450 lines
15 KiB
Python
450 lines
15 KiB
Python
#!/usr/bin/env python3
|
||
"""
|
||
批量回写 2026-05-11 ~ 2026-05-19 的问题描述（AI 归纳版）
|
||
|
||
对每个日期：
|
||
1. 从 MySQL 读取消息 → 聚类 → 生成问题簇
|
||
2. 调用 DeepSeek API 为每个簇生成精炼问题描述
|
||
3. 用 AI 描述重新生成完整归纳内容（替代脚本默认的 generate_problem_description）
|
||
4. 覆盖写入飞书知识库对应日期的子文档
|
||
"""
|
||
|
||
import sys, os, json, urllib.request, subprocess, time, re
|
||
from datetime import datetime, date, timedelta
|
||
from collections import defaultdict
|
||
|
||
sys.path.insert(0, os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "skills", "feishu-feedback-sync", "scripts"))
|
||
|
||
from sync_feishu_feedback import (
|
||
get_db_connection, sort_threads, sort_cluster_msgs,
|
||
extract_location_elements, extract_conclusion, classify_problem,
|
||
get_tenant_token, list_child_nodes, create_child_doc,
|
||
SUMMARY_PARENT_NODE, SUMMARY_SPACE_ID, DISPATCH_CRED_DIR, XIAOKUI_BOT_OPEN_ID,
|
||
CLI, get_env,
|
||
)
|
||
from priority_classifier import compute_final_priority, sort_by_priority
|
||
|
||
# === 配置 ===
|
||
DEEPSEEK_API_KEY = "sk-7cf94305fb12473b956fd2ed2a6db05b"
|
||
DEEPSEEK_BASE_URL = "https://api.deepseek.com/v1"
|
||
DEEPSEEK_MODEL = "deepseek-v4-pro"
|
||
|
||
START_DATE = "2026-05-11"
|
||
END_DATE = "2026-05-19"
|
||
|
||
SYSTEM_PROMPT = """你是一个游戏产品的问题归纳助手。阅读来自测试群的多人对话，用一句中文描述他们讨论的具体问题。
|
||
|
||
要求：
|
||
1. 只描述问题本身，不评价、不建议
|
||
2. 包含关键要素：在哪个端/环节、什么表现
|
||
3. 有频率信息（偶现/频繁/必现）要体现
|
||
4. 仅输出一句中文，不加任何前缀、编号、引号或换行
|
||
5. 如果对话全是无实质内容的闲聊，输出"无明确问题"
|
||
6. 如果是打包/热更类问题，说清楚是哪个版本/分支的包
|
||
7. 如果是语音识别问题，说清楚识别什么内容、识别成了什么"""
|
||
|
||
|
||
def query_date_messages(date_str):
|
||
"""读取指定日期消息"""
|
||
conn = get_db_connection()
|
||
cursor = conn.cursor()
|
||
next_date = (datetime.strptime(date_str, "%Y-%m-%d") + timedelta(days=1)).strftime("%Y-%m-%d")
|
||
cursor.execute(
|
||
"""SELECT message_id, sender_name, msg_type, content, media_url, quote_message_id,
|
||
msg_time, msg_timestamp
|
||
FROM lark_group_message
|
||
WHERE msg_time >= %s AND msg_time < %s
|
||
ORDER BY msg_time ASC""",
|
||
(f"{date_str} 00:00:00", f"{next_date} 00:00:00"),
|
||
)
|
||
rows = cursor.fetchall()
|
||
conn.close()
|
||
# 格式化时间为字符串
|
||
formatted_rows = []
|
||
for r in rows:
|
||
r_list = list(r)
|
||
if r_list[6] and hasattr(r_list[6], 'strftime'):
|
||
r_list[6] = r_list[6].strftime('%Y-%m-%d %H:%M:%S')
|
||
formatted_rows.append(tuple(r_list))
|
||
return formatted_rows
|
||
|
||
|
||
def build_ai_prompt(cluster):
|
||
"""为单个问题簇构建 LLM prompt"""
|
||
lines = []
|
||
lines.append(f"优先级: {cluster.get('priority', '?')}")
|
||
lines.append(f"分类: {cluster.get('category', '?')}")
|
||
lines.append(f"排查结论: {cluster.get('conclusion', '无')}")
|
||
lines.append("--- 对话 ---")
|
||
|
||
for msg in cluster.get("messages", []):
|
||
sender = msg.get("sender", "?")
|
||
content = msg.get("content", "").strip()
|
||
mtype = msg.get("msg_type", "text")
|
||
t = msg.get("time", "")
|
||
|
||
if mtype in ("image", "post_image", "media", "file") and not content:
|
||
continue
|
||
if not content:
|
||
continue
|
||
if len(content) > 200:
|
||
content = content[:197] + "..."
|
||
|
||
lines.append(f"[{t}] {sender}: {content}")
|
||
|
||
return "\n".join(lines)
|
||
|
||
|
||
def call_deepseek(user_prompt, max_retries=2):
|
||
"""调用 DeepSeek 生成问题描述"""
|
||
body = json.dumps(
|
||
{
|
||
"model": DEEPSEEK_MODEL,
|
||
"messages": [
|
||
{"role": "system", "content": SYSTEM_PROMPT},
|
||
{"role": "user", "content": user_prompt},
|
||
],
|
||
"temperature": 0.3,
|
||
"max_tokens": 256,
|
||
}
|
||
).encode()
|
||
|
||
for attempt in range(max_retries + 1):
|
||
try:
|
||
req = urllib.request.Request(
|
||
f"{DEEPSEEK_BASE_URL}/chat/completions",
|
||
data=body,
|
||
headers={"Authorization": f"Bearer {DEEPSEEK_API_KEY}", "Content-Type": "application/json"},
|
||
method="POST",
|
||
)
|
||
resp = urllib.request.urlopen(req, timeout=60)
|
||
data = json.loads(resp.read())
|
||
content = data["choices"][0]["message"]["content"].strip()
|
||
content = content.strip('"\'""'' \n')
|
||
return content
|
||
except Exception as e:
|
||
if attempt < max_retries:
|
||
print(f" ⚠️ 重试 {attempt + 1}: {e}")
|
||
time.sleep(3)
|
||
else:
|
||
raise
|
||
|
||
|
||
def is_valid_description(desc):
|
||
"""验证 AI 生成的描述是否有效"""
|
||
if not desc or not desc.strip():
|
||
return False
|
||
# 过滤明显的垃圾输出
|
||
garbage_patterns = [
|
||
r'^(没有回答|没有回复|不知道|不确定|无法判断|不太清楚)[。！]?$',
|
||
r'^(没有回答。|没有回复。){3,}', # 重复"没有回答"
|
||
r'^[。！，\s]+$', # 纯标点
|
||
r'^[?？]+$', # 纯问号
|
||
]
|
||
for pat in garbage_patterns:
|
||
if re.search(pat, desc.strip()):
|
||
return False
|
||
# 太短（<3个中文字符）
|
||
chinese_chars = len(re.findall(r'[\u4e00-\u9fff]', desc))
|
||
if chinese_chars < 3:
|
||
return False
|
||
return True
|
||
|
||
|
||
def generate_ai_description(cluster_data, cluster_msgs):
|
||
"""为单个簇调用 AI 生成描述，失败时回退到关键词生成"""
|
||
prompt = build_ai_prompt(cluster_data)
|
||
try:
|
||
desc = call_deepseek(prompt)
|
||
if is_valid_description(desc):
|
||
return desc
|
||
else:
|
||
print(f" ⚠️ AI 输出无效，使用关键词 fallback")
|
||
# 回退到关键词生成
|
||
from sync_feishu_feedback import generate_problem_description, extract_location_elements
|
||
loc = extract_location_elements(cluster_msgs)
|
||
return generate_problem_description(cluster_msgs, loc, "")
|
||
except Exception as e:
|
||
print(f" ❌ AI 调用失败: {e}，使用关键词 fallback")
|
||
from sync_feishu_feedback import generate_problem_description, extract_location_elements
|
||
loc = extract_location_elements(cluster_msgs)
|
||
return generate_problem_description(cluster_msgs, loc, "")
|
||
|
||
|
||
def build_summary_markdown(valid_clusters):
|
||
"""
|
||
用 AI 描述重新生成归纳内容（替代 generate_summary）
|
||
valid_clusters: list of {cluster_id, msgs, earliest_time, priority_info, ai_description, category}
|
||
"""
|
||
lines = ["## 今日问题归纳\n"]
|
||
|
||
# 按优先级+分类分组
|
||
grouped = defaultdict(list)
|
||
for vc in valid_clusters:
|
||
p = vc.get("priority_info", {}).get("priority", "P2")
|
||
grouped[p].append(vc)
|
||
|
||
priority_headers = {
|
||
"P0": "⚠️ P0级核心问题（需优先处理）",
|
||
"P1": "⚡ P1级重要问题",
|
||
"P2": "📌 P2级一般问题",
|
||
"P3": "📝 P3级低优先级",
|
||
}
|
||
|
||
for p_level in ["P0", "P1", "P2", "P3"]:
|
||
items = grouped.get(p_level, [])
|
||
if not items:
|
||
continue
|
||
lines.append(f"**{priority_headers[p_level]}**")
|
||
|
||
by_category = defaultdict(list)
|
||
for vc in items:
|
||
by_category[vc["category"]].append(vc)
|
||
|
||
cat_idx = 0
|
||
for cat_name, cat_items in by_category.items():
|
||
cat_idx += 1
|
||
lines.append(f"{cat_idx}. **{cat_name}**")
|
||
for vc in cat_items:
|
||
desc = vc.get("ai_description", "") or vc.get("fallback_description", "未知问题")
|
||
lines.append(f" - {desc}")
|
||
lines.append("")
|
||
|
||
# 问题拆解
|
||
lines.append("## 今日问题拆解\n")
|
||
|
||
idx = 0
|
||
for vc in valid_clusters:
|
||
idx += 1
|
||
pi = vc.get("priority_info", {})
|
||
priority_label = pi.get("priority", "P2")
|
||
emoji = pi.get("emoji", "📌")
|
||
desc = vc.get("ai_description", "") or "未知问题"
|
||
|
||
sorted_msgs = sort_cluster_msgs(vc["msgs"])
|
||
|
||
lines.append(f"### {emoji} {priority_label}")
|
||
lines.append("")
|
||
lines.append(f"**{idx}，问题描述：** {desc}")
|
||
lines.append("")
|
||
conclusion = extract_conclusion(sorted_msgs)
|
||
lines.append(conclusion)
|
||
lines.append("")
|
||
lines.append("| 发言人 | 对话信息 |")
|
||
lines.append("|--------|---------|")
|
||
|
||
first_speaker = sorted_msgs[0][1]
|
||
last_speaker = sorted_msgs[-1][1]
|
||
seen_speakers = set()
|
||
|
||
for i, m in enumerate(sorted_msgs):
|
||
name = m[1]
|
||
text = str(m[3]).replace("\n", " ").replace("\r", " ").strip() if m[3] else ""
|
||
text = re.sub(r"\[Image:[^\]]+\]", "", text)
|
||
text = re.sub(r"https?://\S+", "", text)
|
||
text = re.sub(r"\s+", " ", text)
|
||
media_url = str(m[4]) if m[4] else ""
|
||
|
||
info_parts = []
|
||
if text:
|
||
if len(text) > 80:
|
||
text = text[:77] + "..."
|
||
info_parts.append(text)
|
||
if media_url:
|
||
label = "图片" if media_url.lower().endswith((".png", ".jpg", ".jpeg", ".gif", ".webp")) else "文件"
|
||
info_parts.append(f"📎 [{label}]({media_url})")
|
||
if not info_parts:
|
||
info_parts.append("[图片]")
|
||
dialogue_info = "<br>".join(info_parts) if len(info_parts) > 1 else info_parts[0]
|
||
|
||
role_tag = ""
|
||
if name == first_speaker and name not in seen_speakers:
|
||
role_tag = "🚩 报告："
|
||
elif name == last_speaker and i == len(sorted_msgs) - 1:
|
||
role_tag = "✅ "
|
||
|
||
seen_speakers.add(name)
|
||
lines.append(f"| {name} | {role_tag}{dialogue_info} |")
|
||
|
||
lines.append("")
|
||
lines.append("---")
|
||
lines.append("")
|
||
|
||
return "\n".join(lines)
|
||
|
||
|
||
def write_to_kb(date_str, markdown_content):
|
||
"""覆盖写入知识库子文档"""
|
||
title = f"{date_str} 问题反馈"
|
||
children = list_child_nodes()
|
||
|
||
if title in children:
|
||
obj_token = children[title]["obj_token"]
|
||
print(f" 📝 更新: {title}")
|
||
else:
|
||
obj_token = create_child_doc(title)
|
||
if not obj_token:
|
||
children = list_child_nodes()
|
||
if title in children:
|
||
obj_token = children[title]["obj_token"]
|
||
else:
|
||
print(f" ❌ 无法创建/找到子文档: {title}")
|
||
return False
|
||
print(f" ➕ 新建: {title}")
|
||
|
||
# 写入（lark-cli --markdown @file 要求相对路径）
|
||
os.makedirs("tmp", exist_ok=True)
|
||
tmp_md = f"tmp/_xiaokui_backfill_{date_str}.txt"
|
||
with open(tmp_md, "w", encoding="utf-8") as f:
|
||
f.write(markdown_content)
|
||
|
||
env = get_env()
|
||
result = subprocess.run(
|
||
[CLI, "docs", "+update", "--doc", obj_token, "--as", "bot", "--mode", "overwrite", "--markdown", f"@{tmp_md}"],
|
||
env=env,
|
||
capture_output=True,
|
||
text=True,
|
||
timeout=30,
|
||
)
|
||
os.unlink(tmp_md)
|
||
|
||
# lark-cli 输出可能在 stdout 或 stderr
|
||
output = (result.stdout + result.stderr).strip()
|
||
try:
|
||
d = json.loads(output)
|
||
if d.get("ok"):
|
||
print(f" ✅ 写入成功")
|
||
return True
|
||
else:
|
||
print(f" ❌ 写入失败: {d.get('error', {}).get('message', output)[:200]}")
|
||
return False
|
||
except json.JSONDecodeError:
|
||
print(f" ❌ 响应解析失败: {output[:200]}")
|
||
return False
|
||
|
||
|
||
def process_date(date_str, dry_run=False):
|
||
"""处理单个日期"""
|
||
print(f"\n{'=' * 60}")
|
||
print(f"📅 {date_str}")
|
||
print(f"{'=' * 60}")
|
||
|
||
# 1. 读取消息
|
||
rows = query_date_messages(date_str)
|
||
if not rows:
|
||
print(" ⚠️ 无消息")
|
||
return None
|
||
print(f" 📊 {len(rows)} 条消息")
|
||
|
||
# 2. 聚类
|
||
sorted_msgs, clusters, cluster_order = sort_threads(rows)
|
||
print(f" 🔗 {len(clusters)} 个簇")
|
||
|
||
# 3. 收集有效簇（≥2条消息）
|
||
valid_clusters = []
|
||
for cid in cluster_order:
|
||
cmsgs = clusters[cid]
|
||
if len(cmsgs) < 2:
|
||
continue
|
||
|
||
pi = compute_final_priority(cmsgs)
|
||
cat = classify_problem(cmsgs)
|
||
|
||
# 构建簇消息摘要（用于 AI prompt）
|
||
cluster_data = {
|
||
"cluster_id": cid,
|
||
"msgs": cmsgs,
|
||
"earliest_time": min(m[6] for m in cmsgs),
|
||
"priority_info": pi,
|
||
"category": cat,
|
||
"conclusion": extract_conclusion(sort_cluster_msgs(cmsgs)),
|
||
"messages": [
|
||
{
|
||
"sender": m[1],
|
||
"content": str(m[3]) if m[3] else "",
|
||
"msg_type": str(m[2]),
|
||
"media_url": str(m[4]) if m[4] else "",
|
||
"time": str(m[6]),
|
||
}
|
||
for m in cmsgs
|
||
],
|
||
}
|
||
valid_clusters.append(cluster_data)
|
||
|
||
if not valid_clusters:
|
||
print(" ⚠️ 无有效问题簇（需≥2条消息）")
|
||
return None
|
||
|
||
# 按优先级排序
|
||
valid_clusters = sort_by_priority(valid_clusters)
|
||
|
||
# 4. AI 生成描述
|
||
for vc in valid_clusters:
|
||
idx = valid_clusters.index(vc) + 1
|
||
print(f" 🤖 簇 #{idx}/{len(valid_clusters)}...")
|
||
if not dry_run:
|
||
desc = generate_ai_description(vc, vc["msgs"])
|
||
if desc and desc != "（待归纳）":
|
||
vc["ai_description"] = desc
|
||
print(f" ✅ {desc[:80]}...")
|
||
else:
|
||
print(f" ⚠️ AI 失败，使用空描述")
|
||
vc["ai_description"] = "（待归纳）"
|
||
# API 限速保护
|
||
time.sleep(0.5)
|
||
else:
|
||
vc["ai_description"] = f"[DRY-RUN] 待AI归纳"
|
||
print(f" [DRY-RUN]")
|
||
|
||
# 5. 生成完整 markdown
|
||
markdown = build_summary_markdown(valid_clusters)
|
||
|
||
if dry_run:
|
||
print(f"\n --- 预览前 500 字符 ---")
|
||
print(markdown[:500])
|
||
return markdown
|
||
|
||
# 6. 写入知识库
|
||
write_to_kb(date_str, markdown)
|
||
return markdown
|
||
|
||
|
||
def main():
|
||
import argparse
|
||
parser = argparse.ArgumentParser()
|
||
parser.add_argument("--dry-run", action="store_true")
|
||
parser.add_argument("--date", help="仅处理指定日期")
|
||
args = parser.parse_args()
|
||
|
||
if args.date:
|
||
dates = [args.date]
|
||
else:
|
||
start = datetime.strptime(START_DATE, "%Y-%m-%d")
|
||
end = datetime.strptime(END_DATE, "%Y-%m-%d")
|
||
dates = [(start + timedelta(days=i)).strftime("%Y-%m-%d") for i in range((end - start).days + 1)]
|
||
|
||
print(f"🚀 批量 AI 归纳回写 {'[DRY-RUN]' if args.dry_run else ''}")
|
||
print(f" 日期范围: {dates[0]} ~ {dates[-1]} ({len(dates)} 天)")
|
||
|
||
results = {}
|
||
for d in dates:
|
||
try:
|
||
r = process_date(d, dry_run=args.dry_run)
|
||
results[d] = "✅" if r else "⏭️"
|
||
except Exception as e:
|
||
print(f" ❌ 异常: {e}")
|
||
results[d] = f"❌ {e}"
|
||
|
||
if not args.dry_run:
|
||
# 避免 API 频率限制
|
||
time.sleep(1)
|
||
|
||
print(f"\n{'=' * 60}")
|
||
print("📊 汇总:")
|
||
for d, status in results.items():
|
||
print(f" {d}: {status}")
|
||
|
||
|
||
if __name__ == "__main__":
|
||
main()
|