feat(memory,conversation): 记忆富化/证据包、时间线幂等字段与对话分段全链路

数据库 - 新增迁移 0003：timeline_events.memory_source_id 外键 → memory_sources，便于按 ingest 源做时间线幂等后端 - 记忆 - 新增 ingest 后 LLM 富化（摘要/事实/时间线），可配置开关与最大字符数 - 新增证据包组装：合并 chunk、摘要、事实、时间线、故事等检索结果；支持空 query 时是否仍带 rolling 等开关 - repo/retriever/service/router/schemas/summarizer/timeline/extractor 等扩展；文档 memory-retrieval.md 更新后端 - 对话 WS - 增加 PING/PONG；分段 ASR 日志与空音频处理；转写失败与「无助手回复」错误提示更明确 - 助手多段回复持久化使用统一分隔符，与分段逻辑一致后端 - Agent - reply_limits：按 [SPLIT] 与段落拆段，并保证非空 fallback，供 WS 与 TTS 多段下发后端 - 回忆录任务 - transcript ingest 记录 source_id；任务成功结?
2026-03-27 16:01:28 +08:00
parent 1374f6e8f5
commit e4bf0710c7
70 changed files with 3404 additions and 557 deletions
--- a/api/app/adapters/asr/tencent_asr.py
+++ b/api/app/adapters/asr/tencent_asr.py
@@ -38,7 +38,7 @@ class TencentASRProvider:
    async def transcribe(self, audio: bytes, format: str = "m4a") -> str:
        client = self._get_client()
        if not client:
-            return ""
+            return "转写失败: 腾讯云 ASR 客户端未初始化（请检查密钥与依赖）"
        try:
            from tencentcloud.asr.v20190614 import models

@@ -46,12 +46,26 @@ class TencentASRProvider:
            req = models.SentenceRecognitionRequest()
            req.EngSerViceType = "16k_zh"
            req.SourceType = 1
-            req.VoiceFormat = format
+            # 小写；与文档一致。iOS 常见为 m4a(AAC) 容器，与 16k 引擎匹配
+            req.VoiceFormat = (format or "m4a").lower()
            req.Data = audio_base64
            req.DataLen = len(audio)

            resp = client.SentenceRecognition(req)
-            return (resp.Result or "").strip()
+            text = (resp.Result or "").strip()
+            if text:
+                return text
+            err = getattr(resp, "Error", None) or getattr(resp, "Message", None)
+            logger.warning(
+                "Tencent ASR empty Result, audio_len={} format={} err={}",
+                len(audio),
+                req.VoiceFormat,
+                err,
+            )
+            return (
+                "转写失败: 腾讯云返回空文本（常见原因：采样率与 16k_zh 不匹配、"
+                "格式不受支持或音频无效；请确认客户端为 16kHz 单声道 m4a）"
+            )
        except Exception as e:
-            logger.error("Tencent ASR transcribe failed: {}", e)
-            return ""
+            logger.error("Tencent ASR transcribe failed: {}", e, exc_info=True)
+            return f"转写失败: {e}"[:500]
--- a/api/app/adapters/asr/whisper_local.py
+++ b/api/app/adapters/asr/whisper_local.py
@@ -1,11 +1,42 @@
 """Local faster-whisper ASR adapter — implements ASRProvider port."""

-from app.core.logging import get_logger
+from __future__ import annotations
+
+import asyncio
 import os
+import re
 import tempfile
+from typing import Any, Iterable
+
+from app.core.logging import get_logger

 logger = get_logger(__name__)

+_SUBTITLE_WATERMARK_RE = re.compile(
+    r"(字幕|听译|压制|字幕组).{0,20}(by|BY|By)|字幕\s*by",
+    re.UNICODE,
+)
+
+
+def _looks_like_subtitle_hallucination(text: str) -> bool:
+    """静音时第二遍易吐出视频字幕水印；仅丢弃此类短句。"""
+    t = (text or "").strip()
+    if len(t) > 48:
+        return False
+    if _SUBTITLE_WATERMARK_RE.search(t):
+        return True
+    if len(t) <= 12 and "字幕" in t and not re.search(r"[？?！!。，、]", t):
+        return True
+    return False
+
+
+def _join_segment_text(segments: Iterable[Any]) -> tuple[str, int]:
+    segs = list(segments)
+    return "".join(str(getattr(seg, "text", "") or "") for seg in segs).strip(), len(
+        segs
+    )
+
+
 _DEFAULT_CACHE_DIR = os.path.normpath(
    os.path.join(
        os.path.dirname(os.path.abspath(__file__)),
@@ -70,30 +101,95 @@ class WhisperASRProvider:
        return self._load_model()

    async def transcribe(self, audio: bytes, format: str = "m4a") -> str:
+        # 与 v1.1.0 相同的单次 transcribe；推理放线程池，避免阻塞 asyncio（tag 上为同步调用）。
        self._load_model()
        if not self._model:
            return ""

-        tmp_path = None
-        try:
-            with tempfile.NamedTemporaryFile(suffix=f".{format}", delete=False) as tmp:
-                tmp.write(audio)
-                tmp_path = tmp.name
+        model = self._model

-            segments, _info = self._model.transcribe(
-                tmp_path,
-                language="zh",
-                beam_size=5,
-                vad_filter=True,
-                vad_parameters={"min_silence_duration_ms": 500},
-            )
-            return "".join(seg.text for seg in segments).strip()
-        except Exception as e:
-            logger.error("Whisper transcribe failed: {}", e)
-            return ""
-        finally:
-            if tmp_path and os.path.exists(tmp_path):
-                try:
-                    os.remove(tmp_path)
-                except OSError:
-                    pass
+        def _sync_transcribe() -> str:
+            tmp_path = None
+            try:
+                with tempfile.NamedTemporaryFile(
+                    suffix=f".{format}", delete=False
+                ) as tmp:
+                    tmp.write(audio)
+                    tmp_path = tmp.name
+
+                segments, _info = model.transcribe(
+                    tmp_path,
+                    language="zh",
+                    beam_size=5,
+                    vad_filter=True,
+                    vad_parameters={
+                        "min_silence_duration_ms": 500,
+                        "threshold": 0.35,
+                        "min_speech_duration_ms": 200,
+                    },
+                )
+                text, pass1_seg_count = _join_segment_text(segments)
+                used_second_pass = False
+                pass2_seg_count = 0
+                pass3_seg_count = 0
+
+                if not text:
+                    logger.info(
+                        "Whisper VAD pass 无文本，关闭 VAD 再试一次（短录音易被 VAD 判为静音）"
+                    )
+                    segments2, _info2 = model.transcribe(
+                        tmp_path,
+                        language="zh",
+                        beam_size=5,
+                        vad_filter=False,
+                        condition_on_previous_text=False,
+                        # 略抬高：减少边界片段被标成 no_speech 而整段为空
+                        no_speech_threshold=0.85,
+                    )
+                    raw2, pass2_seg_count = _join_segment_text(segments2)
+                    used_second_pass = True
+                    if raw2 and _looks_like_subtitle_hallucination(raw2):
+                        logger.info(
+                            "Whisper 丢弃疑似字幕水印幻听: {!r}",
+                            raw2[:120],
+                        )
+                        text = ""
+                    else:
+                        text = raw2
+
+                if not text and used_second_pass:
+                    try:
+                        from faster_whisper import decode_audio
+
+                        audio_np = decode_audio(tmp_path, sampling_rate=16000)
+                        segments3, _info3 = model.transcribe(
+                            audio_np,
+                            language="zh",
+                            beam_size=5,
+                            vad_filter=False,
+                            condition_on_previous_text=False,
+                            no_speech_threshold=0.85,
+                        )
+                        raw3, pass3_seg_count = _join_segment_text(segments3)
+                        if raw3 and _looks_like_subtitle_hallucination(raw3):
+                            logger.info(
+                                "Whisper decode_audio 回退仍是疑似字幕水印幻听: {!r}",
+                                raw3[:120],
+                            )
+                        elif raw3:
+                            text = raw3
+                    except Exception as ex:
+                        logger.warning("Whisper decode_audio 回退失败: {}", ex)
+
+                return text
+            except Exception as e:
+                logger.error("Whisper transcribe failed: {}", e)
+                return ""
+            finally:
+                if tmp_path and os.path.exists(tmp_path):
+                    try:
+                        os.remove(tmp_path)
+                    except OSError:
+                        pass
+
+        return await asyncio.to_thread(_sync_transcribe)