Skip to main content
โ† Back to Gallery

๐Ÿ“„ Real Estates Dossier

Italian real estate auction PDF extraction workflow. Pages are rendered to PNG, OCR'd individually with Cirrascale's olmOCR-2-7B-1025 model, and joined into raw and corrected text outputs. Structured extraction is opt-in. When enabled, document classification feeds a TypeSafe Jev decision pass first; a user-defined policy can then skip or trigger the Gemma full JSON extraction pass. v3 improvements over v2: - Cirrascale OCR with per-page processing and response capture - Text corrections and intermediate artifacts - Structured logging with proper exception hierarchy Features: - Battle-tested OCR pipeline with text corrections - Parallel page processing - Comprehensive caching and workspace management

Workflow Details

Workflow ID: real_estates_extraction_v3
Version: 3.1.0
Total Steps: 3

Workflow Steps

1
Ocr Extraction
python
โ†“
2
Document Classification
python
โ†“
3
Ensemble Extraction
python

Input Configuration

Select an instance to apply configuration overrides.
Supported formats: PDF, DOC, DOCX, TXT, PNG, JPG
Path to PDF document
Enable text corrections
Fix encoding issues
Digit correction mode: auto (impact-based), on (always apply), off (never apply)
Re-run OCR even when artifacts already exist
Enable DSE: classify the document, run JEV first, then conditionally run Gemma.
Minimum document-classification confidence required for DSE.
Gemma model used when the JEV policy says full extraction is needed.
First-stage TypeSafe decision model.
Optional TypeSafe questions JSON; blank derives questions from the classified schema.
Policy JSON controlling whether Gemma runs after JEV.
JEV failure handling: fail, or run_gemma to fall back to Gemma.

Execution Preview

This workflow will execute the following steps in order:

1 Ocr Extraction
Processor: python
โ†“
2 Document Classification
Processor: python
โ†“
3 Ensemble Extraction
Processor: python
Total Steps: 3

Example Output

{
  "structured_data": {
    "sale_id": "RG 123/2024",
    "tribunal": "Tribunale di Roma",
    "lot_number": "1",
    "assets": [...],
    "valuation": {
      "market_value": 250000.00,
      "auction_base_price": 187500.00
    }
  },
  "metadata": {
    "processing_time": 24.7,
    "llm_provider": "mistralai/Mistral-7B-Instruct-v0.3"
  }
}