增加多模态模型识别功能，可以正确显示置信度了。

exp_multimodal/vlm_classifier.py

@@ -1,10 +1,39 @@
 import json
+import re
 from typing import Dict, List, Optional
 
 from .ollama_client import chat_vision
 from .prompts import build_closedset_prompt
 
 
+def _extract_json_obj(text: str):
+    """尽量从返回文本中提取出合法 JSON（支持代码块/前后说明）。"""
+    if not isinstance(text, str):
+        return None
+    s = text.strip()
+    # 去掉 Markdown 代码块包裹
+    if s.startswith("```"):
+        s = re.sub(r"^```[a-zA-Z]*\n|\n```$", "", s).strip()
+    # 直接解析
+    try:
+        obj = json.loads(s)
+        if isinstance(obj, dict):
+            return obj
+    except Exception:
+        pass
+    # 提取第一个花括号对象
+    m = re.search(r"\{[\s\S]*?\}", s)
+    if m:
+        frag = m.group(0)
+        try:
+            obj = json.loads(frag)
+            if isinstance(obj, dict):
+                return obj
+        except Exception:
+            pass
+    return None
+
+
 def classify_image(
     image_path: str,
     labels: List[str],
@@ -25,38 +54,37 @@ def classify_image(
 
     print("[VLM] Calling chat_vision...")
     text = chat_vision(prompt, [image_path], temperature=0.1)
-    _preview_text = (text[:120]).replace("\n", " ")
+    _preview_text = (str(text)[:200]).replace("\n", " ")
-    print(f"[VLM] Received text length={len(text)} preview={_preview_text}...")
+    print(f"[VLM] Received text length={len(str(text))} preview={_preview_text}...")
 
-    # 主路径：严格 JSON 解析
+    # 解析 JSON（更健壮）
-    try:
+    obj = _extract_json_obj(str(text))
-        obj = json.loads(text)
+    if obj and "label" in obj:
-        if isinstance(obj, dict) and "label" in obj:
+        label = obj.get("label", "")
-            label = obj.get("label", "")
+        # 若输出不在封闭集，尝试包含匹配；否则 Unknown
-            # 纠偏：若输出不在封闭集，尝试包含匹配；否则 Unknown
+        if label not in labels:
-            if label not in labels:
+            for lb in labels:
-                for lb in labels:
+                if lb in str(label):
-                    if lb in str(label):
+                    label = lb
-                        label = lb
+                    break
-                        break
+            else:
-                else:
+                label = "Unknown"
-                    label = "Unknown"
+        conf = obj.get("confidence", 0.0)
-            conf = obj.get("confidence", 0.0)
+        try:
-            try:
+            conf = float(conf)
-                conf = float(conf)
+        except Exception:
-            except Exception:
+            conf = 0.0
-                conf = 0.0
+        result = {"label": label, "confidence": conf}
-            result = {"label": label, "confidence": conf}
+        print(f"[VLM] Parsed JSON result={result}")
-            print(f"[VLM] Parsed JSON result={result}")
+        return result
-            return result
-    except Exception as e:
-        print(f"[VLM] JSON parse failed: {e}")
 
-    # 回退：字符串包含匹配
+    print("[VLM] JSON parse failed or missing 'label', enter fallback")
+
+    # 回退：字符串包含匹配（无法确定置信度时返回 0.0）
     for lb in labels:
-        if lb in text:
+        if lb in str(text):
             print(f"[VLM] Fallback match label={lb}")
-            return {"label": lb, "confidence": 0.5}
+            return {"label": lb, "confidence": 0.0}
 
     print("[VLM] No match, return Unknown")
     return {"label": "Unknown", "confidence": 0.0}