Update api/clare_core.py

api/clare_core.py

@@ -15,8 +15,24 @@ from .config import (
     LEARNING_MODE_INSTRUCTIONS,
 )
 
-from langsmith import traceable
-from langsmith.run_helpers import set_run_metadata
+# ----------------------------
+# Tracing toggle (LangSmith)
+# ----------------------------
+# Default OFF for speed + stability in HF cold start environments.
+ENABLE_TRACING = os.getenv("CLARE_ENABLE_TRACING", "0").strip() == "1"
+
+if ENABLE_TRACING:
+    from langsmith import traceable  # type: ignore
+    from langsmith.run_helpers import set_run_metadata  # type: ignore
+else:
+    # no-op decorators / funcs
+    def traceable(*args, **kwargs):  # type: ignore
+        def _decorator(fn):
+            return fn
+        return _decorator
+
+    def set_run_metadata(**kwargs):  # type: ignore
+        return None
 
 
 # ----------------------------
@@ -26,6 +42,9 @@ MAX_HISTORY_TURNS = int(os.getenv("CLARE_MAX_HISTORY_TURNS", "10"))
 MAX_RAG_CHARS_IN_PROMPT = int(os.getenv("CLARE_MAX_RAG_CHARS", "2000"))
 DEFAULT_MAX_OUTPUT_TOKENS = int(os.getenv("CLARE_MAX_OUTPUT_TOKENS", "384"))
 
+# Similarity knobs
+ENABLE_EMBEDDING_SIM = os.getenv("CLARE_ENABLE_EMBEDDING_SIMILARITY", "0").strip() == "1"
+
 
 # ---------- syllabus 解析 ----------
 def parse_syllabus_docx(file_path: str, max_lines: int = 15) -> List[str]:
@@ -251,6 +270,12 @@ def find_similar_past_question(
     embedding_threshold: float = 0.85,
     max_turns_to_check: int = 6,
 ) -> Optional[Tuple[str, str, float]]:
+    """
+    Fast path:
+    - Always do Jaccard on normalized text for up to max_turns_to_check.
+    Optional path (disabled by default for speed/stability):
+    - Embedding-based similarity if ENABLE_EMBEDDING_SIM=1
+    """
     norm_msg = _normalize_text(message)
     if not norm_msg:
         return None
@@ -279,6 +304,10 @@ def find_similar_past_question(
     if best_pair_j and best_sim_j >= jaccard_threshold:
         return best_pair_j[0], best_pair_j[1], best_sim_j
 
+    # Optional: embedding similarity (OFF by default)
+    if not ENABLE_EMBEDDING_SIM:
+        return None
+
     if not history:
         return None
 
@@ -462,6 +491,7 @@ def chat_with_clare(
     cognitive_state: Optional[Dict[str, int]],
     rag_context: Optional[str] = None,
 ) -> Tuple[str, List[Tuple[str, str]]]:
+    # avoid any tracing overhead when disabled (set_run_metadata is no-op in that case)
     try:
         set_run_metadata(
             learning_mode=learning_mode,
@@ -469,6 +499,7 @@ def chat_with_clare(
             doc_type=doc_type,
         )
     except Exception as e:
+        # safe even if tracing enabled but misconfigured
         print(f"[LangSmith metadata error in chat_with_clare] {repr(e)}")
 
     messages = build_messages(