CoquiTitle
CoquiTitle automates the generation of Puerto Rico property title study reports ("Informes de Estudio de Titulo") from official property registry documents.
What It Does
- Ingests PDF documents from the Registro de la Propiedad (KARIBE system)
- Extracts structured data using OCR and multi-pass LLM extraction with inline evidence
- Processes pending documents (documentos presentados) to capture real-time registry state
- Derives ownership rights using a deterministic engine
- Resolves evidence references to pixel-level bounding boxes
- Generates legal prose with embedded citation spans
- Maps every assertion back to source documents with pixel-level citations
Key Architecture Decisions
Evidence-First Extraction
Evidence is extracted WITH facts in the same LLM call, not retroactively mapped. This ensures consistency and avoids re-scanning documents.
Visual + Text Extraction
The multimodal LLM extracts facts from both visual PDF features (handwriting, stamps) and OCR text, enabling high accuracy even when OCR is imperfect.
Two-Pass Extraction with Parallelization
Pass 1 extracts summaries (3 parallel calls), Pass 2 extracts details (N parallel calls, up to 12 concurrent).
Deterministic Title State
Separate title-state-builder stage derives ownership rights without LLM, ensuring auditable and reproducible results.
Multi-Pass Report Generation
3 passes (header, prose, annotations) with passes 1+2 in parallel for optimal performance.
Pipeline Overview
Performance Characteristics
| Step | Duration | Notes |
|---|---|---|
| OCR (per document) | 10-30s | Depends on page count |
| Extraction Pass 1 | ~20-30s | 3 parallel calls |
| Extraction Pass 2 | ~15-40s | N parallel calls |
| Title State Building | <1s | Deterministic, no LLM |
| Evidence Resolution | 2-15s | Serial + parallel LLM fallback |
| Report Generation | ~25-35s | 3 calls (2 parallel) |
| Total (typical) | 3-5 min | End-to-end processing |
Cost per Case
| Component | Estimated Cost |
|---|---|
| Document AI OCR | ~$0.09 (9 pages) |
| Gemini extraction | ~$0.03-0.08 |
| Pending docs processing | ~$0.02-0.05 |
| Report generation | ~$0.01-0.02 |
| Evidence resolution | ~$0.00-0.01 |
| Total | ~$0.20-0.30 |