修正 Ollama fallback 與 PPT vision payload

services/ollama_service.py

@@ -90,17 +90,32 @@ def _mark_unhealthy_best_effort(host: str) -> None:
         logger.debug("[OllamaHost] mark_unhealthy failed for host=%s", host, exc_info=True)
 
 
+def _normalize_host(host: str) -> str:
+    return (host or '').rstrip('/')
+
+
+def _canonical_host_chain() -> List[str]:
+    """Return the approved static fallback chain without duplicates."""
+    chain: List[str] = []
+    for host in (OLLAMA_HOST_PRIMARY, OLLAMA_HOST_SECONDARY, OLLAMA_HOST_FALLBACK):
+        clean = _normalize_host(host)
+        if clean and clean not in chain:
+            chain.append(clean)
+    return chain
+
+
 def _is_unhealthy(host: str) -> bool:
     """檢查 host 是否在 unhealthy TTL 內"""
     import time
     if not host:
         return False
-    ts = _unhealthy_marks.get(host.rstrip('/'))
+    clean_host = _normalize_host(host)
+    ts = _unhealthy_marks.get(clean_host)
     if ts is None:
         return False
     if time.time() - ts >= _UNHEALTHY_TTL:
         # TTL 過期，清除
-        _unhealthy_marks.pop(host.rstrip('/'), None)
+        _unhealthy_marks.pop(clean_host, None)
         return False
     return True
 
@@ -323,12 +338,27 @@ class OllamaService:
         # HOTFIX 三主機 retry 鏈
         attempted_hosts: List[str] = []
         last_error: Optional[str] = None
+        canonical_hosts = _canonical_host_chain()
 
         for attempt in range(3):
-            current_host = self.host  # property 每次 lazy resolve
+            current_host = _normalize_host(self.host)  # property 每次 lazy resolve
             if current_host in attempted_hosts:
-                # 已試過同主機（cache 還沒過期），跳出避免無限迴圈
-                break
+                # 已試過同主機時，若是標準三主機鏈且 caller 沒指定 host，
+                # 改走尚未嘗試的下一台。避免 request timeout(60s) 大於
+                # unhealthy TTL(30s) 時第三輪又 resolve 回 primary，導致 111
+                # final fallback 永遠沒被打到。
+                next_host = None
+                if self._explicit_host is None and current_host in canonical_hosts:
+                    next_host = next((host for host in canonical_hosts if host not in attempted_hosts), None)
+                if not next_host:
+                    # 非標準 host 或 explicit host 維持原行為：跳出避免無限迴圈。
+                    break
+                logger.info(
+                    "[Ollama] resolver returned previously attempted host=%s; forcing next fallback host=%s",
+                    current_host,
+                    next_host,
+                )
+                current_host = next_host
             attempted_hosts.append(current_host)
 
             logger.info(f"[Ollama] 嘗試 #{attempt+1}/3 host={current_host} model={model} timeout={request_timeout}s")

services/ppt_vision_service.py

@@ -23,6 +23,7 @@ import base64
 import logging
 import shutil
 import threading
+from io import BytesIO
 from dataclasses import dataclass, field
 from typing import Optional, Dict, Any, List, Sequence
 
@@ -32,7 +33,9 @@ logger = logging.getLogger(__name__)
 # Feature flag + 配置
 # ─────────────────────────────────────────────────────────────────────────────
 PPT_VISION_MODEL = os.getenv('PPT_VISION_MODEL', 'minicpm-v:latest')
-PPT_VISION_TIMEOUT = int(os.getenv('PPT_VISION_TIMEOUT', '60'))
+PPT_VISION_TIMEOUT = int(os.getenv('PPT_VISION_TIMEOUT', '45'))
+PPT_VISION_IMAGE_MAX_EDGE = int(os.getenv('PPT_VISION_IMAGE_MAX_EDGE', '1280'))
+PPT_VISION_IMAGE_QUALITY = int(os.getenv('PPT_VISION_IMAGE_QUALITY', '82'))
 _AUDIT_LOCK = threading.Lock()
 _LAST_AUDIT_RUN: Dict[str, Any] | None = None
 
@@ -105,6 +108,27 @@ class PPTVisionService:
     def is_available(self) -> bool:
         return is_ppt_vision_enabled()
 
+    def _encode_image_for_vision(self, image_path: str) -> str:
+        """Compress slide screenshots before sending to Ollama vision."""
+        try:
+            from PIL import Image
+
+            with Image.open(image_path) as im:
+                image = im.convert('RGB')
+                image.thumbnail((PPT_VISION_IMAGE_MAX_EDGE, PPT_VISION_IMAGE_MAX_EDGE))
+                buffer = BytesIO()
+                image.save(
+                    buffer,
+                    format='JPEG',
+                    quality=max(50, min(PPT_VISION_IMAGE_QUALITY, 95)),
+                    optimize=True,
+                )
+                return base64.b64encode(buffer.getvalue()).decode('ascii')
+        except Exception:
+            # Pillow is an optimization, not a hard dependency for the vision path.
+            with open(image_path, 'rb') as f:
+                return base64.b64encode(f.read()).decode('ascii')
+
     def check_ppt_file(self, pptx_path: str, max_slides: int = 5) -> Dict[str, Any]:
         """檢查整份 .pptx — Phase 26 整合到 PPT 生成流程。
 
@@ -330,11 +354,9 @@ class PPTVisionService:
                 error=f'image not found: {image_path}',
             )
 
-        # 讀檔並 base64 編碼
+        # 讀檔並 base64 編碼；可用時先壓縮縮圖，避免 Ollama vision 被大圖拖慢。
         try:
-            with open(image_path, 'rb') as f:
-                img_bytes = f.read()
-            img_b64 = base64.b64encode(img_bytes).decode('ascii')
+            img_b64 = self._encode_image_for_vision(image_path)
         except Exception as e:
             return VisionResult(
                 success=False,
@@ -351,7 +373,8 @@ class PPTVisionService:
                 system_prompt=PPT_VISION_SYSTEM_PROMPT,
                 temperature=0.2,
                 timeout=PPT_VISION_TIMEOUT,
-                options={'num_predict': 512},
+                keep_alive='5m',
+                options={'num_predict': 256},
                 images=[img_b64],
             )
             duration_ms = int((time.monotonic() - start) * 1000)