V6: Multi-model — HY3 for Strategist + GPT-OSS-120B for Executor SVG (2-3x faster, higher quality)

interface/app_gradio.py

@@ -33,10 +33,10 @@ load_prefixed_env_file(("OPENROUTER_", "COMFYUI_", "IMAGE_", "PEXELS_", "PIXABAY
 # LLM Client
 # ============================================================
 
-def llm_chat(prompt, system=None, max_tokens=32000, temperature=0.3):
+def llm_chat(prompt, system=None, max_tokens=32000, temperature=0.3, model_override=None):
     import requests
     api_key = os.environ.get("OPENROUTER_API_KEY", "")
-    model = os.environ.get("OPENROUTER_MODEL", "tencent/hy3-preview:free")
+    model = model_override or os.environ.get("OPENROUTER_MODEL", "tencent/hy3-preview:free")
     base_url = os.environ.get("OPENROUTER_BASE_URL", "https://openrouter.ai/api/v1").rstrip("/")
     if not api_key:
         raise RuntimeError("OPENROUTER_API_KEY non configuré dans .env")
@@ -44,12 +44,14 @@ def llm_chat(prompt, system=None, max_tokens=32000, temperature=0.3):
     if system:
         messages.append({"role": "system", "content": system})
     messages.append({"role": "user", "content": prompt})
+    payload = {"model": model, "messages": messages, "temperature": temperature, "max_tokens": max_tokens}
+    # For reasoning models, exclude internal reasoning from output
+    if "hy3" in model or "qwen3" in model:
+        payload["reasoning"] = {"exclude": True}
     r = requests.post(f"{base_url}/chat/completions",
         headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json",
                  "HTTP-Referer": "http://localhost/ppt-master", "X-Title": "ppt-master-gradio"},
-        json={"model": model, "messages": messages, "temperature": temperature,
-              "max_tokens": max_tokens, "reasoning": {"exclude": True}},
-        timeout=600)
+        json=payload, timeout=600)
     if r.status_code >= 400:
         raise RuntimeError(f"OpenRouter erreur {r.status_code}: {r.text[:500]}")
     data = r.json()
@@ -58,6 +60,10 @@ def llm_chat(prompt, system=None, max_tokens=32000, temperature=0.3):
         raise RuntimeError(f"Pas de contenu retourné par le LLM")
     return content
 
+# Model selection: use the best model for each task
+MODEL_STRATEGIST = "tencent/hy3-preview:free"       # Good at planning/reasoning
+MODEL_EXECUTOR = "openai/gpt-oss-120b:free"         # Best at SVG code generation (fast + high quality)
+
 
 # ============================================================
 # STEP 1: STRATEGIST — Generate design_spec + spec_lock + outline
@@ -117,7 +123,7 @@ Rules:
 - page_rhythm: "anchor" for cover/closing, "dense" for data-heavy, "breathing" for impact pages
 - Reply ONLY with raw JSON, no markdown code blocks, no explanation"""
 
-    result = llm_chat(prompt, max_tokens=32000)
+    result = llm_chat(prompt, max_tokens=32000, model_override=MODEL_STRATEGIST)
 
     # Parse JSON
     try:
@@ -317,7 +323,7 @@ DESIGN THIS SLIDE:
 Reply with ONLY the SVG. Start <svg, end </svg>. No explanation."""
 
         try:
-            svg_content = llm_chat(prompt, max_tokens=8000, temperature=0.3)
+            svg_content = llm_chat(prompt, max_tokens=10000, temperature=0.3, model_override=MODEL_EXECUTOR)
 
             # Extract SVG
             svg_start = svg_content.find('<svg')
@@ -439,8 +445,8 @@ def step2_handler(data_json, style, image_mode, progress=gr.Progress()):
 
     # Phase 2: Executor (SVG generation via LLM)
     progress(0.2, desc=f"Génération SVG par LLM (0/{num_slides})...")
-    log.append("━━━ PHASE 2: Executor (SVG via LLM) ━━━")
-    log.append(f"  ⏱️ Estimation: ~60-90s par slide, {num_slides} slides\n")
+    log.append("━━━ PHASE 2: Executor (SVG via GPT-OSS-120B) ━━━")
+    log.append(f"  ⏱️ Estimation: ~30-45s par slide, {num_slides} slides\n")
 
     slide_count = [0]
     def svg_cb(msg):
@@ -505,7 +511,7 @@ with gr.Blocks(title="PPT Master — Générateur IA") as app:
 
         with gr.Tab("2️⃣ Générer le PPTX"):
             gr.Markdown("### 🚀 Pipeline complet: Images + SVG (LLM) + Export")
-            gr.Markdown("⏱️ **Temps estimé**: ~2 min (images) + ~8-12 min (SVG par LLM) + ~30s (export)")
+            gr.Markdown("⏱️ **Temps estimé**: ~2 min (images) + ~5-7 min (SVG via GPT-OSS-120B) + ~30s (export)")
             btn_gen = gr.Button("⚡ Lancer la génération complète", variant="primary", size="lg")
             status2 = gr.Textbox(label="Statut", interactive=False)
             with gr.Row():
@@ -516,16 +522,16 @@ with gr.Blocks(title="PPT Master — Générateur IA") as app:
             gr.Markdown("""
 ## Workflow PPT Master complet
 
-1. **Strategist** → Génère design_spec + spec_lock (palette, typo, icônes, rythme)
+1. **Strategist** (HY3-preview) → design_spec + spec_lock (palette, typo, icônes, rythme)
 2. **Image Acquisition** → ComfyUI GPU ou recherche web
-3. **Executor** → Le LLM génère chaque SVG individuellement en respectant le spec_lock
+3. **Executor** (GPT-OSS-120B) → Génère chaque SVG individuellement, qualité premium
 4. **Post-processing** → finalize_svg.py (icônes, images, rounded rects)
 5. **Export** → svg_to_pptx.py → PPTX natif éditable
 
-**Qualité**: Chaque slide est conçue par le LLM avec dégradés, décorations, typographie variée.
+**Approche multi-modèles**: HY3 pour la planification (raisonnement), GPT-OSS-120B pour le code SVG (rapide + précis).
 Le spec_lock garantit la cohérence couleurs/fonts sur l'ensemble du deck.
 
-**Stack**: OpenRouter (tencent/hy3-preview:free) + ComfyUI local (RTX 5070 Ti)
+**Stack**: OpenRouter (2 modèles gratuits) + ComfyUI local (RTX 5070 Ti)
 """)
 
     btn_plan.click(fn=step1_handler, inputs=[subject, num_slides, style, language, audience, image_mode],