convez.ai
PluginFinanceiroAvançado
Juadsuarezsan

GraphRAG sobre SEC EDGAR 10-K

Juadsuarezsan/graphrag-sec-edgar

Inteligência financeira via grafo de conhecimento + busca híbrida em 10-Ks

Ver no GitHub
InicialPluginScore · de 0 a 100

Plugin que constrói grafos de conhecimento a partir de relatórios 10-K da SEC (S&P 500), combinando extração de entidades com Claude, armazenamento em Neo4j e busca híbrida graph+vector. Responde perguntas multi-hop complexas sobre relacionamentos corporativos, cadeias de suprimento e exposições regulatórias.

Instalação

$npx convez-ai Juadsuarezsan/graphrag-sec-edgar

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 1 mês
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Due diligence de M&A e análise de cadeia de suprimento
  • Mapeamento de relacionamentos corporativos e dependências entre S&P 500
  • Identificação de exposições regulatórias compartilhadas entre empresas

Veredito

Projeto ambicioso em estágio early (planejado), ainda sem implementação completa. Endereça um problema real em inteligência financeira (quem depende de quem, exposições compartilhadas), combinando GraphRAG + Cypher + embeddings em stack maduro (Neo4j + Claude). Requer expertise profunda em data eng e NLP; adequado para times fintech ou consultoria que precisam de análise relacional de earnings reports em escala.

README

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

Project 08 — GraphRAG over SEC EDGAR

Knowledge graph from public 10-K filings of the S&P 500. Answers complex multi-hop questions about the company ecosystem: who supplies whom, who sits on competing boards, which companies share regulatory exposure.

Status DB Technique

Industrial use case: Financial intelligence (Visible Alpha, Tegus, AlphaSense), strategic consulting, M&A.

What this project does

Ingests 10-K filings (annual reports filed with the SEC) from S&P 500 companies. Extracts entities (companies, persons, subsidiaries, products) and relationships (board memberships, supplier ties, competitive overlap) using Claude. Stores in Neo4j with native vector index for hybrid graph+vector retrieval. Answers questions like "which S&P 500 companies depend on TSMC and what is their combined exposure?"

Architecture

FASE DE INGESTA (one pass over corpus):
   10-K PDF/HTML
   │
   ▼
[Document Chunker] respects 10-K sections (Item 1, Item 1A, etc.)
   │
   ▼
[Entity Extractor] Claude + Pydantic
   → entities: [{id, type, properties}]
   → relationships: [{from, to, type, properties}]
   │
   ▼
[Graph Merger] Neo4j MERGE
   → no duplicates; embedding of name+desc stored as node property
   │
   ▼
[Staleness Tracker] each node has updated_at + source_filing_date

FASE DE CONSULTA:
   Question: "Which S&P 500 companies depend on TSMC?"
   │
   ▼
[Query Classifier] Claude → multi_hop + aggregation
   │
   ▼
[Entry Point Finder] vector search on Neo4j → seed nodes
   │
   ▼
[Graph Traversal] Cypher
   MATCH (tsmc)-[:SUPPLIED_BY]<-(c:Company {in_sp500: true})
   │
   ▼
[Result Aggregator] → ordering, totals, ratios
   │
   ▼
[Synthesizer] Claude → natural-language answer citing nodes
   │
   ▼
[Visualization] subgraph for frontend

Roadmap to v1.0.0

  1. SEC EDGAR API integration — download 10-Ks for top 100 S&P 500 companies, last 3 years
  2. Document chunker that respects 10-K section structure
  3. Entity extraction with Pydantic schemas (Company, Person, Product, Risk, Subsidiary)
  4. Neo4j with APOC plugin + GDS for community detection
  5. Embedding ingestion (Voyage AI voyage-3 → Neo4j native vector index)
  6. Query classifier + traversal + synthesizer pipeline
  7. Eval set: 100 multi-hop questions across 4 categories with manual ground truth
  8. Comparative table: Vector RAG vs GraphRAG vs Hybrid
  9. Next.js demo with react-force-graph-2d visualization
  10. Staleness detector flags nodes with data >365 days old
  11. Graph schema documented in docs/graph_schema.md

Stack

LayerTechnology
Graph DBNeo4j 5+ with APOC + GDS plugins
Query languageCypher
Vector indexNative in Neo4j 5.13+
LLMClaude Sonnet 4.5 (entity extraction + synthesis)
SchemasPydantic v2
EmbeddingsVoyage AI voyage-3 or bge-large-en-v1.5
FrameworkLangChain neo4j-graphrag or LlamaIndex KnowledgeGraphIndex
FrontendNext.js + react-force-graph-2d
AuditPostgreSQL for staleness tracking
ObservabilityLangSmith

Definition of Done — project-specific

  • Knowledge graph populated with top 100 S&P 500 + 3 years of 10-Ks (≥5K nodes, ≥20K edges)
  • Eval set of 100 multi-hop questions with manual ground truth
  • Vector RAG vs GraphRAG comparison with reproducible numbers
  • Demo with interactive graph visualization
  • Pre-baked queries covering 4 categories (lookup, 2-hop, 3+ hop, aggregation)
  • Staleness detector active
  • Graph schema documented

Plus the 12 universal DoD blocks.

License

MIT.

Ver o original no GitHub

Bom para

Relacionados