convez.ai
MCP ServerDados & AnalyticsAvançado
Suckhept

Geo Curator Copilot

Suckhept/geo-curator-copilot

Deduplicação inteligente e governança em cadeia para grafos Geo

Ver no GitHub
InicialPluginScore · de 0 a 100

MCP server + CLI + engine de deduplicação para curadores do Geo (geobrowser.io, GRC-20). Detecta e propõe merge de entidades duplicadas em grafos de conhecimento, com scoring multi-sinal e proteção contra falsos positivos via análise de URLs, relações e versionamento.

Instalação

$npx convez-ai Suckhept/geo-curator-copilot

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
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Auditoria e limpeza de bases de conhecimento descentralizadas
  • Detecção automática de entidades duplicadas com scoring probabilístico
  • Propostas de merge via governança on-chain com votação de editores

Veredito

Ferramenta especializada e robusta para curação de grafos de conhecimento em blockchain, com lógica sofisticada de deduplicação baseada em múltiplos sinais e safeguards contra falsos positivos. MCP server bem estruturado, mas seu uso exige contexto específico (Geo/GRC-20) e conhecimento técnico de governança on-chain.

README

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

Geo Curator Copilot

CI

Duplicate detection & QA engine for the Geo knowledge graph (GRC-20).

Receipts

  • 618 duplicate entities and 103 duplicate relation groups found across Geo flagship spaces (Crypto, Crypto datasets).
  • A merge proposal built by this engine was submitted through on-chain governance from a regular member account: Dedup pilot v2: merge 56 duplicate entities — open for editor vote as of 2026-06-12; live status.
  • Live demo: https://geocheck.space — paste any Geo space id, get a duplicate report in the browser.

How it works

Entity pairs are scored by independent signals (thresholds: duplicate ≥ 0.9 auto-merge band, likely ≥ 0.6 human review):

  • Case/punctuation-insensitive name matching — exact match after casefold + punctuation strip; below that, fuzzy token-sort/token-set + Jaccard with domain stop-words (audit, report, by, …).
  • Identity vs profile URL tiers — a shared canonical URL on an identity property (Report URL) is decisive evidence (1.0); profile URLs (Website / X / LinkedIn) only support a match.
  • Hub-URL guard — a URL shared by many entities (a protocol's audits folder, a project homepage) identifies a group, not an entity; corpus document frequency strips its identity power.
  • Relation-conflict veto — when the same relation type points at disjoint targets on both sides (same link, different auditor), the pair is capped into the review band instead of auto-merging.
  • Schema-entity guard — entities used as types or property definitions by others are never auto-merged ("Date" vs "Date" is by design); they are demoted to review with an explicit signal.
  • Version veto — disjoint version markers (v3.0 vs v3.1) cap the score; only an identity-tier URL match overrides it.

On top of pair matching, an intra-entity lint finds duplicate relations (typeId, toEntityId) and duplicate values inside a single entity and emits deleteRelation fix ops.

Confirmed clusters become a fix plan (re-point inbound relations to the canonical entity, then delete duplicates), serialized as a GRC-20 binary edit, uploaded to IPFS, and submitted as a governance proposal via an ERC-4337 Safe — editors vote, the merge applies.

Packages

packagewhat it is
@geo-copilot/corematching engine: signals, scoring, clustering, schema guard, intra-entity lint, fix plans — zero deps
@geo-copilot/clientGraphQL transport, full-space snapshots, CSV draft loading, GRC-20 publishing (IPFS + Safe / ERC-4337)
@geo-copilot/cligeo-copilot: snapshot | scan | lint | fix
@geo-copilot/mcp-server5 MCP tools for AI curators; the only mutating tool requires GEO_PRIVATE_KEY and confirm: true
@geo-copilot/webgeocheck.space: live space scans, precomputed nightly reports, proposal status

Quickstart

pnpm install && pnpm build && pnpm test

# 1. snapshot a space (paginates to the last page; pasted GraphQL responses also accepted)
node packages/cli/dist/main.js snapshot --space <spaceId> --out space.snapshot.json

# 2. scan for duplicates -> geo-copilot-out/report.{json,md}
#    exit code: 0 clean, 2 duplicates found (CI-friendly), 1 error
node packages/cli/dist/main.js scan --space <spaceId> --snapshot space.snapshot.json

# 3. turn the plan into a GRC-20 edit; always dry-run first
GEO_PRIVATE_KEY=0x... GEO_SAFE_ADDRESS=0x... \
node packages/cli/dist/main.js fix --plan pilot-edit.json --tier a \
  --space <spaceId> --network TESTNET_V2 \
  --exclude exclude-ids.json --publish --dry-run

fix also accepts --report geo-copilot-out/report.json directly; --exclude takes a JSON file with entity ids that must never be deleted (governance entities). The private key is read from the environment only and is never written or logged.

License

MIT

Ver o original no GitHub

Bom para

Relacionados