From 28fcd5373be56c2c6a2383e2f0b6851173ea7091 Mon Sep 17 00:00:00 2001 From: stefanfeng Date: Tue, 8 Sep 2026 10:23:34 +0800 Subject: [PATCH] fix(avatar): follow language changes each turn --- digital-avatar-app/backend/routers/chat.py | 24 +++++++++++-- .../backend/tests/test_chat_orchestration.py | 36 +++++++++++++++++++ 2 files changed, 58 insertions(+), 2 deletions(-) diff --git a/digital-avatar-app/backend/routers/chat.py b/digital-avatar-app/backend/routers/chat.py index be5a7c1..fece73e 100644 --- a/digital-avatar-app/backend/routers/chat.py +++ b/digital-avatar-app/backend/routers/chat.py @@ -478,6 +478,24 @@ def _qa_requires_language_adaptation(question: str, answer: str) -> bool: ) +def _qa_requires_per_turn_rendering( + question: str, + answer: str, + history: list[Any], +) -> bool: + """Keep the direct QA fast path only when no conversation can bias language.""" + return bool(history) or _qa_requires_language_adaptation(question, answer) + + +def _per_turn_language_instruction() -> str: + 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。" + "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；" + "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" + "自然语言。不要说明你检测、切换或翻译了语言。" + ) + + def _canonicalize_question(value: str) -> str: value = _normalize_question(value) replacements = ( @@ -709,6 +727,8 @@ def _build_prompt( messages = [{"role": "system", "content": system}] 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": "user", "content": question.strip()}) return messages @@ -845,7 +865,7 @@ def _resolve_reply( qa_pairs = db.query(QAPair).filter(QAPair.avatar_id == avatar.id).all() matched = _match_standard_qa(question, qa_pairs) adapt_qa_language = bool( - matched and _qa_requires_language_adaptation(question, matched.answer) + matched and _qa_requires_per_turn_rendering(question, matched.answer, history) ) if matched and not adapt_qa_language and not image_contexts: return {"answer": matched.answer, "source": "qa", "references": []} @@ -940,7 +960,7 @@ def _stream_reply( qa_pairs = db.query(QAPair).filter(QAPair.avatar_id == avatar.id).all() matched = _match_standard_qa(question, qa_pairs) adapt_qa_language = bool( - matched and _qa_requires_language_adaptation(question, matched.answer) + matched and _qa_requires_per_turn_rendering(question, matched.answer, history) ) messages, reservation = [], None if matched and not adapt_qa_language and not image_contexts: diff --git a/digital-avatar-app/backend/tests/test_chat_orchestration.py b/digital-avatar-app/backend/tests/test_chat_orchestration.py index 835d2ed..a4a66d4 100644 --- a/digital-avatar-app/backend/tests/test_chat_orchestration.py +++ b/digital-avatar-app/backend/tests/test_chat_orchestration.py @@ -11,6 +11,7 @@ from routers.chat import ( _match_standard_qa, _public_avatar_payload, _qa_requires_language_adaptation, + _qa_requires_per_turn_rendering, _require_owned_avatar, _resolve_reply, ) @@ -86,6 +87,41 @@ class ChatOrchestrationTests(unittest.TestCase): self.assertTrue(_qa_requires_language_adaptation("안녕하세요", "你好")) self.assertFalse(_qa_requires_language_adaptation("你好", "您好")) + def test_conversation_qa_is_rendered_for_the_current_turn_language(self): + history = [SimpleNamespace(role="user", content="Please answer in English.")] + self.assertTrue(_qa_requires_per_turn_rendering("Quelle est votre adresse ?", "Our address is Test Road 1.", history)) + + fake_model = Mock(return_value="Notre adresse est Test Road 1.") + result = _resolve_reply( + None, + self.avatar, + "Quelle est votre adresse ?", + history, + qa_pairs=[SimpleNamespace(question="Quelle est votre adresse ?", answer="Our address is Test Road 1.", enabled=True)], + search_fn=Mock(), + model_client=fake_model, + ) + + self.assertEqual(result["source"], "qa") + self.assertEqual(result["answer"], "Notre adresse est Test Road 1.") + 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("不要沿用上一轮语言", messages[-2]["content"]) + + def test_latest_user_message_has_an_adjacent_language_override(self): + history = [ + SimpleNamespace(role="user", content="请用中文回答"), + SimpleNamespace(role="assistant", content="好的，请问有什么可以帮你？"), + ] + messages = _build_prompt(self.avatar, history, "What can you help me with?", []) + + 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("立即切换到相同语言", messages[-2]["content"]) + 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"): with self.subTest(question=question):