convez.ai
SkillBusca & PesquisaIntermediário
PlutoLei

Busca Multi-Fonte de Artigos Acadêmicos

PlutoLei/multi-source-paper-search

Busca acadêmica cross-source com dedup inteligente e consenso citado

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill que expande uma única consulta de literatura em busca paralela em quatro fontes acadêmicas (OpenAlex, Semantic Scholar, arXiv, Perplexity), deduplica determinísticamente por DOI/arXiv-id/título, e gera relatório de consenso com contribuições únicas destacadas. Transforma pesquisa fragmentada em síntese validada e citada.

Instalação

$npx convez-ai PlutoLei/multi-source-paper-search

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
1
Acompanhando
0
Saúde
42%
Último commit
há 3 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Revisões sistemáticas de literatura sem viés de fonte única
  • Pesquisa de estado-da-arte em tópicos emergentes
  • Validação de claims acadêmicos contra múltiplas bases
  • Síntese de divergências entre diferentes repositórios

Veredito

Ferramenta sólida que resolve um problema real: pesquisa acadêmica fonte-única é enviesada. A deduplicação determinística (union-find por DOI/arXiv-id) é mecanicamente robusta, e o split entre work mecânico (Python puro) e semântico (agent) é bem-arquitetado. Requer setup de chaves opcionais, mas degrada graciosamente — funciona só com OpenAlex se nenhuma chave estiver configurada.

README

A tradução para o português deste README ainda está na fila.

multi-source-paper-search

A Claude Code skill that turns one literature request into a cross-source scan. It fans out to four academic sources at once, dedupes the results deterministically, and reconciles them into a single cited synthesis — telling you what's corroborated, what's contested, and what only one source saw.

一次请求,跨 4 个学术源检索 → 确定性去重(DOI / arXiv-id / 标题并查集)→ 共识 / 分歧 / 独有贡献对齐 → 带引用的综合报告。多源去重与共识对齐是这个 skill 的核心价值,不是手搓拼凑能替代的。

What it does

Searching one source at a time gives you a partial, source-biased view. This skill:

  1. Expands your query into a few method-alias variants (recall insurance — dedup across variants is free).
  2. Fans out to OpenAlex, Semantic Scholar, arXiv (via Valyu), and Perplexity in one shot.
  3. Dedupes deterministically — scripts/fanout.py unions records by DOI / arXiv-id / normalized title (union-find), so a paper seen by 3 sources is counted once with a consensus = 3.
  4. Cross-validates claims — the agent reconciles abstracts and the Perplexity narrative into consensus vs. divergence.
  5. Writes a cited report — 检索概况 → 共识发现 → 分歧高亮 → 每源独有贡献 → 论文排名表 → 参考文献.

The split is deliberate: mechanical work (fan-out, normalization, dedup) lives in the script and is exact; semantic work (query expansion, claim-level reconciliation, the report) is the agent's half.

Sources & keys

SourceKeyNotes
OpenAlexnone240M+ works, citation counts. Always live — this is the floor; at least one source always runs.
Semantic Scholaroptional S2_API_KEYWorks keyless but the unauthenticated pool 429s readily; a free key makes it reliable. Best dedup anchor (exposes both DOI and arXiv id).
arXiv (Valyu)Valyu keyPreprints + full text for claim re-checking. Configured via the arxiv-search skill's setup (see below), not an env var.
PerplexityOPENROUTER_API_KEYWeb-grounded synthesis + recent work via OpenRouter (needs credits).

Graceful degradation: fanout.py probes each source and skips any with a missing key/script — it never crashes the run. Skipped sources are reported under skipped_sources, and the report always states "used X / 4 sources" so consensus counts are read in the right context. With zero keys you still get a full OpenAlex-only run.

External dependency — arXiv via arxiv-search

The arXiv source is the one piece that isn't self-contained: fanout.py shells out to the arxiv-search skill's Node locator, found dynamically with:

find ~/.claude -path "*/arxiv-search/*/scripts/search" -type f | head -1

If arxiv-search isn't installed, or its Valyu key isn't configured, the arXiv source simply lands in skipped_sources and the other three sources produce the report as normal. To enable it:

SEARCH=$(find ~/.claude -path "*/arxiv-search/*/scripts/search" -type f | head -1)
"$SEARCH" setup <your-valyu-key>   # free $10 credit at https://platform.valyu.ai

Everything else (fanout.py) is pure Python stdlib — no pip install, runs on any python3.

Install

# 1. Drop the skill into your Claude Code skills directory
cp -r multi-source-paper-search ~/.claude/skills/

# 2. Add your keys (optional — OpenAlex works with none)
cd ~/.claude/skills/multi-source-paper-search
cp .env.example .env
# edit .env: set OPENROUTER_API_KEY and/or S2_API_KEY

fanout.py auto-loads .env from the skill root on startup without overriding your shell environment — no source, no profile edits. .env is gitignored; only .env.example is committed.

Usage

In Claude Code — just ask, in any language:

"多源搜一下 retrieval-augmented generation 在临床问答上的论文" "find papers on X across sources — what's the consensus vs. what's contested?"

The skill triggers on cross-source literature intent (多源 / 交叉 / triangulated / corroborated …) and prefers itself over any single-source skill when the goal is a literature scan rather than fetching one known paper.

Standalone — the fan-out script runs on its own:

python scripts/fanout.py --depth 20 "retrieval augmented generation" "RAG clinical QA"

It emits one JSON object: papers (deduped, ranked, each with consensus = # distinct sources and per_source_unique), sources_used, skipped_sources, and perplexity_narrative. Depth defaults to 20/source — drop to ~10 for a quick scan, raise to 30+ for an exhaustive one.

Repo layout

multi-source-paper-search/
├── SKILL.md                      # the skill definition (procedure, anti-patterns, fallbacks)
├── scripts/fanout.py             # deterministic fan-out + dedup (pure stdlib)
├── references/
│   ├── sources.md                # per-source invocation, keys, how to add a 5th source
│   └── report-template.md        # the report structure the agent follows
├── evals/evals.json              # evaluation prompts
├── .env.example                  # key template (copy to .env)
└── LICENSE                       # MIT

Adding a fifth source is a contained change — see references/sources.md (Crossref / PubMed / CORE are good candidates).

License

MIT © 2026 PlutoLei

Ver o original no GitHub

Bom para

Relacionados