IGM Extractor

Proof-of-concept project showcasing the capability of extracting text from IGM (Import General Manifest) documents using Vision Language Models (VLMs).

How It Works

• Upload an IGM PDF via the web UI.
• Pages are rendered to images and sent to Gemini for extraction.
• The app merges per-page results (manifest + containers), applies fixes, and stores data in an embedded H2 database.
• Review and edit extracted fields, then save.

Getting Started

Prerequisites

• Java 17+ and Maven

• You need to enable Gemini API from here (Google Cloud Console).

• Then, get your API key from here (Google AI Studio).

Build

mvn clean package -DskipTests

You can directly download the .jar file from latest release and run that. (Recommended)

Run

$env:GEMINI_API_KEY = "your_gemini_api_key"
java -jar target\igm-extractor-0.0.1-SNAPSHOT.jar

Or

java -jar target/igm-extractor-0.0.1-SNAPSHOT.jar --gemini.api.key=YOUR_API_KEY

Open http://localhost:8080 to access the app.

Notes

• H2 database: ./data; uploads: ./uploads.
• Reset app state: stop and delete ./data.
• H2 Console: http://localhost:8080/h2-console (JDBC: jdbc:h2:file:./data/igm-extractor, user: sa).

Technologies Used

Core Framework:

• Spring Boot 3.2.1 (Java 21)
• Spring Boot Starter Web (REST APIs, MVC)
• Spring Boot Starter Data JPA (ORM, database access)
• Spring Boot Starter Thymeleaf (server-side HTML templates)

AI/ML:

• Google Gemini API (gemini-2.0-flash-exp model via REST API)
• Spring RestTemplate (HTTP client for Gemini API calls)
• Jackson (JSON parsing and data binding)

Database:

• H2 Database (embedded file-based SQL database)

Document Processing:

• Apache PDFBox 2.0.31 (PDF rendering and image extraction)

Frontend:

• Thymeleaf templates
• Bootstrap 5.3 (responsive UI)
• Bootstrap Icons

Development Tools:

• Spring Boot DevTools (hot reload, debugging)

Performance

1. File Save - Time to save the uploaded PDF file to disk and create the IGM record

2. Render Pages - Time to convert PDF pages to images (PNG format at 150 DPI)

3. Gemini API Calls - Time for all API calls to Google Gemini (parallel processing for multiple pages)

4. Parse & Transform - Time to parse JSON responses and transform into entity objects

5. DB Save - Time to save all manifest and container data to the database

   Step	Duration	% of Total
   File Save	84 ms	0.4%
   Render Pages	4,326 ms	19.1%
   Gemini API Calls	18,156 ms	80.3%
   Parse & Transform	10 ms	0.0%
   DB Save	28 ms	0.1%
   Total	22,604 ms	100%