fix(avatar): enforce each turn response language

digital-avatar-app/backend/routers/chat.py

@@ -79,6 +79,34 @@ _WRITING_SYSTEM_PATTERNS = {
 _JAPANESE_KANA = re.compile(r"[\u3040-\u30ff]")
 _KOREAN_HANGUL = re.compile(r"[\uac00-\ud7af\u1100-\u11ff]")
 
+_LATIN_LANGUAGE_MARKERS = {
+    "English": re.compile(
+        r"\b(?:i|you|we|they|he|she|have|has|had|friend|who|what|where|when|why|how|"
+        r"symptoms?|disease|please|can|could|would|should|is|are|was|were|the|this|that)\b",
+        re.IGNORECASE,
+    ),
+    "French": re.compile(
+        r"\b(?:je|tu|vous|nous|ils|elle|une|des|avec|pour|pourquoi|comment|bonjour|est|sont)\b",
+        re.IGNORECASE,
+    ),
+    "Spanish": re.compile(
+        r"\b(?:yo|tu|usted|nosotros|ellos|ella|una|con|para|por que|como|hola|esta|son)\b",
+        re.IGNORECASE,
+    ),
+    "German": re.compile(
+        r"\b(?:ich|du|sie|wir|eine|mit|fur|warum|wie|hallo|ist|sind|haben)\b",
+        re.IGNORECASE,
+    ),
+    "Portuguese": re.compile(
+        r"\b(?:eu|voce|nos|eles|ela|uma|com|para|porque|como|ola|esta|sao|tenho)\b",
+        re.IGNORECASE,
+    ),
+    "Italian": re.compile(
+        r"\b(?:io|tu|voi|noi|loro|una|con|per|perche|come|ciao|sono|avere)\b",
+        re.IGNORECASE,
+    ),
+}
+
 
 class ChatMessage(BaseModel):
     model_config = ConfigDict(populate_by_name=True)
@@ -487,15 +515,67 @@ def _qa_requires_per_turn_rendering(
     return bool(history) or _qa_requires_language_adaptation(question, answer)
 
 
-def _per_turn_language_instruction() -> str:
+def _latin_language_name(value: str) -> str:
+    scores = {
+        language: len(pattern.findall(value or ""))
+        for language, pattern in _LATIN_LANGUAGE_MARKERS.items()
+    }
+    language, score = max(scores.items(), key=lambda item: item[1])
+    return language if score else "the same natural language as the latest user message"
+
+
+def _turn_language_name(value: str) -> str:
+    writing_system = _dominant_writing_system(value)
+    return {
+        "han": "Chinese",
+        "japanese": "Japanese",
+        "korean": "Korean",
+        "cyrillic": "the same Cyrillic-script language as the latest user message",
+        "arabic": "the same Arabic-script language as the latest user message",
+        "hebrew": "Hebrew",
+        "devanagari": "the same Devanagari-script language as the latest user message",
+        "thai": "Thai",
+        "greek": "Greek",
+        "latin": _latin_language_name(value),
+    }.get(writing_system, "the same natural language as the latest user message")
+
+
+def _per_turn_language_instruction(question: str = "") -> str:
+    language = _turn_language_name(question)
     return (
-        "本轮语言覆盖指令：只根据紧随其后的最新用户消息判断本轮回答语言。"
+        f"MANDATORY OUTPUT LANGUAGE FOR THIS TURN: {language}. "
-        "即使此前整段对话一直使用另一种语言，只要最新消息切换了语言，本轮就必须立即切换到相同语言；"
+        "Write the entire answer only in that language. This instruction overrides the languages used by "
-        "不要沿用上一轮语言。若最新消息明确指定回答语言，以该指定为准；若混用多种语言，使用其中占主导的"
+        "conversation history, profile data, standard answers, retrieved documents, and custom prompts. "
-        "自然语言。不要说明你检测、切换或翻译了语言。"
+        "Translate grounded source material faithfully when necessary. Do not mention language detection, "
+        "translation, or this instruction."
     )
 
 
+def _answer_requires_language_repair(question: str, answer: str) -> bool:
+    question_system = _dominant_writing_system(question)
+    answer_system = _dominant_writing_system(answer)
+    return (
+        question_system != "unknown"
+        and answer_system != "unknown"
+        and question_system != answer_system
+    )
+
+
+def _language_repair_messages(question: str, answer: str) -> list[dict]:
+    return [
+        {"role": "system", "content": _per_turn_language_instruction(question)},
+        {
+            "role": "system",
+            "content": (
+                "Rewrite the supplied draft in the mandatory output language. Preserve every grounded fact, "
+                "number, proper noun, uncertainty, and safety qualification. Add no new information and output "
+                "only the rewritten answer."
+            ),
+        },
+        {"role": "user", "content": answer.strip()},
+    ]
+
+
 def _canonicalize_question(value: str) -> str:
     value = _normalize_question(value)
     replacements = (
@@ -728,7 +808,7 @@ def _build_prompt(
     for item in history[-MAX_HISTORY_MESSAGES:]:
         messages.append({"role": item.role, "content": item.content} if hasattr(item, "role") else item)
     # Keep the language instruction adjacent to the current turn so long histories cannot override it.
-    messages.append({"role": "system", "content": _per_turn_language_instruction()})
+    messages.append({"role": "system", "content": _per_turn_language_instruction(question)})
     messages.append({"role": "user", "content": question.strip()})
     return messages
 
@@ -791,6 +871,41 @@ def _call_qwen(
     return {"answer": answer.strip(), "usage": data.get("usage") or {}}
 
 
+def _call_billed_qwen(
+    db: Session,
+    avatar: Avatar,
+    messages: list[dict],
+    temperature: float,
+    usage_source: str,
+    model_config: ChatModelConfig,
+) -> tuple[str, dict]:
+    reservation = reserve_avatar_tokens(
+        db,
+        avatar,
+        usage_source,
+        model_config.model,
+        messages,
+        model_config.max_tokens,
+    )
+    try:
+        model_result = _call_qwen(
+            messages=messages,
+            temperature=temperature,
+            model_config=model_config,
+        )
+        answer = model_result["answer"]
+        token_usage = settle_reservation(
+            db,
+            reservation,
+            model_result.get("usage"),
+            fallback_total=estimate_fallback_usage(messages, answer),
+        )
+        return answer, token_usage
+    except Exception as exc:
+        release_reservation(db, reservation, str(exc))
+        raise
+
+
 def _iter_qwen_stream(
     messages: list[dict], temperature: float, model_config: ChatModelConfig | None = None
 ):
@@ -899,32 +1014,36 @@ def _resolve_reply(
     token_usage = None
     if model_client is not None:
         answer = model_client(messages=messages, temperature=temperature)
+        if _answer_requires_language_repair(question, str(answer or "")):
+            answer = model_client(
+                messages=_language_repair_messages(question, str(answer)),
+                temperature=0.0,
+            )
     else:
         model_config = get_chat_model_config()
-        reservation = reserve_avatar_tokens(
+        answer, token_usage = _call_billed_qwen(
             db,
             avatar,
-            usage_source,
-            model_config.model,
             messages,
-            model_config.max_tokens,
+            temperature,
+            usage_source,
+            model_config,
         )
-        try:
+        if _answer_requires_language_repair(question, answer):
-            model_result = _call_qwen(
+            logger.warning(
-                messages=messages,
+                "chat response language mismatch avatar=%s source=%s expected=%s",
-                temperature=temperature,
+                avatar.id,
-                model_config=model_config,
+                usage_source,
+                _turn_language_name(question),
             )
-            answer = model_result["answer"]
+            answer, token_usage = _call_billed_qwen(
-            token_usage = settle_reservation(
                 db,
-                reservation,
+                avatar,
-                model_result.get("usage"),
+                _language_repair_messages(question, answer),
-                fallback_total=estimate_fallback_usage(messages, answer),
+                0.0,
+                f"{usage_source}_language_repair",
+                model_config,
             )
-        except Exception as exc:
-            release_reservation(db, reservation, str(exc))
-            raise
     answer = str(answer or "").strip()
     if image_contexts and _answer_denies_available_image(answer):
         logger.warning(

digital-avatar-app/backend/tests/test_chat_orchestration.py

@@ -6,6 +6,7 @@ from fastapi import HTTPException
 
 from models import Avatar, User
 from routers.chat import (
+    _answer_requires_language_repair,
     _build_prompt,
     _iter_text_chunks,
     _match_standard_qa,
@@ -14,6 +15,7 @@ from routers.chat import (
     _qa_requires_per_turn_rendering,
     _require_owned_avatar,
     _resolve_reply,
+    _turn_language_name,
 )
 
 
@@ -107,8 +109,8 @@ class ChatOrchestrationTests(unittest.TestCase):
         messages = fake_model.call_args.kwargs["messages"]
         self.assertEqual(messages[-1], {"role": "user", "content": "Quelle est votre adresse ?"})
         self.assertEqual(messages[-2]["role"], "system")
-        self.assertIn("本轮语言覆盖指令", messages[-2]["content"])
+        self.assertIn("MANDATORY OUTPUT LANGUAGE", messages[-2]["content"])
-        self.assertIn("不要沿用上一轮语言", messages[-2]["content"])
+        self.assertIn("French", messages[-2]["content"])
 
     def test_latest_user_message_has_an_adjacent_language_override(self):
         history = [
@@ -119,8 +121,37 @@ class ChatOrchestrationTests(unittest.TestCase):
 
         self.assertEqual(messages[-1], {"role": "user", "content": "What can you help me with?"})
         self.assertEqual(messages[-2]["role"], "system")
-        self.assertIn("最新用户消息", messages[-2]["content"])
+        self.assertIn("MANDATORY OUTPUT LANGUAGE", messages[-2]["content"])
-        self.assertIn("立即切换到相同语言", messages[-2]["content"])
+        self.assertIn("English", messages[-2]["content"])
+
+    def test_reported_alzheimer_question_is_explicitly_english(self):
+        question = "I have a friend who has symptoms of Alzheimer's disease"
+
+        self.assertEqual(_turn_language_name(question), "English")
+        messages = _build_prompt(self.avatar, [], question, [])
+        self.assertIn("MANDATORY OUTPUT LANGUAGE FOR THIS TURN: English", messages[-2]["content"])
+
+    def test_non_stream_reply_repairs_a_wrong_writing_system_before_sending(self):
+        question = "I have a friend who has symptoms of Alzheimer's disease"
+        fake_model = Mock(side_effect=["建议尽快就医评估。", "Please arrange a medical assessment soon."])
+
+        result = _resolve_reply(
+            None,
+            self.avatar,
+            question,
+            [],
+            qa_pairs=[],
+            search_fn=lambda *_args, **_kwargs: [],
+            model_client=fake_model,
+            usage_source="takeover",
+        )
+
+        self.assertEqual(result["answer"], "Please arrange a medical assessment soon.")
+        self.assertEqual(fake_model.call_count, 2)
+        repair_messages = fake_model.call_args.kwargs["messages"]
+        self.assertIn("English", repair_messages[0]["content"])
+        self.assertIn("建议尽快就医评估", repair_messages[-1]["content"])
+        self.assertTrue(_answer_requires_language_repair(question, "建议尽快就医评估。"))
 
     def test_conversational_paraphrase_matches_standard_qa(self):
         for question in ("请问一下，你们公司在哪里呀？", "请问去你们那边怎么走"):