mirror of
https://github.com/langgenius/dify.git
synced 2026-11-12 14:21:29 +08:00
Charles Yao
231e423431
Track backend integration, Ruff cleanup and migration reconciliation in ENG-1271. Retain required copy contracts without implementing frontend interactions.
68 lines
3.0 KiB
Python
68 lines
3.0 KiB
Python
"""Detect the Builder reply language; fixed copy is localized by Web i18n."""
|
|
|
|
import logging
|
|
import re
|
|
from collections.abc import Callable
|
|
|
|
from services.dify_builder import trace_diagnostics
|
|
from services.dify_builder.agent import llm
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
_DETECT_SYSTEM = (
|
|
"Identify the language of the user's text. Reply with ONLY its IETF BCP-47 "
|
|
"code, e.g. en, zh-Hans, zh-Hant, ja, fr, es. No other text."
|
|
)
|
|
|
|
# A plausible BCP-47 language tag: a 2-3 letter primary subtag optionally
|
|
# followed by one or more hyphenated subtags (script/region/variant, e.g.
|
|
# "zh-Hans", "zh-Hant-HK"). Guards against a model that ignores "reply with
|
|
# ONLY the code" and returns a sentence instead.
|
|
_LANG_CODE_RE = re.compile(r"^[A-Za-z]{2,3}(-[A-Za-z0-9]{1,8})*$")
|
|
|
|
|
|
class LanguageDetector:
|
|
def __init__(self, model_provider: Callable[[], object | None], *, prior_language: str = "") -> None:
|
|
self._model_provider = model_provider
|
|
self._prior_language = prior_language if _LANG_CODE_RE.fullmatch(prior_language) else ""
|
|
|
|
def _fallback(self, text: str) -> str:
|
|
"""Use clear script evidence, then the persisted conversation locale.
|
|
|
|
This is a presentation fallback, not a model, resource or risk decision.
|
|
Short/ambiguous acknowledgements must not reset the conversation locale.
|
|
"""
|
|
trace_diagnostics.report("deterministic_fallback", fallback_used=True)
|
|
# An explicit English request can switch away from the prior locale
|
|
# even when the language-detection provider is unavailable.
|
|
if re.search(r"\b(?:in|use|speak|reply in|respond in) English\b", text, re.IGNORECASE):
|
|
return "en"
|
|
if re.search(r"[\u3040-\u30ff]", text):
|
|
return "ja"
|
|
if re.search(r"[\uac00-\ud7af]", text):
|
|
return "ko"
|
|
if re.search(r"[\u4e00-\u9fff]", text):
|
|
return self._prior_language if self._prior_language.startswith("zh") else "zh-Hans"
|
|
return self._prior_language or "en"
|
|
|
|
def detect_language(self, text: str) -> str:
|
|
if not text or not text.strip():
|
|
return "en"
|
|
try:
|
|
model = self._model_provider()
|
|
if model is None:
|
|
return self._fallback(text)
|
|
raw = llm.invoke_text(model, system=_DETECT_SYSTEM, user=text[:2000]).strip()
|
|
except Exception:
|
|
logger.exception("dify_builder: language detection failed; using conversation fallback")
|
|
return self._fallback(text)
|
|
# Reject anything that isn't a single short, code-shaped token: the model
|
|
# disobeying "reply with ONLY the code" (e.g. "The language is Japanese.")
|
|
# must use the conversation fallback instead of becoming a language tag.
|
|
if not raw or re.search(r"\s", raw) or len(raw) > 35:
|
|
return self._fallback(text)
|
|
code = raw.strip('."')
|
|
if not code or not _LANG_CODE_RE.match(code):
|
|
return self._fallback(text)
|
|
return code
|