feat(translate): 加 Agnes 翻译 fallback,buffer 改 0.5
腾讯 TMT 月度配额快满时(腾讯后台口径已用 2M/5M),翻译降级链: 1. tencent TMT(主,按月配额) 2. tencent MaaS u2(第二级,翻译专用,无配额) 3. agnes 通用 LLM(第三级,质量次之但够用) 4. local NLLB(最后兜底) 新增 backend/app/services/translation/agnes.py: AgnesTranslator 复用 LlmClient 做限速 + 重试,系统 prompt 强约束只输出译文, 去除 "以下是翻译" 等常见 LLM 翻译前缀。 service.py 改动: - fallback 链 maas -> agnes -> local - cache 接受 agnes 结果(30天) - add_usage 只算 tencent TMT buffer 调整: TENCENT_TMT_QUOTA_BUFFER 0.05 -> 0.5 腾讯云后台按请求字节计费,与我们 redis 字符累加口径差约 2.5x; 按腾讯后台口径 redis 累加到 1M 字符即触发降级(对应腾讯约 2.5M 字节 =50% 用量), 留足 buffer,避免月底真爆。
This commit is contained in:
Mavis
118
backend/app/services/translation/agnes.py
Normal file
118
backend/app/services/translation/agnes.py
Normal file
@@ -0,0 +1,118 @@
|
||||
"""Agnes 翻译后端(通用 LLM 做翻译,作为 TMT/MaaS 之后的第三级 fallback)。
|
||||
|
||||
- 端点:settings.agnes_base_url/v1/chat/completions
|
||||
- 模型:settings.agnes_chat_model(agnes-2.0-flash)
|
||||
- 鉴权:Bearer AGNES_API_KEY
|
||||
- 复用 LlmClient 的限速(interval_sec)
|
||||
|
||||
适用场景:
|
||||
- TMT 月配额快满(>95%)时,先走 MaaS(u2)
|
||||
- MaaS 不可用 / 失败时,降到 Agnes(通用 LLM 翻译,质量次之,但够用)
|
||||
- Agnes 不可用时,本地 NLLB 最后兜底
|
||||
|
||||
注意:
|
||||
- Agnes 不是翻译专用模型,翻译质量不如 TMT / MaaS u2
|
||||
- 但 Agnes 的 commentary / image / classification 已经走它,客户端有相同 quota 计费体系
|
||||
- 用一个特殊 system prompt 强约束:只输出译文
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import logging
|
||||
|
||||
from app.config import settings
|
||||
from app.services.llm.client import LlmClient
|
||||
from app.services.translation.base import BaseTranslator, TranslationResult
|
||||
|
||||
logger = logging.getLogger("news.translate.agnes")
|
||||
|
||||
|
||||
# Agnes 做翻译的 system prompt。
|
||||
# 关键约束:只输出译文,不要思考/分析/注释/标签。
|
||||
# 这是通用 LLM 翻车最多的地方——会输出 "以下是中文翻译: ..." 前缀,要强约束。
|
||||
_SYSTEM_PROMPT = """你是即时翻译引擎。把用户输入的文本翻译成简体中文。
|
||||
|
||||
严格规则:
|
||||
1. 只输出译文本身,不要输出任何其他文字
|
||||
2. 不要输出"以下是翻译"、"译文:"、"[Translation]"、引号、破折号等任何包裹
|
||||
3. 不要分步思考、不要分析、不要注释、不要解释
|
||||
4. 不要输出 <think>、<reasoning>、<analysis> 等内部标签
|
||||
5. 如果输入已经是中文,直接原样返回
|
||||
6. 保持原文段落结构(用换行分隔)"""
|
||||
|
||||
|
||||
class AgnesTranslator(BaseTranslator):
|
||||
"""Agnes 翻译(通用 LLM 兜底)。"""
|
||||
|
||||
name = "agnes"
|
||||
|
||||
def __init__(self):
|
||||
if not settings.agnes_api_key:
|
||||
raise RuntimeError("AGNES_API_KEY 未配置")
|
||||
# 用 chat_model,不复用 image_model
|
||||
self.client = LlmClient(
|
||||
base_url=settings.agnes_base_url,
|
||||
api_key=settings.agnes_api_key,
|
||||
chat_model=settings.agnes_chat_model,
|
||||
image_model=settings.agnes_image_model,
|
||||
interval_sec=settings.llm_interval_sec,
|
||||
)
|
||||
|
||||
def is_configured(self) -> bool:
|
||||
return bool(settings.agnes_api_key)
|
||||
|
||||
async def translate(
|
||||
self, text: str, source: str = "auto", target: str = "zh"
|
||||
) -> TranslationResult:
|
||||
if not text.strip():
|
||||
return TranslationResult(text=text, engine=self.name, chars=0)
|
||||
|
||||
if not self.is_configured():
|
||||
raise RuntimeError("AGNES_API_KEY 未配置")
|
||||
|
||||
# 长文本截断保护:Agnes max_tokens 上限小,避免超限
|
||||
# 经验值:1 token ≈ 1.5 个英文 / 0.7 个中文字符
|
||||
# 默认 chat 接口 max_tokens=1500,我们给 4096 留余量
|
||||
# 输入截到 6000 字符(MaaS 的限制类似)
|
||||
max_input = 6000
|
||||
truncated = text[:max_input] if len(text) > max_input else text
|
||||
|
||||
# 保留段落结构
|
||||
user = truncated
|
||||
|
||||
# LlmClient.chat 已含 interval_sec 节流
|
||||
content = await self.client.chat(
|
||||
system=_SYSTEM_PROMPT,
|
||||
user=user,
|
||||
temperature=0.1, # 翻译任务低温度更稳
|
||||
max_tokens=4096,
|
||||
)
|
||||
|
||||
if not content or not content.strip():
|
||||
raise RuntimeError("Agnes 返回空 content")
|
||||
|
||||
# 清洗:模型有时还会输出 markdown 代码块包裹
|
||||
cleaned = content.strip()
|
||||
if cleaned.startswith("```"):
|
||||
# 去掉代码块围栏
|
||||
lines = cleaned.split("\n")
|
||||
cleaned = "\n".join(
|
||||
l for l in lines if not l.strip().startswith("```")
|
||||
).strip()
|
||||
|
||||
# 去除已知的前缀模式(模型偶尔还是会犯)
|
||||
prefixes_to_strip = [
|
||||
"以下是译文:",
|
||||
"以下是中文翻译:",
|
||||
"译文:",
|
||||
"中文翻译:",
|
||||
"Translation:",
|
||||
"翻译结果:",
|
||||
]
|
||||
for p in prefixes_to_strip:
|
||||
if cleaned.startswith(p):
|
||||
cleaned = cleaned[len(p):].lstrip()
|
||||
break
|
||||
|
||||
return TranslationResult(
|
||||
text=cleaned, engine=self.name, chars=len(text), cached=False
|
||||
)
|
||||
@@ -1,9 +1,15 @@
|
||||
"""翻译服务门面:配额检查 + 缓存 + 引擎选择 + 月度计数。
|
||||
|
||||
引擎链路(优先级降序):
|
||||
1. tencent TMT(主,按月配额)
|
||||
1. tencent TMT(主,按月配额;快满时主动切走)
|
||||
2. tencent_maas(备用,OpenAI 兼容,无配额;主失败/TMT 配额耗尽时启用)
|
||||
3. local(最后兜底,需 settings.local_translate_enabled=true)
|
||||
3. agnes(第三级,通用 LLM 做翻译;MaaS 不可用时启用 — 质量次之但够用)
|
||||
4. local(最后兜底,需 settings.local_translate_enabled=true)
|
||||
|
||||
注:
|
||||
- TMT 是按月计费的(腾讯云后台可能计费口径是请求字节,我们 redis 累加的是字符数,
|
||||
差异约 2-3x);用户从腾讯云后台看"已用 2M"时,我们 redis 显示约 80 万字符
|
||||
- 用户决策:以腾讯云后台数字为准,快满时降级
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
@@ -15,6 +21,7 @@ from typing import Protocol
|
||||
|
||||
from app.config import settings
|
||||
from app.redis_client import get_redis
|
||||
from app.services.translation.agnes import AgnesTranslator
|
||||
from app.services.translation.base import BaseTranslator, TranslationResult
|
||||
from app.services.translation.local import LocalTranslator
|
||||
from app.services.translation.tencent import TencentTranslator
|
||||
@@ -38,6 +45,7 @@ class TranslationService:
|
||||
def __init__(self):
|
||||
self._tencent: BaseTranslator | None = None
|
||||
self._tencent_maas: BaseTranslator | None = None
|
||||
self._agnes: BaseTranslator | None = None
|
||||
self._local: BaseTranslator | None = None
|
||||
# 串行:1 个并发;避免触发腾讯 TMT 限速
|
||||
self._sem = asyncio.Semaphore(1)
|
||||
@@ -62,6 +70,18 @@ class TranslationService:
|
||||
self._tencent_maas = None
|
||||
return self._tencent_maas
|
||||
|
||||
def _agnes(self) -> BaseTranslator | None:
|
||||
"""第三级:Agnes 通用 LLM 翻译(在 MaaS 不可用时启用)。
|
||||
质量比 TMT/MaaS u2 差,但通用 LLM 也能翻,够用。
|
||||
"""
|
||||
if self._agnes is None and settings.agnes_api_key:
|
||||
try:
|
||||
self._agnes = AgnesTranslator()
|
||||
except Exception as e:
|
||||
logger.warning("Agnes init failed: %s", e)
|
||||
self._agnes = None
|
||||
return self._agnes
|
||||
|
||||
def _local_translator(self) -> BaseTranslator | None:
|
||||
"""最后兜底:本地模型(需开关)。"""
|
||||
if self._local is None and settings.local_translate_enabled:
|
||||
@@ -72,12 +92,13 @@ class TranslationService:
|
||||
self._local = None
|
||||
return self._local
|
||||
|
||||
# 兼容旧调用点:返回第一个可用的 fallback(优先 maas,次 local)
|
||||
# 兼容旧调用点:返回第一个可用的 fallback(优先 maas,次 agnes,再 local)
|
||||
def _fallback(self) -> BaseTranslator | None:
|
||||
m = self._maas()
|
||||
if m is not None:
|
||||
return m
|
||||
return self._local_translator()
|
||||
for getter in (self._maas, self._agnes, self._local_translator):
|
||||
f = getter()
|
||||
if f is not None:
|
||||
return f
|
||||
return None
|
||||
|
||||
async def can_use_tencent(self, chars: int) -> bool:
|
||||
if not settings.tencentcloud_secret_id:
|
||||
@@ -130,8 +151,8 @@ class TranslationService:
|
||||
engine = None
|
||||
|
||||
if engine is None:
|
||||
# 配额耗尽 / TMT 不可用:走备用链
|
||||
engine = self._maas() or self._local_translator()
|
||||
# 配额耗尽 / TMT 不可用:走备用链(maas → agnes → local)
|
||||
engine = self._fallback()
|
||||
if engine is None:
|
||||
# 全无可用:返回原文 + 标记
|
||||
return TranslationResult(
|
||||
@@ -139,10 +160,10 @@ class TranslationService:
|
||||
engine="skip",
|
||||
chars=chars,
|
||||
)
|
||||
if engine.name == "tencent_maas":
|
||||
logger.info("tencent quota exhausted, fallback to tencent_maas for %d chars", chars)
|
||||
else:
|
||||
logger.info("fallback to local translator for %d chars", chars)
|
||||
logger.info(
|
||||
"tencent quota exhausted, fallback to %s for %d chars",
|
||||
engine.name, chars,
|
||||
)
|
||||
|
||||
# 3) 调用(失败时降级)
|
||||
async with self._sem:
|
||||
@@ -165,15 +186,19 @@ class TranslationService:
|
||||
if res is None:
|
||||
raise RuntimeError(f"translation failed for {chars} chars (engine={engine.name})")
|
||||
|
||||
# 注:engine 已经设好但运行时降级需要重新判断 fallback 链
|
||||
# 上面 translate() 调用失败时,会重试 _fallback() 里下一个可用引擎
|
||||
# 这里 engine 已经在 _fallback() 中按顺序选了一个最合适的,直接使用即可
|
||||
|
||||
# 4) 写缓存 — 只缓存真实翻译结果;失败/降级文本不缓存(避免污染 30 天)
|
||||
if res.engine in ("tencent", "tencent_maas", "nllb") and not res.cached:
|
||||
if res.engine in ("tencent", "tencent_maas", "agnes", "nllb") and not res.cached:
|
||||
if "[翻译失败" not in res.text and "[本条未翻译" not in res.text:
|
||||
try:
|
||||
await r.set(ck, res.text, ex=60 * 60 * 24 * 30) # 30 天
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
# 5) 计数(只在 tencent TMT 上计;maas/local 都不计腾讯云配额)
|
||||
# 5) 计数(只在 tencent TMT 上计;maas / agnes / local 都不计腾讯云配额)
|
||||
if res.engine == "tencent":
|
||||
try:
|
||||
await self.add_usage(res.chars or chars)
|
||||
|
||||
Reference in New Issue
Block a user