DeepXiv · OSV (osv)

About 200k open-source package security advisories — the part of osv.dev that CVE does not cover (synced daily; exact count in corpus). Over 90% are malicious-package reports (MAL-*, overwhelmingly npm); the remaining ~13k are standalone advisories with no CVE id (GHSA / RUSTSEC / PYSEC / GO …). Malicious behaviours (closed set of 13), the legitimate package being impersonated, IOCs (domains / URLs / IPs / filenames, each verified to appear in the advisory text) and one-line summaries are model-extracted. Hybrid vector + lexical search; filters and sorting in Postgres.

Corpus: 199,701 documents.

Setup

pip install -U deepxiv-sdk      # deepxiv registers a token on first use

With a registered key (https://data.rag.ac.cn/register), search and read are fair use: 60/min, 5 at once, 5,000/day; on the auto-registered token they spend the deepxiv daily limit at the credits listed below. ask draws on the daily ask allowance (resets 00:00 UTC; a site-wide daily total can run out early at peak times), or 3 lifetime trial asks on the auto-registered token. On 429, back off for Retry-After seconds. Data goes to stdout; cost, quota and sources go to stderr. Exit 0 ok, 1 API error, 2 usage error. The general CLI guide is the deepxiv-cli skill.

Start here

deepxiv fa search osv "malicious npm package credential theft" -F malicious=true -F ecosystems=npm -k 5
deepxiv fa search osv "*" -F behaviors=credential_theft --sort published -k 5
deepxiv fa read osv MAL-2026-17452 -l brief
deepxiv fa ask "Are there malicious npm packages impersonating popular crypto libraries?" -d osv

Free samples: ids MAL-2026-17452; queries malicious npm package.

Search

deepxiv fa search osv "query" [-F FILTER]... [-k N] [--offset N] [--mode M] [--sort S] [--head] [--json]

Modes: hybrid, dense, bm25, keyword. Up to 200 hits. Cost 2 (with --head: 3). Sorts: relevance, date, published (default relevance). An empty query is rejected; use "*" to list by filters only. Filters combine with AND; an undeclared field exits 2.

Advanced filters (see deepxiv fa spec osv): imitates, package_names, aliases, indicators.

Read

deepxiv fa read osv ID [ID...] [-l brief|detail|full] [-e EXTRA] [-p k=v]... [-f json|md|html]

At -l full, -p section=NAME reads one section and -p characters=N the first N characters.

Facets

deepxiv fa facets osv FIELD lists the legal values and counts, free. Use it instead of guessing a spelling. Fields: source, vuln_class.

Ask

deepxiv fa ask "question" -d osv [-e low|medium|high] [-p k=v]... [--no-stream] [--json]

Agentic Q&A over this domain: hybrid vector + lexical search (filters the vector layer lacks fall back to Postgres full text); the model reads the hits and answers with [osv:<doc_id> <title>] citations; /stream is NDJSON with real token deltas

Always pass -d osv: automatic domain picking is still early.

What this corpus will and will not tell you

• Division of labour with cve: cve holds CVE-numbered vulnerabilities (about 390k from NVD plus KEV and EPSS; exact count in /v1/cve corpus); osv holds what CVE does not cover — overwhelmingly poisoned / impersonating packages (mostly npm), which never get a CVE id. The two domains do not overlap; for supply-chain poisoning, use osv.
• Reverse lookup by IOC: filters.indicators=["some-domain"] finds every package that exfiltrates to it. Each IOC was verified to appear in the advisory text — not model-invented.
• Typosquatting: filters.imitates is a fuzzy match, so "requests" finds the packages impersonating it.
• behaviors and malicious are model-assigned — fine as a summary, not as a quote; the advisory text is details, in level=detail.
• ecosystems and package_names are arrays; passing a list means any-match.
• Displaying OSV data requires attribution (see the attribution field): the data comes from OSV.dev and its upstream databases, and each advisory's original source (GitHub Advisory / RustSec / PyPA …) keeps its own licence.
• Terms are ANDed (every term must match), so a long pile of words easily returns nothing — prefer fewer terms plus filters on the Postgres path. Two paths: with a query, no sort and only filters the vector layer has, search is hybrid vector + lexical (semantic recall, terms need not all match), topped up from Postgres; otherwise (sort given, indicators / imitates / published used, or mode=keyword) it runs in Postgres as an English tsvector and a Chinese zhparser pass, unioned; quoted phrases / OR / -exclusion work. Each hit's layer says which path.
• Model-written fields (summarise, do not quote): behaviors, imitates, impact, indicators, malicious, one_liner, one_liner_zh, vuln_class.
• The field table and enumerations are generated from the store's own self-description (search.field), not hand-written; enumeration counts are as of 2026-10-15.

Citing

Cite records as [MAL-2026-17452]. Say plainly when the corpus has nothing.

Generated by scripts/gen_fa_skills.py from deepxiv fa spec osv --json on 2026-10-19. Field names and prices can change; deepxiv fa spec osv is the live source.