python/Pillow subprocess version was retired.
under the hood
Built like the rest of app.nz: small, fast, self-hosted
Semantic search on gobed
Queries run through the gobed int8 engine with IVF-HNSW-PQ indexes over 725K paper embeddings — millisecond lookups on CPU, GPU CAGRA when attached. The same engine sold as the gobed addon.
Code-linked results
Every paper joins against 613K paper-to-repository links plus methods, datasets, and leaderboard evals, so a search lands you on runnable code, not just a PDF.
Live arXiv crawler
A Go crawler ingests new arXiv submissions continuously with historical backfill, keeping coverage gap-free since the archive dumps ended.
Figure-aware social cards
Every paper gets an OG card composed from its best PDF figure: pages render in-process via MuPDF, embedded images are extracted and scored, and the card is drawn natively in Go — no Python, no subprocesses.
Embedded PDF reader
Open any paper in a viewer with the PDF inline, linked code repos, tasks, and related papers one click away.
One Go binary
fasthttp + SQLite WAL + gobed in a single process. The whole 725K-paper engine, its crawler, and its image pipeline deploy as two binaries and a systemd unit.
api
Everything is an endpoint
| Endpoint | Purpose |
|---|---|
| GET /api/search?q=…&type=papers|methods|datasets | semantic search with similarity scores |
| GET /api/papers?limit=&offset=&q= | paginated paper listing with filters |
| GET /api/paper?id=… | full paper record: abstract, authors, tasks, code links |
| GET /api/eval?task=…&dataset=… | leaderboard results for a task/dataset pair |
| GET /view/paper?id=… | reader UI with embedded PDF |
curl "https://papers.app.nz/api/search?q=state+space+models&type=papers&limit=5"
From query to running code
Papers search is free to use, wired into app.nz agents and Deep Research, and powered by the same gobed engine you can attach to your own apps.