Davidsv/FrenchBee_vision_dataset

FRENCHBEE-TECHNICAL-QUERY-DATASET

This dataset contains a structured collection of technical queries generated from FrenchBee technical documents. It is designed to train and evaluate information retrieval models and improve AI understanding of aerospace technical documentation.

About Me

I'm David Soeiro-Vuong, a third-year Computer Science student working as an apprentice at TW3 Partners, a company specialized in Generative AI. Passionate about artificial intelligence and language models optimization, I focus on creating efficient model merges that balance performance and capabilities.

🔗 Connect with me on LinkedIn

Project Details

Purpose

This tool automatically generates technical queries from aviation documents to:

• Create training datasets for fine-tuning retrieval models
• Develop domain-specific question generation capabilities
• Support multilingual technical understanding in the aerospace domain

Features

• PDF Processing: Processes multi-page technical documents with contextual awareness
• Multimodal Analysis: Uses image-based document understanding with Gemini 1.5 Flash
• Multilingual Support: Generates queries in five languages (French, English, Spanish, German, Italian)
• Query Categorization: Creates four distinct types of technical queries:
  • Main Technical: Focused on specifications and regulations
  • Secondary Technical: Detailed aspects of technical operations
  • Visual Technical: Related to diagrams and technical visualizations
  • Multimodal Semantic: Complex queries combining multiple technical aspects

Implementation Details

• Rate Limiting: Implements an async rate limiter for API request management
• Concurrency: Uses asyncio for parallel document processing
• Image Processing: Captures high-resolution page images using PyMuPDF
• Structured Output: Generates Pydantic-validated query objects
• Results Tracking: Records processing progress and errors in JSONL format

Example Usage

The tool processes a folder of PDF documents and generates multiple technical queries for each relevant page:

# Process a folder of aviation PDFs
PDF_FOLDER = "/path/to/aviation/documents"
OUTPUT_FILE = "/path/to/output/queries.jsonl"

# Run the processing pipeline
asyncio.run(process_pdf_folder(PDF_FOLDER, OUTPUT_FILE))

Example Output

{
  "pdf_name": "operations_manual.pdf",
  "page_number": 42,
  "language": "FR",
  "queries": {
    "language": "FR",
    "query1": "Quelles sont les implications des réglementations ETOPS sur l'exploitation des vols long-courriers vers Tahiti selon le manuel d'exploitation ?",
    "query2": "Comment la gestion des performances des A350 est-elle optimisée pour réduire la consommation de carburant sur les vols transatlantiques ?",
    "query3": "Pouvez-vous expliquer l'interprétation des courbes de consommation spécifique sur le diagramme de performance des A350-900 ?"
  },
  "timestamp": "2025-02-28T14:30:45.123456"
}

Use Cases

This project is valuable for:

• Creating specialized training datasets for aerospace document understanding
• Developing technical query generation capabilities for aviation knowledge bases
• Supporting multilingual technical documentation retrieval systems
• Enhancing AI's ability to process complex aviation regulations and manuals

Dependencies

• Python 3.8+
• PyMuPDF
• aiofiles
• instructor
• LiteLLM
• Pydantic
• Gemini 1.5 Flash API access