paper-search

Search OpenAlex + arXiv, fuse compact query variants, then rank by relevance and impact relative to the paper's own field and age. No API key or auth.

python3 scripts/paper_search.py "agent runtime" -q "LLM agent infrastructure"
python3 scripts/paper_search.py "distributed storage" --institution Alibaba
python3 scripts/paper_search.py "crash consistency" --venue FAST
python3 scripts/paper_search.py "agent memory" --fresh 60 --category cs.AI
python3 scripts/paper_search.py --after arXiv:2307.03172 --about "position bias"
python3 scripts/paper_search.py --selftest

Repeat -q for alternate terminology; papers matching several variants rank higher within their quality bucket. --institution resolves a lab/company and includes its descendants; --institution-type company searches industry broadly. Venue aliases include FAST, OSDI, SOSP, NSDI, MLSys, NeurIPS, ICML, ICLR, VLDB, and SIGMOD. --after takes arXiv:ID, DOI:x, or an OpenAlex Wxxx. Feed a useful displayed topic back through --topic. Use --json for machine output and --help for all flags.

Reading the output

Papers are bucketed, best first. The buckets are the point — recency and quality are independent axes, and collapsing them into one score hides exactly the tradeoff that matters in a fast-moving field.

Each line shows date (age) · citations (velocity) · fwci · [tier] venue · authors h=<max h-index>. FWCI is field-weighted citation impact: 1.0 = exactly the average for that field and year, so it lets a systems paper with 90 citations correctly outrank an LLM paper with 300. # shows OpenAlex's inferred topic and high-confidence keywords; use them to refine a query, not as ground truth. [TOP] includes FAST and the major AI/systems venues.

Traps

• Citation counts are a lagging indicator. In AI/infra, the paper that matters may be 6 weeks old with zero citations. Never conclude "nothing exists" from a citation-ranked list — run --fresh 60 before saying a topic is unexplored.
• FWCI is noise below ~18 months — the expected-citation denominator is near zero, so 7 citations can score FWCI 108. The script ignores it below that age; don't reintroduce it by reading the raw number off a young paper.
• Preprint date ≠ publication date. A 2024 TACL paper may be a 2023 arXiv paper; the idea landed 8 months before the venue date, and in a fast field that lead is the story. Output shows preprint YYYY-MM when they differ — cite the earlier one for priority.
• Author credibility is the weakest signal here; don't lean on it. OpenAlex conflates common names ("Kevin Lin": h=75 across 825 works, several people), and h= is the max across authors — the value conflation inflates. Prefer the @ Stanford, Berkeley affiliation beside it. On --fresh arXiv hits both are absent by design: arXiv exposes no institutions, and name lookup is worse than useless ("Feng Wang" → 4,865 authors, so you'd attach a stranger's h-index). That is what FRESH? means — open the PDF.
• A landmark is cited by every field. --after on a famous paper returns medical and legal applications too; pass --about "<keywords>" to keep the frontier on topic.
• Affiliation is a scope, not a quality score. Resolve a named lab/company with --institution; use --institution-type company only when industry-wide recall is wanted.
• Topics and keywords are inferred and sometimes polysemous. Expand only labels that fit the paper's title/abstract. Never promote a result solely because a tag matches.
• --field cs is the default. Pass --field any for anything else, or results look mysteriously empty. Venue metadata is imperfect regardless — a paper published at EMNLP may still read [PREPRINT] arXiv, so trust the citation numbers over the tier label.

Workflow for a fast-moving topic

1. Turn the question into 2–4 short variants: canonical phrase, acronym, alternate community term, and mechanism. Pass them separately with -q; do not make one long synonym query.
2. Inspect the first pass's titles, topics, and keywords. Rerun only useful discovered terms; add --topic, --institution, --institution-type company, --venue, or --category when asked.
3. paper_search.py --after <landmark-id> --about "<topic>" --since <~12mo ago> — what built on it since, ranked by impact. This is how you avoid citing a superseded result.
4. paper_search.py "<topic>" --fresh 45 — what dropped in the last few weeks, which step 1 structurally cannot see.
5. Read the abstracts, then the two or three papers that actually earned it.