Slice 3: Real OCR Pipeline (Horizontal Text)

[emcifuentesc]

Parent

• PRD: NihongOCR MVP #1 (PRD: NihongOCR MVP)

What to build

Replace the Python sidecar stub with the real OCR pipeline: PaddleOCR for text detection, Sudachi for tokenization and Kana reading extraction, and pykakasi for Romaji conversion. The Python worker receives an image file path via NDJSON stdin and emits a real Reading JSON with lines and tokens. The Rust Sidecar Manager and TypeScript Reading Engine from previous slices consume this real data without changes.

This slice is purely about the Python NLP pipeline — no UI or Rust changes beyond wiring real data through the existing protocol.

Acceptance criteria

• PaddleOCR integrated in Python sidecar with bundled models
• Sudachi integrated with bundled dictionary for tokenization and reading extraction
• pykakasi integrated for Romaji conversion
• Python worker processes real image → real Reading JSON (horizontal text)
• Token schema matches agreed format: surface, kana, romaji per token
• Punctuation included as tokens
• Pure-Kana tokens have kana equal to surface and romaji transliterated
• OCR confidence scores forwarded per token
• Reading schema includes orientation field (default horizontal)
• OCR Worker integration tests verify output shape and approximate correctness on sample images

Blocked by

• Slice 2: Region Capture → Reading Window with Dummy Data #3 (Slice 2: Region Capture → Reading Window with Dummy Data)

Testing notes

Write integration tests using a small suite of sample images (horizontal Japanese, mixed text, empty image). Verify output JSON shape and presence of key tokens rather than exact boundaries (Sudachi tokenization may vary by version).

[emcifuentesc]

This was generated by AI during triage.

Agent Brief

Goal: Replace the Python sidecar stub with the real OCR pipeline: PaddleOCR for text detection, Sudachi for tokenization and Kana reading extraction, and pykakasi for Romaji conversion. The Python worker receives an image file path via NDJSON stdin and emits a real Reading JSON with lines and tokens.

Context:

• Slice 1 (Slice 1: Project Scaffold + Sidecar Lifecycle #2) complete: Tauri scaffold, sidecar protocol, splash screen
• Slice 2 (Slice 2: Region Capture → Reading Window with Dummy Data #3) complete: Capture flow, Reading Engine UI with dummy data
• Domain glossary and ADRs in repo root
• PRD: PRD: NihongOCR MVP #1

Current state:

• python-worker/main.py — stub that returns hardcoded Reading JSON
• Rust SidecarManager sends {"type": "ocr", "image_path": "..."} and expects {"type": "reading", ...}
• Reading Engine renders the JSON schema already agreed upon

What to build:

1. Install PaddleOCR, Sudachi, and pykakasi in the Python environment
2. Load PaddleOCR models and Sudachi dictionary at startup (background warmup)
3. Implement OCR pipeline in main.py:
   • Receive image path from stdin
   • PaddleOCR.detect() → text regions with confidence scores
   • For each region: Sudachi.tokenize() → tokens with surface + reading
   • pykakasi.convert() → Romaji for each token
   • Build Reading JSON with orientation (default horizontal), lines, tokens
4. Token schema: surface, kana, romaji per token
5. Include punctuation as tokens
6. Pure-Kana tokens: kana = surface, romaji = transliteration
7. Forward OCR confidence scores per token
8. Integration tests with sample images

Key acceptance criteria:

• PaddleOCR detects Japanese text from real images
• Sudachi tokenizes and extracts Kana readings
• pykakasi produces correct Romaji
• Output JSON matches agreed Reading schema
• Integration tests pass on sample images
• No changes needed to Rust or TypeScript (drop-in replacement)

Blockers: None — Slice 2 is complete.