feat(kb-deep-analyzer): Smart subtitle chunking strategy

## 改进内容 ### 动态分段 - 根据字符数动态计算分段数 - 每个chunk最多3000字符 - 限制最小3段，最多20段 ### 自然断点分割 - 优先级：段落(双换行) > 句末(。\n) > 句号 > 换行 - 不截断，保持语义完整 ### 时间估算 - 用字符比例估算每段时间 - 长视频时间轴更准确
2026-02-10 16:41:15 +08:00
parent 3cfd36cf4e
commit 08066334c4
1 changed files with 105 additions and 31 deletions
@@ -269,27 +269,62 @@ class DeepAnalyzer:
        return result

    def _analyze_in_chunks(self, title: str, subtitle: str, duration: int) -> AnalysisResult:
-        """分段分析长视频"""
+        """分段分析长视频 - 动态分段策略"""
        result = AnalysisResult()

-        # 确定分段数
-        if duration < 3600:  # <60分钟
-            num_segments = 3
-        elif duration < 7200:  # 60-120分钟
-            num_segments = 4
-        else:  # >120分钟
-            num_segments = 6
+        # 常量配置
+        CHUNK_MAX_CHARS = 3000  # 每个chunk最大字符数
+        MIN_CHUNKS = 3          # 最少分段数
+        MAX_API_CHUNKS = 20     # 最多API调用次数

-        # 生成分段摘要
-        segment_duration = duration // num_segments
+        total_chars = len(subtitle)
+
+        # 1. 动态计算分段数
+        # 计算需要多少个chunk（每个不超过3000字）
+        min_required = total_chars // CHUNK_MAX_CHARS + 1
+        # 根据时长限制最大分段
+        if duration < 1800:      # <30分钟: 最多6段
+            max_chunks = 6
+        elif duration < 7200:    # <2小时: 最多12段
+            max_chunks = 12
+        else:                    # >2小时: 最多20段
+            max_chunks = 20
+
+        num_segments = min(max(min_required, MIN_CHUNKS), MAX_API_CHUNKS)
+
+        print(f"[DeepAnalyzer] 字幕{total_chars}字符，分{num_segments}段分析")
+
+        # 2. 按字符数分段（保持语义完整）
+        chunk_size = total_chars // num_segments
+        chunks = []
+        chunk_boundaries = []  # 记录每段的字符边界
+
+        for i in range(num_segments):
+            start = i * chunk_size
+            if i == num_segments - 1:
+                # 最后一段取到结尾
+                end = total_chars
+                chunk_text = subtitle[start:]
+            else:
+                # 在自然断点处分割
+                end = self._find_breakpoint(subtitle, start, start + chunk_size)
+                chunk_text = subtitle[start:end]
+            chunks.append(chunk_text)
+            chunk_boundaries.append((start, end))
+
+        print(f"[DeepAnalyzer] 各段字符: {[len(c) for c in chunks]}")
+
+        # 3. 用字符比例估算时间
+        char_per_second = total_chars / duration if duration > 0 else 1
+
+        # 合并结果
        chunk_summaries = []

-        # 简单按字符数分割字幕
-        subtitle_chunks = self._chunk_subtitle(subtitle, num_segments)
-
-        for i, chunk_text in enumerate(subtitle_chunks):
-            start_sec = i * segment_duration
-            end_sec = min((i + 1) * segment_duration, duration)
+        for i, chunk_text in enumerate(chunks):
+            start_char, end_char = chunk_boundaries[i]
+            # 用字符位置估算时间
+            start_sec = max(0, int(start_char / char_per_second))
+            end_sec = min(duration, int(end_char / char_per_second))
            start_time = self._format_time(start_sec)
            end_time = self._format_time(end_sec)

@@ -299,6 +334,8 @@ class DeepAnalyzer:
                'index': i,
                'start_time': start_time,
                'end_time': end_time,
+                'start_sec': start_sec,
+                'end_sec': end_sec,
                'summary': chunk_summary
            })

@@ -308,6 +345,58 @@ class DeepAnalyzer:

        return result

+    def _find_breakpoint(self, text: str, start: int, preferred_end: int) -> int:
+        """
+        在自然断点处分割文本
+
+        优先级：
+        1. 双换行（段落分隔）
+        2. 句号+换行（句末）
+        3. 句号（句末）
+        4. 换行（行末）
+        5. 回到preferred_end
+
+        Args:
+            text: 完整文本
+            start: 起始位置
+            preferred_end: 理想结束位置
+
+        Returns:
+            实际结束位置
+        """
+        if not text or start >= len(text):
+            return start
+
+        # 搜索范围：理想位置前后各500字符
+        search_start = max(start, preferred_end - 500)
+        search_end = min(len(text), preferred_end + 500)
+
+        # 优先级1: 双换行
+        double_newline = text.rfind('\n\n', search_start, search_end)
+        if double_newline != -1:
+            return double_newline
+
+        # 优先级2: 句号+换行
+        period_newline = text.rfind('。\n', search_start, search_end)
+        if period_newline != -1:
+            return period_newline + 1  # 包含句号
+
+        # 优先级3: 句号（中文或英文）
+        last_period = max(
+            text.rfind('。', search_start, search_end),
+            text.rfind('.', search_start, search_end)
+        )
+        if last_period != -1:
+            return last_period + 1  # 包含句号
+
+        # 优先级4: 换行
+        last_newline = text.rfind('\n', search_start, search_end)
+        if last_newline != -1:
+            return last_newline
+
+        # 没找到自然断点，返回理想位置
+        return preferred_end
+
    def _analyze_chunk(self, chunk_text: str, chunk_index: int) -> str:
        """分析单个分片"""
        text_sample = self._get_text_sample(chunk_text, max_chars=3000)
@@ -435,21 +524,6 @@ class DeepAnalyzer:
            text[-part_size:]
        )

-    def _chunk_subtitle(self, subtitle: str, num_chunks: int) -> List[str]:
-        """简单分割字幕"""
-        if not subtitle:
-            return [""] * num_chunks
-
-        chunk_size = len(subtitle) // num_chunks
-        chunks = []
-
-        for i in range(num_chunks):
-            start = i * chunk_size
-            end = start + chunk_size if i < num_chunks - 1 else len(subtitle)
-            chunks.append(subtitle[start:end])
-
-        return chunks
-
    def _generate_simple_timeline_from_duration(self, title: str, duration: int) -> List[Dict]:
        """生成简单时间轴"""
        segments = 2 if duration < 300 else 3