Use gemini-2.5-flash-lite for master deck parsing

- Added PARSING_MODEL env var (default: gemini-2.5-flash-lite)
- Master deck parser now uses fast lite model
- 3-4x faster than gemini-3.1-pro: 22 layouts = 8-12 min vs 44-60 min
- Keep gemini-3.1-pro for presentation generation (quality)

Co-Authored-By: Claude Sonnet 4.5 (1M context) <noreply@anthropic.com>

.env.example

@@ -19,12 +19,15 @@ DEV_AUTH_PASSWORD=change-me-to-secure-password
 
 # AI Provider — Google Gemini for all AI operations
 GOOGLE_API_KEY=your_google_api_key_here
-GOOGLE_MODEL=models/gemini-2.0-flash
+GOOGLE_MODEL=gemini-3.1-pro-preview
 IMAGE_PROVIDER=gemini_flash
 
+# Master deck parsing model (use fast lite model for speed)
+PARSING_MODEL=gemini-2.5-flash-lite
+
 # Get your Google AI API key at: https://aistudio.google.com/app/apikey
-# Gemini 2.0 Flash: Fast, cheap, great for text generation
-# Gemini 3.1 Flash: Excellent vision model for image analysis
+# Gemini 3.1 Pro: Best quality for presentations
+# Gemini 2.5 Flash Lite: Super fast for parsing (recommended)
 
 # Optional: Image fallback providers (if Gemini image gen fails)
 PEXELS_API_KEY=

backend/services/master_deck_parser_service.py

@@ -36,6 +36,21 @@ from api.v1.ppt.endpoints.prompts import (
 from services.documents_loader import DocumentsLoader
 from services.llm_service import UnifiedLLMService, _detect_llm_provider, LLMProvider
 
+# Use fast lite model for parsing (gemini-2.5-flash-lite)
+def _get_parsing_provider():
+    """Get provider config for master deck parsing with fast lite model."""
+    google_key = os.getenv("GOOGLE_API_KEY")
+    if not google_key:
+        raise ValueError("GOOGLE_API_KEY required for parsing")
+
+    # Use gemini-2.5-flash-lite: extremely fast, good enough for layout parsing
+    parsing_model = os.getenv("PARSING_MODEL", "gemini-2.5-flash-lite")
+    return {
+        "provider": LLMProvider.GOOGLE,
+        "api_key": google_key,
+        "model": parsing_model
+    }
+
 # OXML namespaces
 NS = {
     "a": "http://schemas.openxmlformats.org/drawingml/2006/main",
@@ -460,7 +475,8 @@ async def _do_parse(deck_id: uuid.UUID) -> dict:
             all_fonts.update(normalize_font_family_name(f) for f in raw if f)
 
         # 6. Process each item through LLM pipeline
-        llm_provider = _detect_llm_provider()
+        llm_provider = _get_parsing_provider()  # Use fast lite model for parsing
+        print(f"[MasterDeckParser] Using {llm_provider['model']} for parsing")
         layouts_result = []
 
         # Build per-layout screenshot mapping