feat: language-aware routing (English -> llama3:8b-instruct; else Qwen 7B/14B). Docs updated

server/config.py

@@ -9,6 +9,7 @@ class Settings:
     ollama_host: str = os.getenv("OLLAMA_HOST", "http://localhost:11434")
     base_model: str = os.getenv("BASE_MODEL", "qwen2.5:7b-instruct")
     boost_model: str = os.getenv("BOOST_MODEL", "qwen2.5:14b-instruct")
+    english_model: str = os.getenv("ENGLISH_MODEL", "llama3:8b-instruct")
     embedding_model: str = os.getenv("EMBEDDING_MODEL", "nomic-embed-text")
     index_path: str = os.getenv("INDEX_PATH", "data/index.jsonl")
 

server/main.py

@@ -84,9 +84,16 @@ def search(req: SearchRequest) -> Dict[str, Any]:
 def chat(req: ChatRequest) -> Dict[str, Any]:
     model = req.model
     if not model:
-        # 라우팅: 메시지 길이/force_boost 기준 간단 분기
-        total_chars = sum(len(m.get("content", "")) for m in req.messages)
-        model = settings.boost_model if (req.force_boost or total_chars > 2000) else settings.base_model
+        # 언어 감지(매우 단순): 영문 비율이 높으면 영어 모델, 아니면 기본/부스팅
+        user_text = "\n".join(m.get("content", "") for m in req.messages if m.get("role") == "user")
+        ascii_letters = sum(ch.isascii() and ch.isalpha() for ch in user_text)
+        non_ascii_letters = sum((not ch.isascii()) and ch.isalpha() for ch in user_text)
+        english_ratio = ascii_letters / max(ascii_letters + non_ascii_letters, 1)
+        total_chars = len(user_text)
+        if english_ratio > 0.8:
+            model = settings.english_model
+        else:
+            model = settings.boost_model if (req.force_boost or total_chars > 2000) else settings.base_model
 
     context_docs: List[str] = []
     if req.use_rag and index.rows: