convez.ai
PluginAutomaçãoAvançado
Juadsuarezsan

Document Intelligence Pipeline

Juadsuarezsan/document-intelligence-pipeline

Extração de campos estruturados de PDFs complexos com OCR, Vision e validação Pydantic

Ver no GitHub
InicialPluginScore · de 0 a 100

Pipeline de processamento inteligente de documentos que extrai campos estruturados de PDFs complexos (contratos, relatórios financeiros, formulários escaneados). Combina análise de layout, OCR com fallback para Claude Vision, validação Pydantic e roteamento baseado em confiança para casos de uso em legal, finanças e healthcare.

Instalação

$npx convez-ai Juadsuarezsan/document-intelligence-pipeline

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

  • Processamento automatizado de contratos legais e extrair cláusulas-chave
  • Digitalização de relatórios financeiros e invoices com extração de valores e tabelas
  • Processamento de formulários médicos e administrativos escaneados

Veredito

Pipeline industrial bem arquitetado para IDP em produção, com três caminhos otimizados (text-based, OCR, Vision) e validação cruzada. Projeto ainda em fase de planejamento, mas roadmap técnico claro e stack moderno (Pydantic v2, Claude Vision, docling). Recomendado para times que precisam processar documentos variados em escala com alta confiabilidade.

README

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

Project 04 — Document Intelligence Pipeline

Intelligent Document Processing for complex PDFs: contracts, financial reports, medical forms, scanned forms with tables. Layout analysis → OCR → Claude Vision fallback → Pydantic-validated extraction → confidence-gated routing.

Status LLM Domain

Industrial use case: Hyperscience, Rossum, Klarity — IDP for legal-tech, fintech, healthcare, insurance.

What this project does

Extracts structured fields from messy real-world PDFs. Routes simple text PDFs through unstructured/docling (fast, cheap), scanned ones through OCR (Tesseract/PaddleOCR), and the hard ones through Claude Vision. Validates extractions with Pydantic schemas + cross-field rules. Auto-approves high-confidence; routes the rest to a human reviewer.

Architecture

PDF input
   │
   ▼
[Document Classifier] Claude Vision → type: contract | invoice | form | report
   │
   ▼
[Layout Analyzer]
   ├─ text-based → unstructured / docling
   ├─ scanned   → OCR (Tesseract or PaddleOCR)
   └─ hard      → Claude Vision direct
   │
   ▼
[Table Extractor] Camelot or Table Transformer (TATR)
   │
   ▼
[Field Extractor] Claude + Pydantic schema per document type
   │
   ▼
[Validator]
   ├─ regex (NIT, dates, amounts)
   ├─ cross-field (subtotal + tax = total)
   └─ LLM-as-judge for coherence
   │
   ▼
[Confidence Scorer] per-field + global
   ├─ > 0.9 → auto-approve
   ├─ 0.7-0.9 → human review queue
   └─ < 0.7 → re-process with more expensive VLM

Roadmap to v1.0.0

  1. Download FUNSD (199 scanned forms) + DocVQA (12K docs) + PubLayNet (360K pages)
  2. Document type classifier (Claude Vision)
  3. Three-path layout analyzer (unstructured | OCR | Vision)
  4. Table extractor (Camelot for text PDFs, TATR for image-based)
  5. Pydantic schemas per document type (invoice, contract, form, report)
  6. Regex + cross-field validators
  7. Confidence scorer + routing
  8. Eval: field-level F1 on FUNSD, doc-level accuracy on DocVQA, table F1 on FinTabNet
  9. Next.js demo with PDF viewer side-by-side with JSON output + confidence per field
  10. Gallery of 10 pre-processed documents accessible from the demo

Stack

LayerTechnology
Parsingunstructured 0.16+ or docling (IBM, 2024)
OCRTesseract 5 or PaddleOCR
VLMClaude Sonnet 4.5 with vision input
TablesCamelot or Table Transformer
LayoutLayoutParser or detectron2
SchemasPydantic v2
ReasoningClaude Sonnet 4.5
StoragePostgreSQL
FrontendNext.js with react-pdf viewer
ObservabilityLangSmith

Definition of Done — project-specific

  • Pipeline works on all 3 input types: clean text, scanned, complex layout
  • Field-level F1 reported on FUNSD test set
  • Document-level accuracy reported on DocVQA
  • Table extraction evaluated on FinTabNet or equivalent
  • Demo: upload a PDF, see extraction with confidence scores in real time
  • Gallery of 10 pre-processed sample documents linked from demo
  • Comparative table: unstructured-only vs unstructured + Vision fallback vs Vision direct

Plus the 12 universal DoD blocks.

License

MIT.

Ver o original no GitHub

Bom para

Relacionados