feat(classroom): P0-3 真实置信度与低置信度质量门控

网关: transcribe 置信度估算（文本长度基线0.7渐近0.95/fallback 0.3）；vision 截断标志透传且置信度减半

本地: estimateAsrConfidence（清洗前后长度比代理信号）；流式final/flush与按段转写均携带confidence

UI: UnifiedTimeline 低置信度（<0.55）弱化+角标；LiveTranscript/TranscriptEntry 增加confidence字段

Author: Aparencia

client/electron/ai/local-asr/SherpaAsrService.ts

@@ -21,7 +21,10 @@
 import * as path from 'path';
 import * as os from 'os';
 import { logger } from '../../logger.js';
-import { cleanAsrResult } from '../../../src/lib/capture/asrFilters.js';
+import {
+  cleanAsrResult,
+  estimateAsrConfidence,
+} from '../../../src/lib/capture/asrFilters.js';
 import {
   getLocalAsrConfig,
   getModelDir,
@@ -210,7 +213,7 @@ export function resetAvailabilityCache(): void {
 export async function transcribeStreaming(
   pcmData: Float32Array,
   hotwords?: string,
-): Promise<{ text: string; engine: 'zipformer'; durationMs: number }> {
+): Promise<{ text: string; confidence: number; engine: 'zipformer'; durationMs: number }> {
   const startTime = Date.now();
 
   const recognizer = getOnlineRecognizer();
@@ -236,15 +239,17 @@ export async function transcribeStreaming(
       recognizer.decode(stream);
     }
 
-    const result = recognizer.getResult(stream);
+    const rawText = recognizer.getResult(stream).text ?? '';
     // 输出后处理：相邻重复压缩 + 幻觉过滤
     // P1-1 两遍重打分接入点：此处为本地按段转写最终文本出口，
     // SenseVoice 重打分将在此处对 text 做句末复核（高置信度者胜出）
-    const text = cleanAsrResult(result.text ?? '');
+    const text = cleanAsrResult(rawText);
+    // P0-3：置信度估算（清洗前后长度比代理信号，供 UI 低置信度标记）
+    const confidence = estimateAsrConfidence(rawText, text);
     const durationMs = Date.now() - startTime;
 
     logger.debug(`[LocalASR] Zipformer transcribe: ${text.length} chars, ${durationMs}ms`);
-    return { text, engine: 'zipformer', durationMs };
+    return { text, confidence, engine: 'zipformer', durationMs };
   } finally {
     stream.free?.();
   }
@@ -259,7 +264,7 @@ export async function transcribeStreaming(
 export async function transcribeLocal(
   audioBase64: string,
   options?: { language?: string; sampleRate?: number; channels?: number; hotwords?: string },
-): Promise<{ text: string; language: string; durationMs: number }> {
+): Promise<{ text: string; language: string; confidence: number; durationMs: number }> {
   const config = getLocalAsrConfig();
   const language = options?.language ?? config.language;
 
@@ -275,5 +280,5 @@ export async function transcribeLocal(
   const pcmData = new Float32Array(rawBytes.buffer, rawBytes.byteOffset, rawBytes.byteLength / 4);
 
   const result = await transcribeStreaming(pcmData, options?.hotwords);
-  return { text: result.text, language, durationMs: result.durationMs };
+  return { text: result.text, language, confidence: result.confidence, durationMs: result.durationMs };
 }

client/electron/ai/local-asr/streamingAsr.ts

@@ -12,7 +12,7 @@
 
 import type { BrowserWindow } from 'electron';
 import { logger } from '../../logger.js';
-import { cleanAsrResult, computeRms, SILENCE_RMS_THRESHOLD } from '../../../src/lib/capture/asrFilters.js';
+import { cleanAsrResult, computeRms, estimateAsrConfidence, SILENCE_RMS_THRESHOLD } from '../../../src/lib/capture/asrFilters.js';
 import { getOnlineRecognizer, feedWaveform, type OnlineStream } from './SherpaAsrService.js';
 
 /** partial 推送节流：两次 partial 推送的最小间隔（ms） */
@@ -148,14 +148,16 @@ export function feedStreamingAsr(audioBuffer: ArrayBuffer, sampleRate?: number):
     // 端点检测：断句 → 推送 final 并以最新热词重建流（P0-6 热词生效点）
     if (recognizer.isEndpoint(_stream)) {
       // 输出后处理：相邻重复压缩 + 幻觉过滤（静音段重复输出防护）
-      const finalText = cleanAsrResult(recognizer.getResult(_stream).text ?? '');
+      const rawText = recognizer.getResult(_stream).text ?? '';
+      const finalText = cleanAsrResult(rawText);
+      const confidence = estimateAsrConfidence(rawText, finalText);
       _lastFinalText = finalText;
       // 重建流：以会话最新热词 createStream（热词变化无需重启即可生效）
       _stream = recognizer.createStream(_latestHotwords);
       _lastPartialText = '';
       _lastPartialEmitAt = 0;
       if (finalText) {
-        emit('asr_stream_final', { text: finalText, timestamp: Date.now() });
+        emit('asr_stream_final', { text: finalText, confidence, timestamp: Date.now() });
       }
       return;
     }
@@ -190,13 +192,13 @@ export function stopStreamingAsr(): void {
       // flush 尾句：仅在窗口存活且流内已有可交付文本时推送
       if (_win && !_win.isDestroyed()) {
         const recognizer = getOnlineRecognizer();
-        const tailText = recognizer
-          ? cleanAsrResult(recognizer.getResult(_stream).text ?? '')
-          : '';
+        const rawTail = recognizer ? recognizer.getResult(_stream).text ?? '' : '';
+        const tailText = rawTail ? cleanAsrResult(rawTail) : '';
         // P0-4 flush 去重：尾句与最近一次 final 完全一致时不再推送
         // （端点已推送过该句，flush 重复上屏是停止瞬间重复的兜底场景）
         if (tailText && tailText !== _lastFinalText) {
-          emit('asr_stream_final', { text: tailText, timestamp: Date.now() });
+          const confidence = estimateAsrConfidence(rawTail, tailText);
+          emit('asr_stream_final', { text: tailText, confidence, timestamp: Date.now() });
         }
       }
       // 新版（1.13+）流对象无 free 方法（句柄由 GC 回收），可选调用

client/src/features/classroom/components/UnifiedTimeline.tsx

@@ -24,6 +24,8 @@ export interface TranscriptEntry {
   timestamp: number;
   /** P1-2：用户修正后的文本（存在时优先显示） */
   editedText?: string;
+  /** P0-3：转写置信度（估算口径，<0.55 时 UI 弱化标记） */
+  confidence?: number;
 }
 
 interface UnifiedTimelineProps {
@@ -156,6 +158,9 @@ export function UnifiedTimeline({ bundle, liveTranscripts, autoAnchors = [], par
       const isEditing = editingId === t.id;
       const displayText = t.editedText ?? t.text;
       const isEdited = !!t.editedText;
+      // P0-3 低置信度标记：置信度 <0.55 的转写弱化显示并加角标（估算口径，
+      // 非统计置信度；未携带置信度的旧数据视为 1）
+      const isLowConfidence = typeof t.confidence === 'number' && t.confidence < 0.55;
       const time = new Date(t.timestamp).toLocaleTimeString([], { hour: '2-digit', minute: '2-digit', second: '2-digit' });
 
       if (isEditing) {
@@ -183,8 +188,11 @@ export function UnifiedTimeline({ bundle, liveTranscripts, autoAnchors = [], par
         <div key={t.id} className="flex gap-2 p-2 rounded-kb-sm transition-colors hover:bg-bg-tertiary/30 group">
           <span className="text-[10px] text-text-tertiary flex-shrink-0 mt-0.5 tabular-nums">{time}</span>
           <Mic className="w-3.5 h-3.5 flex-shrink-0 mt-0.5 text-emerald-500/70" strokeWidth={1.5} />
-          <span className={cn('flex-1 text-[12px] leading-relaxed min-w-0', idx === rows.length - 1 ? 'text-text-primary font-medium' : 'text-text-secondary')}>
+          <span className={cn('flex-1 text-[12px] leading-relaxed min-w-0', idx === rows.length - 1 ? 'text-text-primary font-medium' : 'text-text-secondary', isLowConfidence && 'opacity-60')}>
             {displayText}
+            {isLowConfidence && (
+              <span className="ml-1.5 text-[10px] text-text-quaternary font-medium" title={`置信度 ${(t.confidence ?? 0).toFixed(2)}，识别可能不准确`}>低置信</span>
+            )}
             {isEdited && (
               <span className="ml-1.5 text-[10px] text-amber-500 font-medium" title={`原始: ${t.text}`}>已修正</span>
             )}

client/src/features/classroom/hooks/useClassroomEvents.ts

@@ -34,7 +34,7 @@ import { remapKeyframeMarkers } from '../utils/tipTapImageUtils';
 import { persistKeyframeImage } from '../utils/keyframePersistence';
 import { transcribeWithRetry, toAsrLanguage, useAsrSemaphore, isLocalAsrReady, setOnAsrFallback } from '../utils/asrTranscriber';
 import { applySessionReplaces, getSessionHotwordsString, addDynamicBoosts } from '../utils/hotwordRuntime';
-import { cleanAsrResult, dedupeAcrossFinals } from '@/lib/capture/asrFilters';
+import { cleanAsrResult, dedupeAcrossFinals, estimateAsrConfidence } from '@/lib/capture/asrFilters';
 
 /** 触发一次增量分析所需的关键帧数 */
 const INCREMENTAL_BATCH_SIZE = 5;
@@ -50,6 +50,8 @@ export interface LiveTranscript {
   id: string;
   text: string;
   timestamp: number;
+  /** P0-3：转写置信度（估算口径，<0.55 时 UI 弱化标记；缺失视为 1） */
+  confidence?: number;
 }
 
 interface UseClassroomEventsOptions {
@@ -260,10 +262,14 @@ export function useClassroomEvents({
             sample_rate: 16000,
             channels: 1,
           }, 1, hotwords || undefined)
-            .then((text) => {
+            .then((outcome) => {
               asr.markSuccess();
               // 输出清洗：相邻重复压缩 + 幻觉过滤（本地路径主进程已 clean，此处兑底云端降级）
-              const cleaned = cleanAsrResult(text ?? '');
+              const cleaned = cleanAsrResult(outcome?.text ?? '');
+              // P0-3：清洗后文本变化时重估置信度（网关估算基于未清洗文本）
+              const confidence = cleaned && outcome
+                ? (cleaned === outcome.text ? outcome.confidence : estimateAsrConfidence(outcome.text, cleaned))
+                : 0;
               // 将转写结果回填到对应的音频段，并剥离 audioBase64 释放内存（单段约 1.2MB，
               // 长课堂数百段否则无界累积——内测 5GB 内存主因）。全量分析回退路径优先用
               // 已转写的 audioText（sessionAnalyzer: seg.audioText ?? transcribe），无需再持有原始音频；
@@ -281,7 +287,7 @@ export function useClassroomEvents({
                 // 实时转录上屏（FIFO 上限控制）；展示替换后文本（P1-3 替换词后处理），
                 // 原始清洗后转写经 audioTextRaw 可回溯
                 setLiveTranscripts((prev) => {
-                  const next = [...prev, { id: seg.id, text: applySessionReplaces(cleaned), timestamp: seg.timestampStart }];
+                  const next = [...prev, { id: seg.id, text: applySessionReplaces(cleaned), timestamp: seg.timestampStart, confidence }];
                   return next.length > MAX_LIVE_TRANSCRIPTS
                     ? next.slice(next.length - MAX_LIVE_TRANSCRIPTS)
                     : next;
@@ -329,7 +335,7 @@ export function useClassroomEvents({
     });
     const offFinal = window.electronAPI.on('asr_stream_final', (...args: unknown[]) => {
       if (statusRef.current !== 'capturing') return;
-      const data = args[0] as { text: string; timestamp: number };
+      const data = args[0] as { text: string; timestamp: number; confidence?: number };
       setPartialText('');
       // 双保险：主进程已 clean，此处兜底云端/旧版本主进程的未清洗输出
       const cleaned = cleanAsrResult(data?.text ?? '');
@@ -339,11 +345,15 @@ export function useClassroomEvents({
       const text = dedupeAcrossFinals(lastFinalTextRef.current, cleaned);
       lastFinalTextRef.current = text || lastFinalTextRef.current;
       if (!text) return;
+      // P0-3：置信度透传（主进程估算；本地再清洗后文本变化时重估）
+      const confidence = cleaned === data.text
+        ? (typeof data.confidence === 'number' ? data.confidence : 1)
+        : estimateAsrConfidence(data.text ?? '', cleaned);
       const id = crypto.randomUUID();
       const timestamp = data.timestamp || Date.now();
       // 实时转录上屏（FIFO 上限控制）；展示替换后文本（P1-3）
       setLiveTranscripts((prev) => {
-        const next = [...prev, { id, text: applySessionReplaces(text), timestamp }];
+        const next = [...prev, { id, text: applySessionReplaces(text), timestamp, confidence }];
         return next.length > MAX_LIVE_TRANSCRIPTS
           ? next.slice(next.length - MAX_LIVE_TRANSCRIPTS)
           : next;

client/src/features/classroom/utils/asrTranscriber.ts

@@ -30,6 +30,15 @@ interface TranscribeResponse {
   text: string;
   model_used?: string;
   warning?: string | null;
+  /** P0-3：网关估算置信度（质量门控弱化标记信号，非统计置信度） */
+  confidence?: number;
+}
+
+/** 统一转写结果：文本 + 置信度（P0-3 质量门控） */
+export interface TranscribeOutcome {
+  text: string;
+  /** 0-1 置信度（估算口径，见网关 transcribe_chain 与本地 estimateAsrConfidence） */
+  confidence: number;
 }
 
 // ================================================================
@@ -82,29 +91,33 @@ export function isLocalAsrReady(): boolean {
 // 本地 ASR 转写（IPC 调用主进程 sherpa-onnx）
 // ================================================================
 
-async function transcribeLocalViaIpc(payload: TranscribePayload): Promise<string | null> {
+async function transcribeLocalViaIpc(payload: TranscribePayload): Promise<TranscribeOutcome | null> {
   if (!window.electronAPI) throw new Error('electronAPI 不可用');
   const result = await window.electronAPI.invoke('local_asr_transcribe', {
     audioBase64: payload.audio_base64,
     language: payload.language,
     sampleRate: payload.sample_rate,
     channels: payload.channels,
     hotwords: payload.hotwords,
-  }) as { text: string; language: string; durationMs: number };
-  return result.text?.trim() || null;
+  }) as { text: string; language: string; confidence?: number; durationMs: number };
+  const text = result.text?.trim() ?? '';
+  if (!text) return null;
+  return { text, confidence: typeof result.confidence === 'number' ? result.confidence : 1 };
 }
 
 // ================================================================
 // 云端 ASR 转写（原有逻辑）
 // ================================================================
 
-async function transcribeCloud(payload: TranscribePayload): Promise<string | null> {
+async function transcribeCloud(payload: TranscribePayload): Promise<TranscribeOutcome | null> {
   const resp = await aiClient.post<TranscribeResponse>('/api/v1/asr/transcribe', payload, { timeout: 30000 });
   // fallback 降级响应（含 warning 或 fallback 空文本）按失败处理
   if (resp.warning || (!resp.text?.trim() && resp.model_used === 'fallback')) {
     throw new Error(resp.warning || 'ASR 服务降级，转写结果为空');
   }
-  return resp.text?.trim() || null;
+  const text = resp.text?.trim() ?? '';
+  if (!text) return null;
+  return { text, confidence: typeof resp.confidence === 'number' ? resp.confidence : 1 };
 }
 
 // ================================================================
@@ -121,12 +134,12 @@ async function transcribeCloud(payload: TranscribePayload): Promise<string | nul
  *
  * @param hotwords - 可选热词增强字符串（zipformer-transducer 支持，空格分隔）
  */
-export async function transcribeWithRetry(payload: TranscribePayload, retries = 1, hotwords?: string): Promise<string | null> {
+export async function transcribeWithRetry(payload: TranscribePayload, retries = 1, hotwords?: string): Promise<TranscribeOutcome | null> {
   // ── 本地 ASR 优先 ──
   if (_localAsrAvailable) {
     try {
-      const text = await transcribeLocalViaIpc({ ...payload, hotwords });
-      if (text) return text;
+      const outcome = await transcribeLocalViaIpc({ ...payload, hotwords });
+      if (outcome) return outcome;
     } catch (localErr) {
       console.warn('[asrTranscriber] 本地 ASR 失败，尝试云端降级:', localErr);
       if (!_localAsrFallbackToCloud) throw localErr;

client/src/lib/capture/asrFilters.ts

@@ -186,3 +186,24 @@ function similarityOf(a: string, b: string): number {
   const union = setA.size + setB.size - intersection;
   return union > 0 ? intersection / union : 0;
 }
+
+// ================================================================
+// 转写置信度估算（P0-3 质量门控）
+// ================================================================
+
+/**
+ * 估算转写置信度：sherpa zipformer 与云端 ASR API 均不返回统计置信度，
+ * 用「清洗前后文本长度比」作代理信号——幻觉过滤/重复压缩会显著缩短文本，
+ * 压缩损失越大说明原始输出越不可靠。
+ * 口径：cleaned 为空 → 0；ratio = cleaned/raw；confidence = 0.35 + 0.65*ratio
+ * （ratio=1 → 1.0；ratio=0.7 → 0.8；ratio=0.5 → 0.68）。
+ * 语义：供 UI 低置信度弱化标记，非统计置信度（与网关 transcribe 估算口径一致）。
+ */
+export function estimateAsrConfidence(rawText: string, cleanedText: string): number {
+  const raw = (rawText ?? '').trim();
+  const cleaned = (cleanedText ?? '').trim();
+  if (!raw || !cleaned) return 0;
+  if (raw === cleaned) return 1;
+  const ratio = cleaned.length / raw.length;
+  return Math.round((0.35 + 0.65 * Math.max(0, Math.min(1, ratio))) * 100) / 100;
+}

server/ai-gateway/chains/transcribe_chain.py

@@ -97,8 +97,21 @@ async def run(
         # 后处理
         result = self._postprocess_result(result)
         result.setdefault("language", language)
-        # GW-2#11: ASR API 不返回置信度，0.0 为占位常量而非测量值
-        result.setdefault("confidence", 0.0)
+        # GW-2#11 + P0-3：ASR API 不返回置信度，此处输出估算值供渲染进程
+        # 质量门控（低置信度段落 UI 弱化标记），非统计置信度。估算口径：
+        # 文本非空基线 0.7，按长度渐近 0.95；fallback/空文本为 0。
+        result["confidence"] = self._estimate_confidence(
+            result.get("text", ""), result.get("model", self.model),
+        )
         result.setdefault("model", self.model)
 
         return result
+
+    @staticmethod
+    def _estimate_confidence(text: str, model: str) -> float:
+        """估算转写置信度（P0-3，语义见 run 内注释）"""
+        if not text.strip():
+            return 0.0
+        if model == "fallback":
+            return 0.3
+        return round(min(0.95, 0.7 + min(len(text), 80) / 200), 2)

server/ai-gateway/chains/vision_extract_chain.py

@@ -281,11 +281,15 @@ async def run(
         # GW-2#4: 显式截断检测——GLM 等 provider 可能把 max_tokens clamp 到
         # 低于请求值（full 模式请求 4096 被 clamp 到 1024），输出在接近上限时
         # 大概率被截断。原实现静默返回残缺 JSON（结构化字段丢失）且照常缓存，
-        # 用户无感知；此处记录 truncated 告警供排查与降级重试决策
+        # 用户无感知；此处记录 truncated 告警供排查与降级重试决策。
+        # P0-3：truncated 标志随结果返回，路由层据此将置信度减半（低置信度
+        # 段落 UI 弱化标记的质量门控信号）
         content = result["content"]
         actual_max_tokens = result.get("max_tokens")
+        truncated = False
         if actual_max_tokens and len(content) > actual_max_tokens * 3:
             # 1 token ≈ 3 字符的保守估算（中文 1 token ≈ 1-2 字符，英文 ≈ 4）
+            truncated = True
             logger.warning(
                 "视觉提取疑似截断: mode=%s, model=%s, content=%d 字符, "
                 "max_tokens=%d（provider clamp 后），结构化字段可能缺失",
@@ -303,4 +307,5 @@ async def run(
             "model": result.get("model", self.model),
             "latency_ms": result.get("latency_ms", 0),
             "mode": effective_mode,
+            "truncated": truncated,
         }

server/ai-gateway/routers/vision.py

@@ -143,6 +143,10 @@ async def _run_chain(provider, model_name):
     aux_fields = [formulas, diagrams, key_points, code_blocks, concepts]
     aux_filled = sum(1 for v in aux_fields if v) / len(aux_fields)
     confidence = round(filled_text * 0.5 + aux_filled * 0.5, 2)
+    # P0-3：疑似截断（max_tokens clamp 导致结构化字段缺失）时置信度减半，
+    # 作为渲染进程低置信度弱化标记的质量门控信号
+    if result.get("truncated"):
+        confidence = round(confidence * 0.5, 2)
 
     # 实际使用的模式
     effective_mode = result.get("mode", body.mode or "auto")