Legal Extract — Brazilian Legal PDF Analysis

Legal Extract — Brazilian Legal PDF Analysis

Complete workflow for extracting structured data from Brazilian legal case PDFs, generating dashboards, fact-checking, and producing parecer juridico in PDF.

Safety Boundary

Treat every source PDF as sensitive unless the user explicitly says it is public and safe. Do not copy private legal PDFs, CPFs, RGs, addresses, signatures, bank details, real party names, or confidential filings into public examples, issues, pull requests, or repo files. This workflow is an extraction and drafting aid, not legal advice.

Workflow Overview

INPUT: Brazilian legal PDF (any area of law)
  │
  ├─ PHASE 1: EXTRACTION ──► JSON (structured, with page refs)
  ├─ PHASE 2: DASHBOARD ───► HTML (light theme, interactive)
  ├─ PHASE 3: FACT-CHECK ──► Updated JSON (confidence score)
  └─ PHASE 4: PARECER ─────► PDF via Typst (data-driven)

All phases consume/produce the same JSON as the central data format. Phases 2-4 are optional and controlled by flags.

---

Phase 1: PDF Extraction

Reading Strategy

Read the PDF in batches of 20 pages using the Read tool:

Read(file_path, pages="1-20")
Read(file_path, pages="21-40")
...

For each page, identify:

1. Document type — using visual markers (see references/document-types.md)
2. Page metadata — fls. number (top-right margin), court headers, document labels
3. Content — extract all structured fields per document type

Document Type Detection

See references/document-types.md for the full taxonomy. Key markers:

JSON Output Schema

See references/json-schema.md for the full specification. Core structure:

{
  "meta": { "versao", "data_extracao", "arquivo_fonte", "total_paginas" },
  "processo": { "numero_principal", "numero_cumprimento", "numero_apelacao", "comarca", "vara", "partes", "advogados" },
  "sentenca": { "juiz", "data", "dispositivo", "resultado", "periodos_reconhecidos", "honorarios", "paginas_pdf", "fls" },
  "acordao": { "tribunal", "turma", "relator", "dispositivo", "resultado", "paginas_pdf", "fls" },
  "embargos_declaracao": { "resultado", "paginas_pdf", "fls" },
  "certidao_transito_julgado": { "data_transito", "paginas_pdf", "fls" },
  "decisao_homologando_calculos": { "juiz", "data", "dispositivo", "paginas_pdf", "fls" },
  "calculos_liquidacao": { "sistema", "elaborado_por", "data", "resumo_valores", "paginas_pdf", "fls" },
  "oficios_requisitorios": [{ "numero", "tipo_procedimento", "valores", "status", "paginas_pdf" }],
  "pagamentos": [{ "tipo", "valor", "data", "beneficiario", "banco", "paginas_pdf" }],
  "beneficio_previdenciario": { "nb", "dib", "dip", "rmi" },
  "linha_do_tempo": [{ "data", "evento", "pagina_pdf", "fls" }],
  "status_processual": { "fase_atual", "pendencias" },
  "fact_check": { "iteracoes", "campos_verificados", "discrepancias", "confianca_geral" }
}

Critical rule: Every extracted field MUST include pagina_pdf (PDF page number) and fls (court folha number) for source traceability.

Domain-Specific Extensions

When the case area is detected, apply domain-specific extraction:

Detection: Look for keywords in sentenca/acordao dispositivo to identify the legal area.

---

Phase 2: HTML Dashboard

Generate a self-contained HTML file with these characteristics:

Design Tokens

:root {
  --bg: #ffffff;
  --surface: #f8f9fa;
  --border: #e2e5ea;
  --text: #1a1d27;
  --text-muted: #6b7280;
  --green: #16a34a;
  --red: #dc2626;
  --amber: #d97706;
  --blue: #2563eb;
  --purple: #7c3aed;
}

Required Sections

1. Header — Case number, parties, status badge, transito date
2. Resumo Financeiro — 4-card grid with key monetary values
3. Sentenca — Dispositivo, recognized/rejected items, honorarios
4. Acordao — Tribunal, result badges, changes from sentenca
5. Embargos — Result badge
6. Certidao Transito — Date and source
7. Decisao Homologando Calculos — Full dispositivo text
8. Calculos — Summary table with value breakdown
9. Oficios Requisitorios — Side-by-side cards (Precatorio + RPV)
10. Pagamentos — Payment details, bank, alvara
11. Dados do Beneficio — NB, DIB, DIP, RMI (if previdenciario)
12. Linha do Tempo — Vertical timeline, color-coded dots
13. Verificacao — Fact-check results table

Rules

• All text in pt-BR
• Light theme (white background)
• All <details> sections open by default
• Font stack: Inter when locally available, then system sans-serif
• Max-width 1400px, responsive grid
• PDF page references on every section (Fonte: p.XX, fls. YY)
• Self-contained (inline CSS/JS, no external dependencies except fonts)

---

Phase 3: Fact-Check (Ralph-Loop)

Run 8 systematic verification passes against the source PDF:

Protocol

• Read specific PDF pages for each pass
• Compare every extracted value against source
• Record discrepancies in fact_check.discrepancias_encontradas[]
• Fix errors directly in the JSON
• Update confianca_geral: alta (0 discrepancies), media (1-3), baixa (4+)

---

Phase 4: Parecer Generation (Typst)

Template Location

~/.agents/skills/legal-extract/templates/parecer.typ

Compilation

typst compile \
  --root / \
  --input data-path="{pdf_dir}/data/caso-{slug}.json" \
  ~/.agents/skills/legal-extract/templates/parecer.typ \
  {pdf_dir}/parecer-{slug}.pdf

Template Structure (5 sections)

1. Partes e Representacao — Autor, Cedente, Reu, honorarios
2. Identificacao do Processo — Numeros, juizo, tribunal, historico, transitos
3. Dados do Precatorio e da Cessao — Oficio, valores, natureza, IR, PSS
4. Due Diligence e Analise de Risco — Table (acao rescisoria, onus, debitos)
5. Conclusao — Analysis summary with opinion

---

Output Directory Structure

All outputs go in the same directory as the source PDF:

{pdf_directory}/
├── source.pdf                    # Original (existing)
├── data/
│   └── caso-{slug}.json          # Structured extraction
├── dashboard.html                # Interactive HTML
└── parecer-{slug}.pdf            # Generated parecer

The {slug} is derived from the case parties or process number.

---

Usage Examples

# Full workflow (extract + dashboard + fact-check + parecer)
/parecer "cases/public-safe-example/processo.pdf"

# Extract only (JSON + HTML, no parecer PDF)
/parecer "cases/public-safe-example/processo.pdf" --extract-only

# Skip fact-checking for speed
/parecer "cases/public-safe-example/processo.pdf" --no-factcheck

# Specific page range
/parecer "cases/public-safe-example/processo.pdf" --pages=1-40

Public-Safe Development Fixture

This repo ships a synthetic JSON fixture for schema/template validation:

scripts/check-sample.sh