# ocr/app/engines/

OCR engine abstraction layer. Two categories of engines:

1. **OcrEngine subclasses** (image-to-text): PaddleOCR, Google Vision, Hybrid. Accept image bytes, return text + confidence + word boxes.
2. **GeminiEngine** (PDF-to-structured-data and VIN decode): Standalone module for maintenance schedule extraction and VIN decoding via google-genai SDK. Accepts PDF bytes or VIN strings, returns structured JSON. Not an OcrEngine subclass because the interface signatures differ.

## Files

| File | What | When to read |
| ---- | ---- | ------------ |
| `__init__.py` | Public engine API exports (OcrEngine, create_engine, exceptions) | Importing engine interfaces |
| `base_engine.py` | OcrEngine ABC, OcrConfig, OcrEngineResult, WordBox, exception hierarchy | Engine interface contract, adding new engines |
| `paddle_engine.py` | PaddleOCR PP-OCRv4 primary engine | Local OCR debugging, accuracy tuning |
| `cloud_engine.py` | Google Vision TEXT_DETECTION fallback engine (WIF authentication) | Cloud OCR configuration, API quota |
| `hybrid_engine.py` | Combines primary + fallback engine with confidence threshold switching | Engine selection logic, fallback behavior |
| `engine_factory.py` | Factory function and engine registry for instantiation | Adding new engine types |
| `gemini_engine.py` | Gemini 2.5 Flash integration for maintenance schedule extraction and VIN decoding (google-genai SDK, 20MB PDF limit, structured JSON output, Google Search grounding for VIN decode) | Manual extraction debugging, VIN decode, Gemini configuration |

## Engine Selection

```
create_engine(config)
    |
    +-- Primary: PaddleOCR (local, fast, no API limits)
    |
    +-- Fallback: Google Vision (cloud, 1000/month limit)
    |
    v
HybridEngine (tries primary, falls back if confidence < threshold)
```

GeminiEngine is created independently by ManualExtractor and the VIN decode router, not through the engine factory.