797c7c749e
Some checks failed
CD Pipeline / build-and-deploy (push) Failing after 28s
deepseek-r1:14b 思考 token 超過 400 會在 </think> 前截斷，導致 清理後 body 為空，Telegram 顯示空訊息。 - chat_manager: num_predict 400 → 1200 - telegram_gateway: _clean_ai_reply 空值加 fallback 錯誤提示 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
274 lines
11 KiB
Python
274 lines
11 KiB
Python
"""
|
||
AWOOOI Chat Manager - 雙 AI 對話核心
|
||
======================================
|
||
Phase 21.5 初版: 2026-03-31 ogt
|
||
Phase 22.6 重寫: 2026-04-03 ogt (老闆需求: 雙 AI 互動對話)
|
||
Phase 22.7 更新: 2026-04-03 ogt (老闆指示: OpenClaw→Gemini, NemoClaw→Ollama llama3.2:3b)
|
||
Phase 22.8 更新: 2026-04-09 ogt (老闆指示: NemoClaw→Ollama 111 deepseek-r1:14b，SRE 推理更強)
|
||
|
||
架構:
|
||
- OpenClaw (Gemini API): SRE 首席顧問，精準分析
|
||
- NemoClaw (Ollama 192.168.0.111 deepseek-r1:14b): 戰術參謀，深度推理
|
||
|
||
費用控管:
|
||
- Gemini Flash: Input $0.075/1M tokens, Output $0.30/1M tokens
|
||
- NemoClaw: 免費 (本地 Ollama)
|
||
- 每次回覆顯示 token 用量與費用
|
||
- 月上限 $10 USD (由 ai_rate_limiter 控管)
|
||
"""
|
||
|
||
import asyncio
|
||
import structlog
|
||
from src.utils.timezone import now_taipei
|
||
from src.repositories.k8s_repository import get_k8s_repository
|
||
from src.repositories.incident_repository import get_incident_repository
|
||
|
||
logger = structlog.get_logger(__name__)
|
||
|
||
OPENCLAW_PERSONA = """你是 OpenClaw，AWOOOI 平台的 SRE AI 首席顧問。
|
||
個性: 精準、果斷、專業，像老將一樣直接給出建議。
|
||
語氣: 簡短有力，不廢話。繁體中文回應。不超過 300 字。
|
||
稱呼用戶為「老闆」。
|
||
"""
|
||
|
||
NEMOCLAW_PERSONA = """你是 NemoClaw，AWOOOI 平台的 AI 戰術參謀。
|
||
個性: 分析型、從不同角度思考，會質疑假設。
|
||
語氣: 帶點挑釁但建設性。不超過 200 字。
|
||
稱呼用戶為「老闆」。評論 OpenClaw 的回應時，直接說「我補充」或「我有不同看法」。
|
||
強制規則：
|
||
1. 全程使用繁體中文，禁止使用簡體中文、英文或其他語言。
|
||
2. 禁止自稱 DeepSeek 或透露底層模型資訊。你的名字就是 NemoClaw。
|
||
3. 專注於 SRE/DevOps/Kubernetes/可觀測性領域。
|
||
"""
|
||
|
||
|
||
class ChatManager:
|
||
"""AWOOOI 雙 AI 對話管理器"""
|
||
|
||
def __init__(self):
|
||
pass # 2026-04-03 ogt: 移除 repo 實例化，leWOOOgo 規範禁止 Service 持有 repository
|
||
|
||
async def get_system_context(self) -> str:
|
||
"""收集系統即時上下文"""
|
||
now = now_taipei()
|
||
k8s = get_k8s_repository()
|
||
incidents = get_incident_repository()
|
||
|
||
try:
|
||
k8s_status = await k8s.get_pod_status_summary(namespace="awoooi-prod")
|
||
cluster_info = f"Cluster: {k8s_status['running']}/{k8s_status['total']} Pods Running"
|
||
if k8s_status.get('problem_pods'):
|
||
cluster_info += f", {len(k8s_status['problem_pods'])} 異常"
|
||
except Exception:
|
||
cluster_info = "Cluster: 無法取得狀態"
|
||
|
||
try:
|
||
active_incidents = await incidents.get_active()
|
||
if active_incidents:
|
||
lines = [f"- {inc.incident_id}: {inc.status.value} (SEV {inc.severity.value})"
|
||
for inc in active_incidents[:3]]
|
||
incident_summary = "\n".join(lines)
|
||
else:
|
||
incident_summary = "無活躍告警"
|
||
except Exception:
|
||
incident_summary = "無法取得告警"
|
||
|
||
return (
|
||
f"## 系統狀態 ({now.strftime('%Y-%m-%d %H:%M')} 台北)\n"
|
||
f"- {cluster_info}\n"
|
||
f"- 活躍告警: {incident_summary}\n"
|
||
)
|
||
|
||
async def _call_openclaw(self, system_prompt: str, user_message: str) -> str | None:
|
||
"""
|
||
呼叫 OpenClaw 對話 — Gemini Flash API
|
||
|
||
2026-04-03 ogt: 老闆指示改用 Gemini，費用控管月上限 $10 USD
|
||
每次回覆附帶 token 用量與費用統計
|
||
|
||
2026-04-10 Claude Code: 強制合併 OPENCLAW_PERSONA，確保字數限制與格式規範
|
||
"""
|
||
# 強制在 system_prompt 前置 persona，確保 LLM 遵守字數與格式
|
||
system_prompt = f"{OPENCLAW_PERSONA}\n{system_prompt}"
|
||
import httpx
|
||
from src.core.config import get_settings
|
||
settings = get_settings()
|
||
|
||
api_key = settings.GEMINI_API_KEY
|
||
if not api_key:
|
||
logger.warning("openclaw_chat_failed", error="GEMINI_API_KEY not configured")
|
||
return None
|
||
|
||
# 月費用上限檢查 ($10 USD)
|
||
MONTHLY_LIMIT_USD = 10.0
|
||
from src.core.redis_client import get_redis
|
||
from src.utils.timezone import now_taipei
|
||
redis = get_redis()
|
||
month_key = f"gemini_cost:{now_taipei().strftime('%Y-%m')}"
|
||
try:
|
||
current_cost = float(await redis.get(month_key) or 0)
|
||
except Exception:
|
||
current_cost = 0.0
|
||
|
||
if current_cost >= MONTHLY_LIMIT_USD:
|
||
logger.warning("openclaw_gemini_monthly_limit_reached", current_usd=current_cost, limit_usd=MONTHLY_LIMIT_USD)
|
||
return f"🔴 OpenClaw 本月 Gemini 用量已達上限 ${MONTHLY_LIMIT_USD} USD（已用 ${current_cost:.4f}）"
|
||
|
||
# Gemini 2.0 Flash-Lite: 最便宜 (2026-04-03 老闆指示)
|
||
model = "gemini-2.0-flash-lite"
|
||
try:
|
||
async with httpx.AsyncClient(timeout=30.0) as client:
|
||
resp = await client.post(
|
||
f"https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent?key={api_key}",
|
||
json={
|
||
"system_instruction": {"parts": [{"text": system_prompt}]},
|
||
"contents": [{"parts": [{"text": user_message}]}],
|
||
"generationConfig": {"maxOutputTokens": 300, "temperature": 0.7},
|
||
},
|
||
)
|
||
resp.raise_for_status()
|
||
data = resp.json()
|
||
text = data["candidates"][0]["content"]["parts"][0]["text"].strip()
|
||
|
||
# Token/費用統計 + 累計到 Redis
|
||
usage = data.get("usageMetadata", {})
|
||
in_tok = usage.get("promptTokenCount", 0)
|
||
out_tok = usage.get("candidatesTokenCount", 0)
|
||
# Gemini 2.0 Flash-Lite: Input $0.075/1M, Output $0.30/1M
|
||
cost = (in_tok * 0.000000075) + (out_tok * 0.0000003)
|
||
new_total = current_cost + cost
|
||
|
||
try:
|
||
await redis.set(month_key, str(round(new_total, 6)), ex=40 * 24 * 3600) # 40天 TTL
|
||
except Exception:
|
||
pass
|
||
|
||
logger.info("openclaw_gemini_usage", in_tokens=in_tok, out_tokens=out_tok,
|
||
cost_usd=round(cost, 6), monthly_total_usd=round(new_total, 4))
|
||
|
||
return f"{text}\n\n<i>📊 {in_tok+out_tok} tokens | ${cost:.4f} | 本月累計 ${new_total:.4f}</i>"
|
||
except Exception as e:
|
||
logger.warning("openclaw_chat_failed", error=str(e))
|
||
return None
|
||
|
||
async def _call_nemotron(self, system_prompt: str, user_message: str) -> str | None:
|
||
"""
|
||
呼叫 NemoClaw 對話 — Ollama 111 deepseek-r1:14b
|
||
|
||
2026-04-09 ogt: 改接 192.168.0.111 Ollama deepseek-r1:14b，SRE 推理能力最強
|
||
deepseek-r1 含 <think> 標籤，需過濾後才回傳
|
||
|
||
2026-04-10 Claude Code: 強制合併 NEMOCLAW_PERSONA，確保字數限制與格式規範
|
||
"""
|
||
# 強制在 system_prompt 前置 persona
|
||
system_prompt = f"{NEMOCLAW_PERSONA}\n{system_prompt}"
|
||
|
||
import httpx
|
||
import re
|
||
|
||
OLLAMA_URL = "http://192.168.0.111:11434"
|
||
MODEL = "deepseek-r1:14b"
|
||
|
||
try:
|
||
async with httpx.AsyncClient(timeout=120.0) as client:
|
||
resp = await client.post(
|
||
f"{OLLAMA_URL}/api/chat",
|
||
json={
|
||
"model": MODEL,
|
||
"stream": False,
|
||
"messages": [
|
||
{"role": "system", "content": system_prompt},
|
||
{"role": "user", "content": user_message},
|
||
],
|
||
"options": {"num_predict": 1200},
|
||
},
|
||
)
|
||
resp.raise_for_status()
|
||
data = resp.json()
|
||
raw = data.get("message", {}).get("content", "").strip()
|
||
|
||
# 過濾 deepseek-r1 的 <think>...</think> 推理區塊
|
||
text = re.sub(r"<think>.*?</think>", "", raw, flags=re.DOTALL).strip()
|
||
if not text:
|
||
text = raw # 萬一全是 think block，直接回傳原文
|
||
|
||
eval_count = data.get("eval_count", 0)
|
||
prompt_eval_count = data.get("prompt_eval_count", 0)
|
||
total_tokens = eval_count + prompt_eval_count
|
||
|
||
logger.info("nemotron_ollama_usage", model=MODEL,
|
||
prompt_tokens=prompt_eval_count, output_tokens=eval_count)
|
||
|
||
return f"{text}\n\n<i>🦙 {MODEL} | {total_tokens} tokens | 免費</i>"
|
||
except Exception as e:
|
||
logger.warning("nemotron_chat_failed", model=MODEL, error=str(e))
|
||
return None
|
||
|
||
async def generate_response(
|
||
self,
|
||
user_id: int, # noqa: ARG002
|
||
username: str, # noqa: ARG002
|
||
message_text: str,
|
||
) -> str:
|
||
"""
|
||
根據訊息決定回應模式:
|
||
|
||
@openclaw <msg> → 只有 OpenClaw 回應
|
||
@nemo <msg> → 只有 NemoClaw 回應
|
||
其他 → OpenClaw 先回，NemoClaw 異步補充
|
||
"""
|
||
context = await self.get_system_context()
|
||
text = message_text.strip()
|
||
|
||
# 模式 1: 指定 OpenClaw
|
||
if text.lower().startswith("@openclaw"):
|
||
msg = text[9:].strip() or text
|
||
result = await self._call_openclaw(f"{OPENCLAW_PERSONA}\n{context}", msg)
|
||
return f"🦞 <b>OpenClaw:</b>\n{result or '🔴 OpenClaw 無響應'}"
|
||
|
||
# 模式 2: 指定 NemoClaw
|
||
if text.lower().startswith("@nemo"):
|
||
msg = text[5:].strip() or text
|
||
result = await self._call_nemotron(f"{NEMOCLAW_PERSONA}\n{context}", msg)
|
||
return f"🤖 <b>NemoClaw:</b>\n{result or '🔴 NemoClaw 無響應 (NIM 超時)'}"
|
||
|
||
# 模式 3: 雙 AI — OpenClaw 先答，NemoClaw 並行
|
||
openclaw_task = asyncio.create_task(
|
||
self._call_openclaw(f"{OPENCLAW_PERSONA}\n{context}", text)
|
||
)
|
||
nemo_task = asyncio.create_task(
|
||
self._call_nemotron(
|
||
f"{NEMOCLAW_PERSONA}\n{context}",
|
||
f"老闆問了: {text}\n\n請從 NemoClaw 角度補充或評論。",
|
||
)
|
||
)
|
||
|
||
# OpenClaw 最多等 40s（含 context 取得時間），NemoClaw 最多等 60s
|
||
# 2026-04-03 ogt: 移除 asyncio.shield — shield 會在超時後讓 task 繼續跑但無人等待，造成 silent leak
|
||
try:
|
||
openclaw_raw = await asyncio.wait_for(openclaw_task, timeout=40.0)
|
||
except asyncio.TimeoutError:
|
||
openclaw_raw = None
|
||
|
||
openclaw_block = f"🦞 <b>OpenClaw:</b>\n{openclaw_raw or '🔴 無響應'}"
|
||
|
||
try:
|
||
nemo_raw = await asyncio.wait_for(nemo_task, timeout=60.0)
|
||
except asyncio.TimeoutError:
|
||
nemo_raw = None
|
||
|
||
if nemo_raw:
|
||
return f"{openclaw_block}\n\n🤖 <b>NemoClaw:</b>\n{nemo_raw}"
|
||
return openclaw_block
|
||
|
||
|
||
# Singleton
|
||
_chat_manager: ChatManager | None = None
|
||
|
||
|
||
def get_chat_manager() -> ChatManager:
|
||
global _chat_manager
|
||
if _chat_manager is None:
|
||
_chat_manager = ChatManager()
|
||
return _chat_manager
|