修复大模型思考输出问题

This commit is contained in:
2026-06-11 14:17:19 +08:00
parent 45bf49998c
commit f1409f9237
3 changed files with 223 additions and 80 deletions

View File

@@ -13,7 +13,7 @@ from typing import Optional
import httpx import httpx
from config import Config from config import Config
from bot.reply_parser import extract_final_reply from bot.reply_parser import parse_reply
logger = logging.getLogger(__name__) logger = logging.getLogger(__name__)
@@ -72,6 +72,46 @@ def clear_session(user_id: str):
_user_sessions.pop(user_id, None) _user_sessions.pop(user_id, None)
# ============================================================
# 显示格式化
# ============================================================
def _format_turn_for_display(text: str, use_emoji: bool = True) -> str:
"""
将一轮的原始文本( thinking... response...)格式化为微信展示文本。
use_emoji=True 用 💭 前缀(流式阶段,企微会渲染为可折叠的"已深度思考"
use_emoji=False 用纯文本标题(最终回复,避免触发"思考中"指示器)
"""
if not text:
return ""
prefix = "💭 思考过程:" if use_emoji else "【思考过程】"
# 查找 response 分隔标记
parts = text.split('\n response\n', 1)
if len(parts) == 2:
think_raw, body_raw = parts
think_clean = think_raw.strip()
if think_clean.startswith(' thinking'):
think_clean = think_clean[len(' thinking'):].lstrip('\n')
think_clean = think_clean.strip()
body_clean = body_raw.strip()
result = ""
if think_clean:
result = f"{prefix}\n{think_clean}\n\n"
if body_clean:
result += body_clean
return result
else:
# 还没有 response全是思考内容
think_clean = text.strip()
if think_clean.startswith(' thinking'):
think_clean = think_clean[len(' thinking'):].lstrip('\n')
think_clean = think_clean.strip()
return f"{prefix}\n{think_clean}" if think_clean else ""
# ============================================================ # ============================================================
# Agent 调用 # Agent 调用
# ============================================================ # ============================================================
@@ -118,20 +158,44 @@ async def call_agent_stream(
return await call_agent_sync(user_id, message) return await call_agent_sync(user_id, message)
stream_id = generate_req_id("stream") stream_id = generate_req_id("stream")
full_text = "" full_text = "" # 累积非 STOP 的增量文本
turn_start = 0 # 当前轮在 full_text 中的起始位置
stop_texts: list[str] = [] # 每轮 STOP 时的完整文本快照
event_count = 0
last_sent_display = "" # 上次已发送到微信的展示文本
has_sent_final = False has_sent_final = False
has_sent_thinking = False first_event_sent = False # 是否已发送过流式更新用于5秒超时保护
async for line in response.aiter_lines(): async for line in response.aiter_lines():
if not line.startswith("data: "): if not line.startswith("data: "):
continue continue
data_str = line[6:] data_str = line[6:]
event_count += 1
try: try:
data = json.loads(data_str) data = json.loads(data_str)
except json.JSONDecodeError: except json.JSONDecodeError:
logger.warning("SSE #%d JSON 解析失败: %s", event_count, data_str[:200])
continue continue
finish_reason = data.get("finishReason", "")
content = data.get("content", {})
parts = content.get("parts", [])
# 日志:每个 SSE 事件的摘要
part_summary = []
for p in parts:
t = p.get("text", "")
part_summary.append(
f"text_len={len(t)}, thought={p.get('thought', False)}, "
f"preview={t[:50]!r}"
)
logger.info(
"SSE #%d | finish=%s | parts=%d | %s",
event_count, finish_reason, len(parts),
" | ".join(part_summary) if part_summary else "(无parts)",
)
# 错误处理 # 错误处理
if "error" in data: if "error" in data:
logger.error("Agent SSE 错误: %s", data["error"]) logger.error("Agent SSE 错误: %s", data["error"])
@@ -139,51 +203,82 @@ async def call_agent_stream(
clear_session(user_id) clear_session(user_id)
return "抱歉,处理消息时出现错误,请稍后重试。" return "抱歉,处理消息时出现错误,请稍后重试。"
# 收集所有文本(含思考过程) is_final = finish_reason == "STOP"
parts = data.get("content", {}).get("parts", [])
for part in parts:
text = part.get("text", "")
if not text:
continue
if part.get("thought"):
continue
full_text += text
# 只发送一次"正在思考"状态 if is_final:
if full_text and not has_sent_thinking and not has_sent_final: # STOP 事件ADK 会重复发送完整文本,不追加到 full_text
try: # 保存当前轮的文本快照
await ws_client.reply_stream( turn_text = full_text[turn_start:]
frame, stream_id, "💭 正在思考...", False if turn_text:
stop_texts.append(turn_text)
logger.info(
">>> STOP #%d,当前轮文本长度: %d前200字: %s",
len(stop_texts), len(turn_text), turn_text[:200],
) )
has_sent_thinking = True turn_start = len(full_text)
except Exception: last_sent_display = "" # 新轮次重置展示缓存
pass else:
# 非 STOP增量文本
for part in parts:
text = part.get("text", "")
if text:
full_text += text
# 最终完成时提取正式回复并发送 # 流式更新微信消息
is_final = data.get("finishReason") == "STOP" turn_text = full_text[turn_start:]
if is_final and full_text and not has_sent_final: display = _format_turn_for_display(turn_text)
clean_reply = extract_final_reply(full_text) if display and display != last_sent_display:
logger.info("Agent 回复完成,长度: %d", len(clean_reply)) try:
try: await ws_client.reply_stream(
await ws_client.reply_stream( frame, stream_id, display, False
frame, stream_id, clean_reply, True )
) last_sent_display = display
has_sent_final = True first_event_sent = True
except Exception as e: except Exception:
logger.warning("推送最终回复失败: %s", e) pass
# 兜底:如果流结束但未发送最终回复 # === SSE 流结束发送最终回复 ===
if full_text and not has_sent_final: logger.info(
clean_reply = extract_final_reply(full_text) "=== SSE 流结束 === %d 个事件, %d 个 STOP, full_text 总长: %d",
logger.info("Agent SSE 兜底发送,长度: %d", len(clean_reply)) event_count, len(stop_texts), len(full_text),
)
for idx, st in enumerate(stop_texts):
tc, bc = parse_reply(st)
logger.info(
"STOP[%d] 解析: 思考=%d, 正文=%d, 正文前100字: %s",
idx, len(tc), len(bc), bc[:100] if bc else "(空)",
)
# 从最后一个有正文的 STOP 构建最终内容
final_body = ""
for stop_text in reversed(stop_texts):
_, body_content = parse_reply(stop_text)
if body_content:
final_body = body_content
break
# 兜底
if not final_body and stop_texts:
_, final_body = parse_reply(stop_texts[-1])
if not final_body and full_text:
_, final_body = parse_reply(full_text[turn_start:])
if not final_body and full_text:
final_body = full_text[turn_start:]
if final_body:
try: try:
# 用纯正文 + finish=True 正常关闭流
# 思考过程在流式阶段已通过 💭 展示finish=True 替换为正文
await ws_client.reply_stream( await ws_client.reply_stream(
frame, stream_id, clean_reply, True frame, stream_id, final_body, True
) )
has_sent_final = True
logger.info(">>> 最终回复发送成功,正文长度: %d", len(final_body))
except Exception as e: except Exception as e:
logger.warning("推送兜底回复失败: %s", e) logger.warning("推送最终回复失败: %s", e)
return extract_final_reply(full_text) if full_text else "" _, body = parse_reply(full_text[turn_start:] if turn_start else full_text)
return body if body else ""
except httpx.ConnectError: except httpx.ConnectError:
logger.error("无法连接到 Agent 服务: %s", Config.AGENT_BASE_URL) logger.error("无法连接到 Agent 服务: %s", Config.AGENT_BASE_URL)
@@ -236,4 +331,4 @@ async def call_agent_sync(user_id: str, message: str) -> str:
return "抱歉,服务暂时不可用,请稍后重试。" return "抱歉,服务暂时不可用,请稍后重试。"
except Exception as e: except Exception as e:
logger.error("同步调用 Agent 失败: %s", e) logger.error("同步调用 Agent 失败: %s", e)
return "抱歉,处理消息时出现错误,请稍后重试。" return "抱歉,处理消息时出现错误,请稍后重试。"

View File

@@ -9,8 +9,6 @@
""" """
import logging import logging
from aibot import generate_req_id
from bot.agent_client import call_agent_stream from bot.agent_client import call_agent_stream
from bot.message_parser import extract_metadata, extract_text_content from bot.message_parser import extract_metadata, extract_text_content
@@ -33,11 +31,8 @@ async def handle_message(ws_client, frame: dict, msg_type: str):
# ---- 调用 Agent ---- # ---- 调用 Agent ----
reply = await call_agent_stream(ws_client, frame, user_id, content) reply = await call_agent_stream(ws_client, frame, user_id, content)
# ---- 兜底:未收到回复 ----
if not reply: if not reply:
stream_id = generate_req_id("stream") logger.warning("Agent 未返回有效回复,用户: %s", user_id)
await ws_client.reply_stream(frame, stream_id, "(未收到回复)", True)
async def _resolve_content( async def _resolve_content(

View File

@@ -1,55 +1,108 @@
""" """
回复文本解析模块 回复文本解析模块
负责处理 Agent 返回的原始文本,提取最终正式回复 负责处理 Agent 返回的原始文本,分离思考内容和正文
模型思考标记格式(按优先级检测 SSE 实际输出格式(从日志确认
1. <think...</think 标签对MiniMax M2.7 等模型 - " thinking\\n...思考内容...\\n response\\n...正文内容..." (纯文本标记
2. 💭 emoji 前缀(旧版格式,保留兼容) - 思考部分: " thinking\\n" 开头
- 正文部分: "\\n response\\n" 之后的所有内容
ADK 最后一个 event 会重复发送完整内容,需要去重 多轮调用时 full_text 会累加,调用方需要自行切分轮次
""" """
import logging
import re import re
logger = logging.getLogger(__name__) # 思考-正文分隔标记:出现在文本中时,之前是思考,之后是正文
_RESPONSE_MARKER = re.compile(r'\n response\n')
def extract_final_reply(text: str) -> str: def parse_reply(text: str) -> tuple[str, str]:
""" """
SSE 收集的完整文本中提取最终正式回复 文本中分离思考内容和正文
策略 支持两种格式
1. 去除 <think...</think 标签及其内容 1. " thinking\\n...\\n response\\n..." SSE 实际纯文本格式)
2. 兼容 💭 emoji 分割(旧版格式 2. " thinking... response" HTML 标签风格
3. 去重
返回: (思考内容, 正文内容)
""" """
if not text: if not text:
return text return "", ""
# 步骤1去除 <think...</think 标签块 think_content = ""
if "<think" in text: body_content = text
text = re.sub(r"<think[\s\S]*?</think\s*>?", "", text)
# 步骤2兼容 💭 emoji 分割(旧版格式) # 格式1匹配 " thinking\\n ... \\n response\\n" 纯文本标记
if "💭" in text: # 从 SSE 日志确认: 思考内容以 "\\n response\\n" 结束
text = text.split("💭")[-1] parts = _RESPONSE_MARKER.split(text, maxsplit=1)
if len(parts) == 2:
think_part, body_part = parts
# 去掉开头的 " thinking\\n"
think_part = think_part.strip()
if think_part.startswith(' thinking'):
think_part = think_part[len(' thinking'):].lstrip('\n')
think_content = think_part.strip()
body_content = body_part.strip()
# 清理前导空白 # 格式2匹配 " thinking... response" HTML 标签风格(兜底)
text = text.strip() if not think_content and "<think" in text:
match = re.search(r"<think([\s\S]*?)</think\s*>", text)
if match:
think_content = match.group(1).strip()
body_content = re.sub(r"<think[\s\S]*?</think\s*>", "", text).strip()
else:
think_content = text.replace("<think", "").strip()
body_content = ""
# 步骤3去重 # 去重
text = _deduplicate(text) body_content = _deduplicate(body_content)
think_content = _deduplicate(think_content)
return text return think_content, body_content
def extract_body_only(text: str) -> str:
"""
从文本中只提取正文内容(不含思考)。
用于流式更新时过滤思考内容。
如果文本中还没有出现 "\\n response\\n" 标记,
说明还在思考阶段,返回空字符串避免把思考内容当正文发送。
"""
if not text:
return ""
# 检查正文标记是否已出现
split_pos = _find_response_marker(text)
if split_pos < 0:
# 还没出现 response 标记,仍在思考阶段,不返回任何内容
return ""
# 返回 response 之后的内容
body = text[split_pos + 1:]
# 去掉开头的 response\n
body = body.lstrip('\n')
if body.startswith(' response'):
body = body[len(' response'):].lstrip('\n')
# 再次检查是否有新的 response 标记(多轮场景)
pos2 = _find_response_marker(body)
if pos2 >= 0:
body = body[pos2 + 1:]
body = body.lstrip('\n')
if body.startswith(' response'):
body = body[len(' response'):].lstrip('\n')
return body.strip()
def _find_response_marker(text: str) -> int:
"""查找 \\n response\\n 标记的位置,返回 response 中 'r' 的索引"""
m = _RESPONSE_MARKER.search(text)
return m.start() if m else -1
def _deduplicate(text: str) -> str: def _deduplicate(text: str) -> str:
""" """对回复文本去重。"""
对回复文本去重。
ADK 最后一个 SSE event 会重复发送完整内容(含思考+回复),
去除 <think...> 后可能出现连续两份相同的正式回复。
支持行级和字符级去重。
"""
if not text: if not text:
return text return text
@@ -66,13 +119,13 @@ def _deduplicate(text: str) -> str:
if non_empty[-overlap_len:] == non_empty[:overlap_len]: if non_empty[-overlap_len:] == non_empty[:overlap_len]:
return "\n".join(non_empty[overlap_len:]) return "\n".join(non_empty[overlap_len:])
# 字符级去重:检查文本是否由两个相同的子串拼接而成 # 字符级去重
length = len(text) length = len(text)
for sub_len in range(length // 2, 0, -1): for sub_len in range(length // 2, 0, -1):
if length % sub_len == 0 and text[:sub_len] * (length // sub_len) == text: if length % sub_len == 0 and text[:sub_len] * (length // sub_len) == text:
return text[:sub_len] return text[:sub_len]
# 部分重叠:前缀和后缀相同,中间有重复 # 部分重叠
for overlap in range(min(length // 2, 500), 0, -1): for overlap in range(min(length // 2, 500), 0, -1):
if text[:overlap] == text[-overlap:]: if text[:overlap] == text[-overlap:]:
trimmed = text[overlap:] trimmed = text[overlap:]
@@ -81,4 +134,4 @@ def _deduplicate(text: str) -> str:
if trimmed_len % sub_len == 0 and trimmed[:sub_len] * (trimmed_len // sub_len) == trimmed: if trimmed_len % sub_len == 0 and trimmed[:sub_len] * (trimmed_len // sub_len) == trimmed:
return trimmed[:sub_len] return trimmed[:sub_len]
return text return text