Skip to content

CHECKLIST.md

Latest commit

 

History

History
120 lines (92 loc) · 4.51 KB

CHECKLIST.md

File metadata and controls

120 lines (92 loc) · 4.51 KB
Raw

TextLexiq Checklist

Vision

Build a mobile-first document intelligence workspace that lets users capture paper documents, clean and OCR them reliably, refine text with on-device or cloud LLMs, and export to PDF/DOCX—while keeping data private and usable offline.

✅ Implemented (Working / Scaffolded)

Project & Architecture

  • MVVM structure with ui/, viewmodel/, data/, ocr/, exporter/, utils/ modules.
  • Navigation graph with Home → Scanner → Crop/Preview → OCR → Document.
  • Centralized ViewModel factory for injection-ready flows.

Scanner & Capture

  • CameraX live preview integrated in Compose.
  • Capture button saves image to internal storage.
  • Basic rectangular overlay guide for edge alignment.
  • Runtime camera permission flow handled.

Crop & Preprocess

  • Crop and preview screen with adjustable horizontal/vertical bounds.
  • OpenCV preprocessing pipeline:
    • Denoise (Gaussian blur)
    • Grayscale conversion
    • Perspective correction (basic contour-based deskew)
    • Adaptive thresholding (binarize for OCR)
  • Processed image is saved to internal storage and forwarded to OCR.

OCR

  • ML Kit Text Recognition implemented in ocr/TextExtractor.kt.
  • OCR returns structured result: OCRResult(text, confidence).
  • OCR screen displays extracted text and confidence.
  • User can edit text before saving.

Persistence

  • Room database with DocumentEntity, DocumentDao, and TextLexiqDatabase.
  • Repository saves OCR output and exposes recent documents.
  • Home screen lists saved documents (basic summary).

Document Viewing & Editing

  • Document view screen loads by ID and displays content.
  • Document editing mode implemented with save/discard.
  • Metadata fields added (source, engine, language, tags).
  • DocumentDao and Repository updated for CRUD operations.

Exporting

  • PDF export implemented using iText7.
  • DOCX export implemented using Apache POI.
  • LaTeX code generation added.
  • Export action menu in UI with format selection.

OCR Enhancements

  • Advanced OpenCV preprocessing: Deskewing (minAreaRect), Binarization (Adaptive Threshold), Grayscale.
  • Text block sorting in TextExtractor for reliable reading order.
  • Skew correction implemented.

LLM Optimization

  • TokenOptimizer: Whitespace compression and stopword removal.
  • SmartModelRouter: Floating access logic (On-Device vs Cloud).
  • LLMEngine architecture defined.

⚠️ Partially Implemented / Needs Hardening

OCR & Confidence

  • Confidence is averaged from blocks. Visual overlays missing.
  • Tesseract (tess-two) implementation optional/pending.
  • Auto-language detection pending.

Error Handling

  • Basic errors surfaced in UI, but no retry strategies or offline recovery.

🚧 Remaining Work (Detailed)

1) OCR Future Work

  • Option to switch between ML Kit and Tesseract (tess-two).
  • Support multiple languages and auto-language detection.
  • Add visual OCR overlays for confidence highlighting.

2) LLM Integration [DONE]

  • Design LLMEngine interface (abstraction for On-Device vs Cloud).
  • Implement TokenOptimizer (Stopword removal, context compression).
  • Implement SmartModelRouter (Floating access: Route based on complexity/cost).
  • Implement LlamaCppClient (On-Device Tier).
  • Implement CloudLLMClient (Paid Tier).
  • Integrate LLM features into UI (Summarize, Simplify).

3) Scanner UX

  • Implement auto edge detection interface (OpenCV contors logic exists, need UI overlay).
  • Auto capture when page is stable inside the guide.
  • Manual crop corners (drag handles) instead of slider-only crop.

6) UI & State

  • Create a unified app state (loading/saving indicators, error toasts/snackbars).
  • Add settings screen bindings to actual config/state.
  • Add onboarding/permissions screen.

7) Testing & Stability

  • Unit tests for preprocessing and repository layers.
  • Instrumentation tests for CameraX and OCR flows.
  • Validate OpenCV initialization on device (not just debug).

8) Cross-Platform & Server (Future - Daytona/LeanMCP)

  • Create TextLexiq Server (Python) for Daytona/LeanMCP.
  • Port OpenCV/OCR pipeline to Python (pip package).
  • Create textlexiq-js client (npm package).
  • Deploy MCP Server to LeanMCP.

Notes

  • Current implementation uses ML Kit for OCR. If you prefer tess-two, a new implementation can be added in ocr/TextExtractor.kt behind an interface.
  • OpenCV initialization is performed in TextLexiqApp and may need runtime loading for production builds.