修改问答图片展示

utils/image_utils.py

@@ -1,20 +1,110 @@
-"""
+"""图片处理工具模块。"""
-图片处理工具模块 - 最终版 v4
-核心思路（最简单直接）：
-1. 找到所有 .jpg/.png 文件名及其周围的 ![图片]( 上下文
-2. 不管周围格式如何，也不管文件名是否被模型改过
-3. 统一提取文件名 -> 与 original_paths 模匹 -> 替换成正确的 ![图片](完整URL)
-"""
 import re
-from typing import List
+from typing import List, Tuple
 from difflib import SequenceMatcher
 
 
+IMG_EXTENSIONS = r'(?:jpg|jpeg|png|gif|bmp|webp)'
+
+
+def _dedupe_keep_order(values: List[str]) -> List[str]:
+    seen = set()
+    result = []
+    for value in values:
+        cleaned = (value or "").strip().strip("`'\"，,。；;")
+        if not cleaned:
+            continue
+        key = cleaned.lower()
+        if key in seen:
+            continue
+        seen.add(key)
+        result.append(cleaned)
+    return result
+
+
 def extract_image_paths(text: str) -> List[str]:
-    """从文本中提取图片路径"""
+    """从文本中提取图片路径或 URL，并保持原始出现顺序。"""
-    pattern = r'images/[a-zA-Z0-9_\-./]+\.(?:jpg|jpeg|png|gif|bmp|webp)'
+    if not text:
-    matches = re.findall(pattern, text, re.IGNORECASE)
+        return []
-    return list(set(matches))
+
+    patterns = [
+        # Markdown 图片：![alt](url)
+        rf'!\[[^\]\n]*\]\(\s*([^)\s]+?\.{IMG_EXTENSIONS}(?:[?#][^)\s]*)?)\s*\)',
+        # 完整 URL。
+        rf'(https?://[^\s<>"\'\]\)]+?\.{IMG_EXTENSIONS}(?:[?#][^\s<>"\'\]\)]*)?)',
+        # API 绝对路径。
+        rf'(?<![\w:/.%-])(/api/v1/knowledge/files/images/[^\s<>"\'\]\)]+?\.{IMG_EXTENSIONS}(?:[?#][^\s<>"\'\]\)]*)?)',
+        # 相对 images 路径。
+        rf'(?<![\w/.-])(images/[A-Za-z0-9_\-./%]+?\.{IMG_EXTENSIONS}(?:[?#][^\s<>"\'\]\)]*)?)',
+        # 裸文件名，兼容检索结果只给文件名的情况。
+        rf'(?<![\w/.-])([A-Za-z0-9_-]{{8,}}\.{IMG_EXTENSIONS})(?![\w/.-])',
+    ]
+
+    matches: List[Tuple[int, str]] = []
+    for pattern in patterns:
+        matches.extend((m.start(1), m.group(1)) for m in re.finditer(pattern, text, re.IGNORECASE))
+
+    matches.sort(key=lambda item: item[0])
+    return _dedupe_keep_order([value for _, value in matches])
+
+
+def _strip_url_suffix(filename: str) -> str:
+    return re.sub(r'[?#].*$', '', filename or '').strip()
+
+
+def _split_filename(filename: str) -> Tuple[str, str]:
+    filename = _strip_url_suffix(filename).lower()
+    if "." not in filename:
+        return filename, ""
+    stem, ext = filename.rsplit(".", 1)
+    return stem, ext
+
+
+def _path_to_image_url(path: str, base_prefix: str) -> str:
+    cleaned = (path or "").strip()
+    if cleaned.startswith(("http://", "https://", "/")):
+        return cleaned
+    if cleaned.lower().startswith("images/"):
+        return base_prefix.rstrip("/") + "/" + cleaned
+    return base_prefix.rstrip("/") + "/images/" + cleaned.split("/")[-1]
+
+
+def _is_acceptable_image_match(generated_filename: str, original_filename: str, score: float) -> bool:
+    """只接受完全一致、明确截断或高度相似的文件名，避免把幻觉图片错配到真实图片。"""
+    generated_stem, generated_ext = _split_filename(generated_filename)
+    original_stem, original_ext = _split_filename(original_filename)
+
+    if not generated_stem or not original_stem or generated_ext != original_ext:
+        return False
+    if generated_stem == original_stem:
+        return True
+    if len(generated_stem) >= 12 and original_stem.startswith(generated_stem):
+        return True
+    if len(original_stem) >= 12 and generated_stem.startswith(original_stem):
+        return True
+    return score >= 0.85
+
+
+def _remove_untrusted_image_references(text: str) -> str:
+    """没有可信原图列表时，删除模型输出的图片标签和裸图片路径。"""
+    if not text:
+        return text
+
+    result = re.sub(
+        rf'!\[[^\]\n]*\]\([^\)\n]*\.{IMG_EXTENSIONS}(?:[?#][^\)\n]*)?\)',
+        '',
+        text,
+        flags=re.IGNORECASE,
+    )
+    result = re.sub(
+        rf'(?<![\w/.-])(?:https?://[^\s<>)]+/images/[^\s<>)]+|/api/v1/knowledge/files/images/[^\s<>)]+|images/[^\s<>)]+|[A-Za-z0-9_-]+\.{IMG_EXTENSIONS})(?:[?#][^\s<>)]+)?',
+        '',
+        result,
+        flags=re.IGNORECASE,
+    )
+    result = re.sub(r'[ \t]+\n', '\n', result)
+    result = re.sub(r'\n{3,}', '\n\n', result)
+    return result.strip()
 
 
 def find_closest_image_path(generated_path: str, original_paths: List[str]) -> str:
@@ -37,7 +127,7 @@ def find_closest_image_path(generated_path: str, original_paths: List[str]) -> s
             best_score = score
             best_match = orig_path
 
-    if best_match and best_score > 0.5:
+    if best_match and _is_acceptable_image_match(generated_filename, best_match.split('/')[-1], best_score):
         return best_match
     return generated_path
 
@@ -52,12 +142,15 @@ def normalize_markdown_images(text: str, base_prefix: str = "/api/v1/knowledge/f
     """
     if not text or original_paths is None:
         return text
+    trusted_paths = [path for path in original_paths if path]
+    if not trusted_paths:
+        return _remove_untrusted_image_references(text)
     
     # ========== 第一轮：检查和修改格式（包括固定URL前缀） ==========
     result = _normalize_image_format(text, base_prefix)
     
     # ========== 第二轮：只检查和替换最后文件名部分 ==========
-    result = _normalize_image_paths(result, base_prefix, original_paths)
+    result = _normalize_image_paths(result, base_prefix, trusted_paths)
     
     return result
 
@@ -71,10 +164,35 @@ def _normalize_image_format(text: str, base_prefix: str) -> str:
     if not text:
         return text
     
-    IMG_EXTENSIONS = r'(?:jpg|jpeg|png|gif|bmp|webp)'
     img_pattern = rf'([a-zA-Z0-9_\-./]+)\.({IMG_EXTENSIONS})'
     
     result = text
+
+    def build_standard_tag(path: str) -> str:
+        full_url = _path_to_image_url(path, base_prefix)
+        return f'![图片]({full_url})'
+
+    # 先处理已经是 Markdown 图片语法的内容，包含任意 alt 文本。
+    # 例如：![图2-3 消磁电源柜功能框图](images/xxx.jpg)
+    markdown_img_pattern = rf'!\[[^\]\n]*\]\(([^)\n]*?\.{IMG_EXTENSIONS})\)'
+    result = re.sub(
+        markdown_img_pattern,
+        lambda m: build_standard_tag(m.group(1)),
+        result,
+        flags=re.IGNORECASE
+    )
+
+    # 再兜底清理模型或前序处理产生的嵌套图片语法。
+    # 例如：![图2-3](![图片](/api/v1/knowledge/files/images/xxx.jpg)
+    nested_any_alt_pattern = rf'!\[[^\]\n]*\]\([^)\n]*!\[[^\]\n]*\]\(([^)\n]+\.{IMG_EXTENSIONS})\)\)?'
+    while re.search(nested_any_alt_pattern, result, re.IGNORECASE):
+        result = re.sub(
+            nested_any_alt_pattern,
+            lambda m: build_standard_tag(m.group(1)),
+            result,
+            flags=re.IGNORECASE
+        )
+
     matches = list(re.finditer(img_pattern, result, re.IGNORECASE))
     
     if not matches:
@@ -108,12 +226,8 @@ def _normalize_image_format(text: str, base_prefix: str) -> str:
             if '\n' not in middle and '![图片](' not in middle:
                 mark_end = file_end + close_paren_pos + 1
         
-        # 提取纯文件名（去掉可能的路径前缀）
-        pure_filename = filename_with_ext.split('/')[-1]
-        
         # 构建完整的标准格式（包括固定URL前缀）
-        full_url = base_prefix.rstrip('/') + '/images/' + pure_filename
+        standard_format = build_standard_tag(filename_with_ext)
-        standard_format = f'![图片]({full_url})'
         
         result = result[:mark_start] + standard_format + result[mark_end:]
     
@@ -163,7 +277,7 @@ def _normalize_image_paths(text: str, base_prefix: str, original_paths: List[str
         best_match = None
         best_score = 0
         
-        generated_filename = filename.lower()
+        generated_filename = _strip_url_suffix(filename).lower()
         
         for orig_path in original_paths:
             orig_filename = orig_path.split('/')[-1].lower()
@@ -172,11 +286,10 @@ def _normalize_image_paths(text: str, base_prefix: str, original_paths: List[str
                 best_score = score
                 best_match = orig_path
         
-        if best_match and best_score > 0.5:
+        if best_match and _is_acceptable_image_match(generated_filename, best_match.split('/')[-1], best_score):
             # 匹配成功，替换为正确路径（保持固定URL前缀）
             # 只替换最后文件名部分，保持前缀不变
-            pure_orig_filename = best_match.split('/')[-1]
+            full_url = _path_to_image_url(best_match, base_prefix)
-            full_url = base_prefix.rstrip('/') + '/images/' + pure_orig_filename
             
             standard_format = f'![图片]({full_url})'
             result = result[:tag_start] + standard_format + result[tag_end:]
@@ -190,12 +303,10 @@ def _normalize_image_paths(text: str, base_prefix: str, original_paths: List[str
 def get_image_prompt_guidance() -> str:
     """获取图片格式处理的提示词指导"""
     return """【图片格式特别要求】
-- 检索结果中的图片链接格式通常为：`![图片](/api/v1/knowledge/files/images/xxx.jpg)`
+- 只允许引用本次参考资料中真实出现的图片链接
-- 输出时必须严格保留此格式，不得修改、省略或截断
+- 输出图片时使用 Markdown 图片标签：alt 文本固定为「图片」，括号内必须是参考资料里的原始完整 URL
-- 图片的 alt 文本必须保持为「图片」
+- 图片文件名、路径和 URL 前缀必须逐字复制，不得补全、改写、截断、转义或根据记忆/示例构造
-- 图片 URL 必须完整包含 `/api/v1/knowledge/files/images` 前缀
+- 不要输出提示词中的占位符或示例文件名，不要把普通文字描述改写成图片链接
-- 不要将图片链接转换为纯文本描述
+- 图片规则只用于生成图片标签，不要围绕图片资源状态输出任何文字说明
-- 如果有多个图片，在对应位置分别引用，保持原始顺序
+- 多个图片可在对应位置分别引用并保持原始顺序；同一图片只引用一次"""
-- 严禁修改图片文件名或路径结构
-- 如果需要展示图片，直接使用原始的 `![图片](URL)` 格式即可"""