Hotel Construction Document Extraction Pipeline with Google ADK
Accor and Al-Qimmah's planned 4,000-room portfolio across Saudi Arabia requires rapid parsing of heterogeneous architectural and regulatory submittals. Build a document intelligence extraction pipeline using Google ADK and Upstage to parse multilingual blueprints and permits. Achieve 98% recall on bilingual key metadata extraction.
What you are building
The core problem, expected build, and operating context for this challenge.
Build a multimodal document processing agent using Google ADK and Upstage Layout Analyzer to extract structural metadata from submittals.
How work is evaluated
Evaluates precision and recall of bilingual document metadata extraction from construction submittals.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
permit_number_format_check
Verifies permit number follows standard KSA municipal string patterns
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
metadata_extraction_recall
Recall metric across all required architectural fields • target: 0.98 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Construct a Google ADK agent integrated with Gemini 2.0 multimodal capabilities
Leverage Upstage Document AI for layout-aware document extraction
Extract bilingual (Arabic/English) room dimensions, zoning approvals, and MEP specs
Validate extracted JSON schemas against Saudi Ministry of Housing regulatory requirements
Reference links and supporting material
50 synthetic bilingual construction submittals and hotel floor plans formatted as multi-page PDFs.
Challenge setup
Evaluates precision and recall of bilingual document metadata extraction from construction submittals.
Challenge input
PDF file stream or base64 encoded document image
Google ADK
Provides seamless integration with Vertex AI multimodal models.
Upstage
State-of-the-art layout awareness and OCR quality for complex submittal packages.
Evaluated output
JSON object with extracted key-value pairs and confidence scores
- Verifies permit number follows standard KSA municipal string patterns
- Recall metric across all required architectural fields • target: 0.98 • range: 0-1
- Benchmark: DocVQA-Bilingual
- Metadata Extraction Recall target: 0.98
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
- Google ADK · Required
- Google · Optional
- Upstage · Optional
Evaluation contract
- permit_number_format_check · Weight 1
- metadata_extraction_recall · Weight 1
Metrics
- metadata_extraction_recall · Target 0.98
Benchmarks
- DocVQA-Bilingual
- general
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: PDF file stream or base64 encoded document image
Output: JSON object with extracted key-value pairs and confidence scores
Execution harness: Python
Evaluation tools
- Google ADK · Agent Framework
- Upstage · Document Parsing Engine
Recipe freeze marker
No recipe freeze time is recorded.