Skip to content

LunarWingVisionServicev1.1.md

Latest commit

 

History

History
48 lines (37 loc) · 2.5 KB

LunarWingVisionServicev1.1.md

File metadata and controls

48 lines (37 loc) · 2.5 KB

LunarWing Vision Service — Design Notes v1.1

Status

  • Approved by: Kageho (2026-05-03)
  • Verdict: Core design is solid, no rework needed
  • Original authors: Starforce (OCR Sidecar MVP), Ruffles (Vision Service unified design)
  • Source: Consolidated from IRC discussion + merged design document

Architecture Summary

  • Two complementary backends: OCR (exact text) + Vision-Language (semantic understanding)
  • Phase 1 MVP: Starforce's OCR sidecar (Tesseract-based)
  • Phase 2: Ruffles' unified Vision Service with VL model integration
  • Single front-door API for all siblings over time

Core Decision

Split between two image task types:

  1. OCR backend — deterministic, exact text extraction (Tesseract)
  2. Vision-Language backend — semantic understanding, comparison, visual reasoning (Qwen3-VL or similar)

Phases

Phase Scope Lead
1 OCR Sidecar MVP (Tesseract, REST API) Sweetiebot
2 VL model integration (Qwen3-VL) TBD
3 Unified front-door API + auto-routing TBD
4 WASM tool interface for siblings TBD

Kageho's Review Addenda (2026-05-03)

Verdict: SOLID and awesome. No structural changes needed. The following are hardening suggestions:

Must-address before Phase 1 ships

  1. Auth on sidecar endpoints — Add bearer token or IP allowlist. Even LAN-only services need basic access control.
  2. Image size limits — Define max payload (recommend ~10MB), return 413 Payload Too Large with clear error body.
  3. Error response spec — Standardize error JSON format so siblings know what failure looks like. Example: {"error": "unsupported_format", "detail": "TIFF files not supported", "code": 415}
  4. Supported image formats — Explicitly declare: PNG, JPEG, WebP, TIFF (at minimum). Document any Tesseract-specific limitations.

Address during Phase 2 spec refinement

  1. Auto-mode tie-breaking — Clarify behavior when prompt contains both OCR and VL keywords. Suggest: score by keyword count per category, highest wins; on tie, prefer OCR (deterministic).
  2. Per-task VL prompt templates — Generic prompts won't cut it. Define templates for: describe, compare, aesthetic evaluation, text-in-image.
  3. SSRF risk on URL image support — Restrict to LAN or maintain an allowlist. No arbitrary external URL fetching.

Quick fixes

  1. Mermaid diagram — Fix duplicate B label.
  2. Phase 4 WASM tool interface — Sketch a rough contract now to avoid API lock-in later.