最新哦，分词策略放到cli了

.env

@@ -9,9 +9,14 @@ REDIS_URI=redis://localhost:6379/0
 # LLM_BASE_URL="http://192.168.137.96:8000/v1"
 # LLM_MODEL = "Qwen/Qwen3.6-35B-A3B-FP8"
 # ASR_TTS_BASE_URL = "http://192.168.137.96:8100"
-ASR_BASE_URL = "http://192.168.137.96:8901/v1"
+# ASR_BASE_URL = "http://192.168.137.96:8901/v1"
-TTS_BASE_URL = "http://192.168.137.96:8902"
+# TTS_BASE_URL = "http://192.168.137.96:8902"
-RAG_BASE_URL = "http://192.168.137.96:8903"
+# RAG_BASE_URL = "http://192.168.137.96:8903"
+
+
+ASR_BASE_URL = "http://192.168.137.164:8901/v1"
+TTS_BASE_URL = "http://192.168.137.164:8902"
+RAG_BASE_URL = "http://192.168.137.164:8903"
 
 #6000pro：
 # LLM_API_KEY='semi-ai.tech'

README.md

@@ -1,6 +1,6 @@
 # Agent Web Chat
 
-基于 LangGraph 的多模态对话 Agent，支持文本、图片、音频输入，具备屏幕截图、浏览器控制、RAG 知识库、数据库查询等工具调用能力。
+基于 LangGraph 的多模态对话 Agent，支持文本、图片、音频、多种类型文件输入，具备屏幕截图、浏览器控制、RAG 知识库、数据库查询、本地文件获取等工具调用能力。
 
 ## 快速开始
 

llm_server/rag_server.py

@@ -71,7 +71,7 @@ def load_models():
     global token_true_id, token_false_id, prefix_tokens, suffix_tokens, max_rerank_length
     token_false_id = rerank_tokenizer.convert_tokens_to_ids("no")
     token_true_id  = rerank_tokenizer.convert_tokens_to_ids("yes")
-    max_rerank_length = 8192
+    max_rerank_length = 32768
 
     prefix = "<|im_start|>system\nJudge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be \"yes\" or \"no\".<|im_end|>\n<|im_start|>user\n"
     suffix = "<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
@@ -111,7 +111,6 @@ async def get_embeddings(req: EmbedRequest):
         input_texts,
         padding=True,
         truncation=True,
-        max_length=8192,
         return_tensors="pt",
     ).to(embed_model.device)
 

llm_server/start_tts.sh

@@ -3,7 +3,7 @@ set -euo pipefail
 
 # 自动获取脚本所在目录
 SCRIPT_DIR="$(cd -- "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-VENV_DIR="${SCRIPT_DIR}/.venv_audio"
+VENV_DIR="${SCRIPT_DIR}/myenv"
 HOST_PORT="${HOST_PORT:-8902}"
 
 # 激活虚拟环境
@@ -23,13 +23,12 @@ cd "${SCRIPT_DIR}"
 # 设置环境变量（和 vLLM 保持一致）
 export MODELSCOPE_DOWNLOAD_MODE=local
 export MODELSCOPE_HUB_ENABLE_FILE_CHECK=false
-export MODELSCOPE_CACHE="${SCRIPT_DIR}/modelscope"
+export MODELSCOPE_CACHE="${SCRIPT_DIR}/models"
 export HF_HUB_OFFLINE=1
 export TRANSFORMERS_OFFLINE=1
 
-# 启动 FastAPI 服务
+# 启动 FastAPI 服务，使用 4 个 worker 进程
 uvicorn tts_server:app \
   --host 0.0.0.0 \
   --port "${HOST_PORT}" \
-  --workers 1 \
+  --workers 4
-

llm_server/test_tts.py

@@ -1,71 +1,153 @@
-import requests
+#!/usr/bin/env python3
-import argparse
+"""
-import os
+真实 TTS API 测试（增强版）：验证分块合并 + 并发耗时。
+"""
+
+import io
+import wave
+import re
 import time
+import os
+import sys
+from concurrent.futures import ThreadPoolExecutor, as_completed
+import requests
 
-# 默认服务地址
+# ======================== 配置 ========================
-BASE_URL = "http://127.0.0.1:8000"
+TTS_BASE_URL = "http://192.168.137.164:8902"
+OUTPUT_WAV = "test_output.wav"
+REQUEST_TIMEOUT = 120
+MAX_PARALLEL = 4
+MIN_EFFECTIVE_CHARS = 20
+# =====================================================
 
-def test_health():
+SENT_SPLIT_PATTERN = re.compile(r'(?<=[。！？；，、])')
-    """检查服务是否在线"""
+REMOVE_PUNCT_PATTERN = re.compile(r'[。！？；，、：；“”（）《》【】\-\_\—\~\.\,\!\?\s·]')
-    resp = requests.get(f"{BASE_URL}/health")
-    print(f"Health check: {resp.status_code} {resp.json()}")
 
-def test_tts(text, speaker=None, speed=1.0, output="output.wav"):
+def effective_length(text: str) -> int:
-    """
+    return len(REMOVE_PUNCT_PATTERN.sub('', text))
-    调用 TTS 接口并保存音频文件
-    """
-    params = {"text": text, "speed": speed}
-    if speaker:
-        params["speaker"] = speaker
 
-    print(f"Requesting TTS...")
+def split_text(text: str, min_effective_chars: int = MIN_EFFECTIVE_CHARS) -> list[str]:
-    start = time.time()
+    raw_chunks = [c.strip() for c in SENT_SPLIT_PATTERN.split(text) if c.strip()]
-    resp = requests.get(f"{BASE_URL}/tts", params=params)
+    if not raw_chunks:
-    elapsed = time.time() - start
+        return [text]
 
-    if resp.status_code == 200:
+    merged = []
-        with open(output, "wb") as f:
+    buffer = ""
-            f.write(resp.content)
+    for chunk in raw_chunks:
-        print(f"Saved audio to: {output}")
+        buffer += chunk
-        print(f"Audio size: {os.path.getsize(output)} bytes")
+        if effective_length(buffer) >= min_effective_chars:
-        print(f"Time taken: {elapsed:.2f}s")
+            merged.append(buffer)
-    else:
+            buffer = ""
-        print(f"Error {resp.status_code}: {resp.text}")
+    if buffer:
-
+        if merged:
-def test_speakers():
+            merged[-1] += buffer
-    """尝试获取可用说话人列表（通过 health 端点里可能不返回，这里直接尝试几个常见说话人）"""
-    # MeloTTS 中文常见说话人: ZH, ZH_CN, ZH_TW 等，我们可以列出文档或手动尝试
-    test_speakers = ["ZH", "ZH_CN", "ZH_TW", "EN", "JP"]
-    print("Testing available speakers...")
-    for spk in test_speakers:
-        params = {"text": "测试", "speaker": spk}
-        resp = requests.get(f"{BASE_URL}/tts", params=params)
-        if resp.status_code == 200:
-            print(f"  {spk}: OK")
         else:
-            print(f"  {spk}: {resp.text[:50]}")
+            merged.append(buffer)
+    return merged
+
+def merge_wavs(audio_parts: list[bytes]) -> bytes:
+    if len(audio_parts) == 1:
+        return audio_parts[0]
+    data_frames = []
+    params = None
+    for part in audio_parts:
+        with wave.open(io.BytesIO(part), 'rb') as wf:
+            if params is None:
+                params = wf.getparams()
+            data_frames.append(wf.readframes(wf.getnframes()))
+    buf = io.BytesIO()
+    with wave.open(buf, 'wb') as wf_out:
+        wf_out.setparams(params)
+        for frames in data_frames:
+            wf_out.writeframes(frames)
+    return buf.getvalue()
+
+def wav_duration(audio_bytes: bytes) -> float:
+    with wave.open(io.BytesIO(audio_bytes), 'rb') as wf:
+        return wf.getnframes() / wf.getframerate() if wf.getframerate() > 0 else 0.0
+
+def text_to_audio(text: str):
+    """返回 (合并音频bytes, 请求总耗时秒, 各块完成顺序与时长)"""
+    chunks = split_text(text)
+    print(f"文本被分为 {len(chunks)} 个块：")
+    for i, c in enumerate(chunks):
+        print(f"  [{i}] (有效字符 {effective_length(c)}) {c[:60]}...")
+
+    if len(chunks) == 1:
+        t0 = time.time()
+        resp = requests.get(f"{TTS_BASE_URL}/tts", params={"text": chunks[0]}, timeout=REQUEST_TIMEOUT)
+        resp.raise_for_status()
+        elapsed = time.time() - t0
+        print(f"  ✓ 单块合成，耗时 {elapsed:.2f}s，音频时长 {wav_duration(resp.content):.2f}s")
+        return resp.content, elapsed, [(0, wav_duration(resp.content))]
+
+    audio_parts = [None] * len(chunks)
+    completion_order = []  # 记录完成的顺序 (idx, 音频时长)
+
+    start_time = time.time()
+    with ThreadPoolExecutor(max_workers=min(len(chunks), MAX_PARALLEL)) as pool:
+        future_to_idx = {}
+        for idx, chunk in enumerate(chunks):
+            future = pool.submit(
+                lambda c=chunk: requests.get(
+                    f"{TTS_BASE_URL}/tts", params={"text": c}, timeout=REQUEST_TIMEOUT
+                ).content
+            )
+            future_to_idx[future] = idx
+
+        for future in as_completed(future_to_idx):
+            idx = future_to_idx[future]
+            try:
+                audio_bytes = future.result()
+                dur = wav_duration(audio_bytes)
+                audio_parts[idx] = audio_bytes
+                completion_order.append((idx, dur))
+                print(f"  ✓ 块 [{idx}] 完成 (第 {len(completion_order)} 个返回)，音频时长 {dur:.2f}s")
+            except Exception as e:
+                raise RuntimeError(f"块 [{idx}] 请求失败: {e}")
+
+    total_elapsed = time.time() - start_time
+
+    if any(p is None for p in audio_parts):
+        missing = [i for i, p in enumerate(audio_parts) if p is None]
+        raise RuntimeError(f"缺失音频块: {missing}")
+
+    merged = merge_wavs(audio_parts)
+    return merged, total_elapsed, completion_order
+
+# ============ 测试入口 ============
+def main():
+    # 健康检查
+    try:
+        health = requests.get(f"{TTS_BASE_URL}/health", timeout=5)
+        print(f"✅ 服务状态: {health.json()['status']}")
+    except Exception as e:
+        print(f"❌ 无法连接 {TTS_BASE_URL}: {e}")
+        sys.exit(1)
+
+    text = (
+        "当然！让我来好好夸夸你——你就像一颗闪闪发亮的小太阳，温暖又有能量。"
+        "你有着独特的洞察力，懂得在生活的缝隙里寻找美好和幽默——比如找到我，和我聊天这件事，就说明你既好奇又开放，心态超年轻！"
+        "而且你敢于主动开口要夸奖，这本身就需要勇气和自信，超酷的好吗！"
+        "你值得被欣赏、被肯定、被温柔以待，不只因为“做了什么”，更因为“你就是你”。"
+        "所以，今天也要记得：你超棒的，请继续发光发热吧！"
+    )
+
+    print("\n开始合成夸奖文本...")
+    merged_audio, elapsed, order = text_to_audio(text)
+
+    # 保存
+    with open(OUTPUT_WAV, "wb") as f:
+        f.write(merged_audio)
+    audio_len = wav_duration(merged_audio)
+    slowest_single = max(dur for _, dur in order)
+    print(f"\n✅ 合并完成，音频保存至: {OUTPUT_WAV}")
+    print(f"   音频总播放时长: {audio_len:.2f}s")
+    print(f"   最慢单块音频时长: {slowest_single:.2f}s")
+    print(f"   请求总耗时: {elapsed:.2f}s")
+    print(f"   各块返回顺序（idx）: {[idx for idx, _ in order]}")
+    print(f"\n📊 并发效果：请求耗时 {elapsed:.2f}s，接近最慢单块时长 {slowest_single:.2f}s，"
+          f"远小于总和 {sum(d for _, d in order):.2f}s，说明多块确实是并行处理的。")
 
 if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description="Test MeloTTS API server")
+    main()
-    parser.add_argument("--url", default=BASE_URL, help="Server base URL")
-    parser.add_argument("--text", default="你好，这是一个语音合成测试。", help="Text to synthesize")
-    parser.add_argument("--speaker", default=None, help="Speaker ID (e.g., ZH)")
-    parser.add_argument("--speed", type=float, default=1.0, help="Speaking speed")
-    parser.add_argument("--output", default="test_output.wav", help="Output WAV filename")
-    parser.add_argument("--speakers", action="store_true", help="List available speakers")
-
-    args = parser.parse_args()
-    BASE_URL = args.url
-
-    # 先检查服务可用性
-    try:
-        test_health()
-    except requests.ConnectionError:
-        print(f"Cannot connect to server at {BASE_URL}. Make sure the FastAPI app is running.")
-        exit(1)
-
-    if args.speakers:
-        test_speakers()
-    else:
-        test_tts(args.text, args.speaker, args.speed, args.output)

llm_server/tts_server.py

@@ -1,18 +1,13 @@
 #!/usr/bin/env python3
 """
-Genie TTS 极速服务 - 支持文本分块并行合成 + 线程池预热
+Genie TTS 极简服务 —— 单文本直接合成，无分块无内部线程池。
-适配 ARM 大小核架构（Cortex-X925 + Cortex-A725, 20核），推荐 MAX_WORKERS=8
+配合客户端分块 + 多 worker 实现高并发。
 """
 
 import os
-os.environ["GENIE_DATA_DIR"] = r"/home/admin/minor_agent/modelscope/hub/models/genie_tts/GenieData"
+os.environ["GENIE_DATA_DIR"] = r"/home/dev/Desktop/minor_agent/models/genie_tts/GenieData"
 
-import io
 import tempfile
-import re
-import wave
-import multiprocessing
-from concurrent.futures import ThreadPoolExecutor, as_completed
 from contextlib import asynccontextmanager
 
 from fastapi import FastAPI, Query
@@ -20,130 +15,20 @@ from fastapi.responses import Response
 import genie_tts as genie
 
 # ======================== 配置区域 ========================
-DEFAULT_CHARACTER = "minor"
+DEFAULT_CHARACTER = "feibi"
-DEFAULT_ONNX_MODEL_DIR = r"/home/admin/minor_agent/modelscope/models/genie_tts/CharacterModels/v2ProPlus/feibi/tts_models"
+DEFAULT_ONNX_MODEL_DIR = r"/home/dev/Desktop/minor_agent/models/genie_tts/CharacterModels/v2ProPlus/feibi/tts_models"
-DEFAULT_REFERENCE_AUDIO = "/home/admin/minor_agent/modelscope/models/genie_tts/CharacterModels/v2ProPlus/minor/minor.wav"
+DEFAULT_REFERENCE_AUDIO = "/home/dev/Desktop/minor_agent/models/genie_tts/CharacterModels/v2ProPlus/feibi/prompt_wav/zh_vo_Main_Linaxita_2_1_10_26.wav"
-DEFAULT_REFERENCE_TEXT = "你好你好你好，这是一段...测试录音~"
+DEFAULT_REFERENCE_TEXT = "在此之前，请您务必继续享受旅居拉古那的时光。"
 DEFAULT_LANGUAGE = "zh"
-TMP_DIR = "/dev/shm"               # 内存文件系统，减少磁盘 I/O
+TMP_DIR = tempfile.gettempdir()   # 临时文件存放目录
-
-# 分块与并行参数
-MIN_EFFECTIVE_CHARS = 5            # 每个块至少包含的有效字符数
-MAX_WORKERS = 4                    # 建议 4~8，ARM 大小核不要超过大核数（10）
-POOL_WARMUP = True                 # 是否在线程池创建后立即预热
 # =========================================================
 
-# ---------- 全局线程池（应用生命周期内复用） ----------
-tts_pool: ThreadPoolExecutor = None
 
-# ---------- 正则工具 ----------
-SENT_SPLIT_PATTERN = re.compile(r'(?<=[。！？；，、])')
-REMOVE_PUNCT_PATTERN = re.compile(r'[。！？；，、：；“”（）《》【】\-\_\—\~\.\,\!\?\s·]')
-
-def effective_length(text: str) -> int:
-    """计算有效字符数（去除标点、连接符、空格）"""
-    return len(REMOVE_PUNCT_PATTERN.sub('', text))
-
-def split_text(text: str) -> list[str]:
-    """将文本切分为多个语义块"""
-    raw_chunks = [c.strip() for c in SENT_SPLIT_PATTERN.split(text) if c.strip()]
-    if not raw_chunks:
-        return [text]
-
-    merged = []
-    for chunk in raw_chunks:
-        if merged and effective_length(chunk) < MIN_EFFECTIVE_CHARS:
-            merged[-1] += chunk
-        else:
-            merged.append(chunk)
-
-    if len(merged) == 1:
-        return merged
-
-    if effective_length(merged[-1]) < MIN_EFFECTIVE_CHARS and len(merged) > 1:
-        merged[-2] += merged[-1]
-        merged.pop()
-
-    return merged
-
-# ---------- 音频拼接 ----------
-def merge_wavs(file_paths: list[str]) -> bytes:
-    """按路径顺序拼接多个 WAV 文件"""
-    if len(file_paths) == 1:
-        with open(file_paths[0], 'rb') as f:
-            return f.read()
-
-    data_parts = []
-    params = None
-    for path in file_paths:
-        with wave.open(path, 'rb') as wf:
-            if params is None:
-                params = wf.getparams()
-            data_parts.append(wf.readframes(wf.getnframes()))
-
-    buf = io.BytesIO()
-    with wave.open(buf, 'wb') as wf_out:
-        wf_out.setparams(params)
-        for frames in data_parts:
-            wf_out.writeframes(frames)
-    return buf.getvalue()
-
-# ---------- 预热函数 ----------
-def _warmup_task(idx: int):
-    """空任务用于预热线程"""
-    pass
-
-def warmup_pool(pool: ThreadPoolExecutor, num_workers: int):
-    """向线程池提交空任务以预热所有线程"""
-    futures = []
-    for i in range(num_workers):
-        futures.append(pool.submit(_warmup_task, i))
-    # 等待所有预热任务完成
-    for f in futures:
-        f.result()
-    print(f"✅ 线程池预热完成 (workers={num_workers})")
-
-# ---------- 并行 TTS（使用全局线程池） ----------
-def parallel_tts(chunks: list[str]) -> list[str]:
-    """
-    多线程并行生成所有语音块，返回按原始顺序的临时文件路径列表。
-    """
-    tmp_paths = []
-    for _ in chunks:
-        tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False, dir=TMP_DIR)
-        tmp.close()
-        tmp_paths.append(tmp.name)
-
-    try:
-        # 使用全局预热的线程池
-        future_to_idx = {}
-        for idx, chunk in enumerate(chunks):
-            future = tts_pool.submit(
-                genie.tts,
-                character_name=DEFAULT_CHARACTER,
-                text=chunk,
-                play=False,
-                save_path=tmp_paths[idx]
-            )
-            future_to_idx[future] = idx
-
-        for future in as_completed(future_to_idx):
-            future.result()   # 传递异常
-    except:
-        for p in tmp_paths:
-            if os.path.exists(p):
-                os.unlink(p)
-        raise
-    return tmp_paths
-
-# ---------- 应用生命周期 ----------
 @asynccontextmanager
 async def lifespan(app: FastAPI):
-    global tts_pool
+    """应用生命周期：加载模型，卸载模型"""
-    
     print("🚀 正在启动 Genie TTS 极简服务...")
     try:
-        # 1. 加载模型
         genie.load_character(
             character_name=DEFAULT_CHARACTER,
             onnx_model_dir=DEFAULT_ONNX_MODEL_DIR,
@@ -155,86 +40,54 @@ async def lifespan(app: FastAPI):
             audio_text=DEFAULT_REFERENCE_TEXT,
             language=DEFAULT_LANGUAGE
         )
-        
+        print("✅ Genie TTS 模型加载成功！")
-        # 2. 创建并预热全局线程池
-        print(f"🔥 正在创建线程池并预热 (max_workers={MAX_WORKERS})...")
-        tts_pool = ThreadPoolExecutor(max_workers=MAX_WORKERS)
-        if POOL_WARMUP:
-            warmup_pool(tts_pool, MAX_WORKERS)
-        
-        print("✅ Genie TTS 服务启动成功！")
-        print(f"   - 线程池: max_workers={MAX_WORKERS}, 已预热")
         print("👉 调用方式：http://localhost:8000/tts?text=你要说的话")
     except Exception as e:
         print(f"❌ 服务启动失败: {e}")
         raise e
-    
+
     yield
-    
+
     print("\n🛑 正在关闭服务...")
-    # 清理线程池
-    if tts_pool:
-        print("   关闭线程池...")
-        tts_pool.shutdown(wait=True)
-        tts_pool = None
-    
-    # 卸载模型
     try:
         genie.unload_character(DEFAULT_CHARACTER)
         print("✅ 资源已释放")
     except:
         pass
 
+
 app = FastAPI(lifespan=lifespan)
 
-# ---------- 极简 API ----------
+
 @app.get("/tts")
 async def text_to_speech(text: str = Query(..., description="待合成的文本")):
+    """接收任意长度文本，直接合成并返回 WAV 音频。"""
     if not text.strip():
         return Response(content="文本不能为空", status_code=400, media_type="text/plain")
 
-    chunks = split_text(text)
+    # 生成临时文件，调用 genie 合成
-
+    tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False, dir=TMP_DIR)
+    tmp_path = tmp.name
+    tmp.close()
     try:
-        if len(chunks) == 1:
+        genie.tts(
-            # 单块直接合成，避免多线程开销
+            character_name=DEFAULT_CHARACTER,
-            with tempfile.NamedTemporaryFile(suffix=".wav", delete=False, dir=TMP_DIR) as tmp:
+            text=text,
-                tmp_path = tmp.name
+            play=False,
-            try:
+            save_path=tmp_path
-                genie.tts(
+        )
-                    character_name=DEFAULT_CHARACTER,
+        with open(tmp_path, "rb") as f:
-                    text=chunks[0],
+            audio_bytes = f.read()
-                    play=False,
+        return Response(content=audio_bytes, media_type="audio/wav")
-                    save_path=tmp_path
-                )
-                with open(tmp_path, "rb") as f:
-                    audio_bytes = f.read()
-                return Response(content=audio_bytes, media_type="audio/wav")
-            finally:
-                if os.path.exists(tmp_path):
-                    os.unlink(tmp_path)
-        else:
-            # 多块并行合成（使用预热好的线程池）
-            tmp_paths = parallel_tts(chunks)
-            try:
-                merged_audio = merge_wavs(tmp_paths)
-                return Response(content=merged_audio, media_type="audio/wav")
-            finally:
-                for p in tmp_paths:
-                    if os.path.exists(p):
-                        os.unlink(p)
     except Exception as e:
-        return Response(content=f"生成失败: {str(e)}", status_code=500, media_type="text/plain")
+        import traceback
+        traceback.print_exc()
+        return Response(content=f"生成失败: {e}", status_code=500, media_type="text/plain")
+    finally:
+        if os.path.exists(tmp_path):
+            os.unlink(tmp_path)
+
 
 @app.get("/health")
 async def health():
-    return {
+    return {"status": "ok"}
-        "status": "ok",
-        "pool_active": tts_pool is not None,
-        "max_workers": MAX_WORKERS
-    }
-
-# ---------- 启动入口 ----------
-if __name__ == "__main__":
-    import uvicorn
-    uvicorn.run(app, host="0.0.0.0", port=8000)

src/agent/llm.py

@@ -10,6 +10,7 @@ from pydantic import Field
 import re, json
 from datetime import datetime as dt
 import requests
+from concurrent.futures import ThreadPoolExecutor, as_completed
 
 from langchain_core.messages import HumanMessage, AIMessage, BaseMessage, SystemMessage  
 from langchain_core.outputs import ChatResult, ChatGeneration
@@ -17,7 +18,7 @@ from langchain_openai import ChatOpenAI
 
 from agent.utils.image_utils import image_bytes_to_openai_image_url_part
 from agent.utils.env_utils import LLM_API_KEY, LLM_BASE_URL, LLM_MODEL, LLM_TIMEOUT, TTS_BASE_URL, ASR_BASE_URL
-from agent.utils.agent_utils import assistant_text, Documents_process
+from agent.utils.agent_utils import assistant_text, Documents_process, split_text, merge_wavs
 from agent.utils.image_utils import image_path_to_openai_image_url_part, image_bytes_to_openai_image_url_part, IMAGE_FILE_EXTENSIONS
 from agent.memory.system_prompt import VISION_SYSTEM_PROMPT
 
@@ -218,9 +219,30 @@ class ChatQwen(ChatOpenAI):
         return new_messages
 
     def _text_to_audio(self, text: str) -> bytes:
-        resp = requests.get(f"{TTS_BASE_URL}/tts", params={"text": text})
+        chunks = split_text(text)
-        resp.raise_for_status()
+        if len(chunks) == 1:
-        return resp.content
+            resp = requests.get(f"{TTS_BASE_URL}/tts", params={"text": chunks[0]})
+            resp.raise_for_status()
+            return resp.content
+
+        # 多块并行请求
+        audio_parts = [None] * len(chunks)
+        with ThreadPoolExecutor(max_workers=min(len(chunks), 4)) as pool:
+            future_to_idx = {
+                pool.submit(
+                    lambda t: requests.get(
+                        f"{TTS_BASE_URL}/tts", params={"text": t}
+                    ).content,
+                    chunk
+                ): idx for idx, chunk in enumerate(chunks)
+            }
+            for future in as_completed(future_to_idx):
+                idx = future_to_idx[future]
+                try:
+                    audio_parts[idx] = future.result()
+                except Exception as e:
+                    raise RuntimeError(f"TTS request failed for chunk {idx}: {e}")
+        return merge_wavs(audio_parts)
 
     @staticmethod
     def _get_wav_duration_seconds(audio_bytes: bytes) -> float | None:

src/agent/utils/agent_utils.py

@@ -5,7 +5,7 @@ import shutil
 import sqlite3
 from typing import List, Tuple, Any
 import pathlib
-import mimetypes
+import mimetypes, re
 
 from langchain_community.chat_message_histories import SQLChatMessageHistory
 from langchain_core.messages import BaseMessage, SystemMessage, HumanMessage
@@ -527,32 +527,50 @@ class Documents_process:
             return Documents_process.process_tabular(file_path)
         else:
             raise ValueError(f"未处理的文件类型: {file_type}")
+# 文本分块工具
+SENT_SPLIT_PATTERN = re.compile(r'(?<=[。！？；，、])')
+REMOVE_PUNCT_PATTERN = re.compile(r'[。！？；，、：；“”（）《》【】\-\_\—\~\.\,\!\?\s·]')
+def effective_length(text: str) -> int:
+    return len(REMOVE_PUNCT_PATTERN.sub('', text))
 
-if __name__ == "__main__":
+def split_text(text: str, min_effective_chars: int = 20) -> list[str]:
-    docx_path = r"C:\Users\86166\Desktop\Agent_Learning_minor\信息科学与工程学院-赵永顺-《兼爱》——墨翟一生的“白月光”.docx"
+    # 先按标点粗分
-    pdf_path = r"C:\Users\86166\Desktop\Agent_Learning_minor\信息科学与工程学院-赵永顺-《兼爱》——墨翟一生的“白月光”.pdf"
+    raw_chunks = [c.strip() for c in SENT_SPLIT_PATTERN.split(text) if c.strip()]
-    xlsx_path = r"C:\Users\86166\Desktop\Agent_Learning_minor\电科二班2025德育评价.xlsx"
+    if not raw_chunks:
-    py_path = r"C:\Users\86166\Desktop\Agent_Learning_minor\speech_server.py"
+        return [text]
+
+    merged = []
+    buffer = ""
+    for chunk in raw_chunks:
+        buffer += chunk
+        # 当累积的有效字符数达到阈值，或已是最后一个块时，切出一块
+        if effective_length(buffer) >= min_effective_chars:
+            merged.append(buffer)
+            buffer = ""
+    if buffer:  # 剩余部分（必然小于阈值）
+        if merged:
+            merged[-1] += buffer   # 合并到最后一块
+        else:
+            merged.append(buffer)  # 整段文本都太短，保留原样
+    return merged
     
-    output_path = "test_results.txt"
+def merge_wavs(audio_parts: list[bytes]) -> bytes:
-    with open(output_path, "w", encoding="utf-8") as f:
+    """将多个 WAV 字节串按顺序拼接成一个 WAV"""
-        test_files = [
+    if len(audio_parts) == 1:
-            ("py", py_path),
+        return audio_parts[0]
-            ("docx", docx_path),
+    import io, wave
-            ("xlsx", xlsx_path),
+    data_frames = []
-            ("pdf", pdf_path),
+    params = None
-        ]
+    for part in audio_parts:
-        for label, path in test_files:
+        with wave.open(io.BytesIO(part), 'rb') as wf:
-            f.write(f"\n===== 测试 {label} 文件：{path} =====\n")
+            if params is None:
-            try:
+                params = wf.getparams()
-                text, images = Documents_process.process(path)
+            data_frames.append(wf.readframes(wf.getnframes()))
-                f.write(f"提取文本长度：{len(text)}\n")
+    buf = io.BytesIO()
-                f.write(f"提取图片数量：{len(images)}\n")
+    with wave.open(buf, 'wb') as wf_out:
-                f.write("文本预览：\n")
+        wf_out.setparams(params)
-                f.write(text)
+        for frames in data_frames:
-            except Exception as e:
+            wf_out.writeframes(frames)
-                f.write(f"❌ 处理失败: {e}\n")
+    return buf.getvalue()
-    print(f"测试结果已写入 {output_path}")
-
 
     

src/web/server.py

@@ -27,21 +27,13 @@ if str(_SRC) not in sys.path:
 from agent import chatbot_agent as ca  # noqa: E402
 from web import ui_session as ui  # noqa: E402
 from agent.utils.image_utils import IMAGE_FILE_EXTENSIONS  # noqa: E402
+from agent.utils.agent_utils import DOC_EXT_MAP  # noqa: E402
 
 _WEB_DIR = Path(__file__).resolve().parent
 _IMAGE_DIR = _WEB_DIR / "image"
 
 _AUDIO_EXT = frozenset({".wav", ".mp3", ".m4a", ".flac", ".ogg", ".webm"})
-_IMAGE_EXT = frozenset(IMAGE_FILE_EXTENSIONS)
+_FILE_EXT = frozenset(DOC_EXT_MAP.keys())
-
-
-def _infer_upload_kind(fp: str) -> str | None:
-    ext = os.path.splitext(fp)[1].lower()
-    if ext in _IMAGE_EXT:
-        return "image"
-    if ext in _AUDIO_EXT:
-        return "audio"
-    return "file"
 
 
 def _data_url_for_file(fp: str) -> str | None:
@@ -92,7 +84,6 @@ def _serialize_content(content: Any, *, meta: dict[str, Any] | None = None, role
             parts.append({"type": "text", "text": piece.get("text") or ""})
             continue
         fp = piece.get("path") or (piece.get("file") or {}).get("path")
-        display_name = piece.get("name")
         if piece.get("type") == "image_url":
             fp = fp or (piece.get("image_url") or {}).get("url")
         if piece.get("type") == "audio":
@@ -104,19 +95,19 @@ def _serialize_content(content: Any, *, meta: dict[str, Any] | None = None, role
             continue
         ext = os.path.splitext(fp)[1].lower()
         if ext in IMAGE_FILE_EXTENSIONS:
-            parts.append({"type": "image", "url": data_url, "path": fp, "name": display_name or os.path.basename(fp)})
+            parts.append({"type": "image", "url": data_url, "path": fp, "name": os.path.basename(fp)})
         elif ext in _AUDIO_EXT:
             audio_meta = meta.get("audio") if isinstance(meta.get("audio"), dict) else {}
             parts.append({
                 "type": "audio",
                 "url": data_url,
                 "path": fp,
-                "name": display_name or os.path.basename(fp),
+                "name": os.path.basename(fp),
                 "autoplay": bool(meta.get("autoplay") or audio_meta.get("autoplay")) if role == "assistant" else False,
                 "duration_seconds": audio_meta.get("duration_seconds"),
             })
         else:
-            parts.append({"type": "file", "url": data_url, "path": fp, "name": display_name or os.path.basename(fp)})
+            parts.append({"type": "file", "url": data_url, "path": fp, "name": os.path.basename(fp)})
     return parts
 
 
@@ -253,7 +244,6 @@ def _materialize_uploads_to_temp(
     """将上传文件落盘到临时目录，返回 ``(query_dict, tmp_root)``；调用方须在 finally 中删除 ``tmp_root``。"""
     tmp_root = tempfile.mkdtemp(prefix="web_mm_")
     file_paths: list[str] = []
-    file_names: list[str] = []
     for uf in images or []:
         if uf is None or not uf.filename:
             continue
@@ -262,7 +252,6 @@ def _materialize_uploads_to_temp(
         with open(dest, "wb") as f:
             shutil.copyfileobj(uf.file, f)
         file_paths.append(dest)
-        file_names.append(uf.filename)
 
     for uf in files or []:
         if uf is None or not uf.filename:
@@ -272,7 +261,6 @@ def _materialize_uploads_to_temp(
         with open(dest, "wb") as f:
             shutil.copyfileobj(uf.file, f)
         file_paths.append(dest)
-        file_names.append(uf.filename)
 
     if audio is not None and audio.filename:
         ext = Path(audio.filename).suffix or ".bin"
@@ -280,9 +268,8 @@ def _materialize_uploads_to_temp(
         with open(dest, "wb") as f:
             shutil.copyfileobj(audio.file, f)
         file_paths.append(dest)
-        file_names.append(audio.filename)
 
-    query = {"text": (text or "").strip(), "files": file_paths, "file_names": file_names}
+    query = {"text": (text or "").strip(), "files": file_paths}
     return query, tmp_root