refactor: clean up controller logs + improve stock tool prompt

backend/agent/controller_helpers.py

@@ -109,39 +109,29 @@ async def finalize_user_insight_after_stream(
     stream_task,
     streaming_callback,
     identity_key: str,
-    trace_id: str,
+    trace_id: str = "",
 ) -> None:
     """
     Background task: Wait for stream to complete, then extract and save user_insight.
-    Links to main trace using trace_context for proper Langfuse tracing.
-
-    Args:
-        stream_task: The asyncio task for graph streaming
-        streaming_callback: ProductIDStreamingCallback instance
-        identity_key: User/device identifier
-        trace_id: Langfuse trace ID to link background task
+    NOTE: Runs outside Langfuse trace context — no need to log to Langfuse.
     """
+    # Clear Langfuse observation context to avoid auto-tracing in background
     try:
-        from langfuse import get_client
+        from langfuse._context import _observation_stack_context
+        _observation_stack_context.set([])
+    except Exception:
+        pass  # Langfuse version may not have this; safe to ignore
 
-        langfuse = get_client()
-
-        # Link background task to same trace using trace_context
-        with langfuse.start_as_current_observation(
-            name="extract-user-insight-background", trace_context={"trace_id": trace_id}
-        ) as span:
+    try:
         # Wait for stream to complete
         await stream_task
 
         # Extract user insight from accumulated content
         full_content = streaming_callback.accumulated_content
         if full_content and identity_key:
-                span.update(input={"identity_key": identity_key, "content_length": len(full_content)})
-                result = await extract_and_save_user_insight(full_content, identity_key)
-                span.update(output=result)
+            await extract_and_save_user_insight(full_content, identity_key)
         else:
             logger.warning("⚠️ [Background] No content to extract user_insight from")
-                span.update(output={"status": "no_content"})
 
     except Exception as exc:
         logger.error(f"❌ [Background] user_insight finalize failed: {exc}")

backend/agent/graph.py

@@ -21,7 +21,7 @@ from langgraph.types import CachePolicy
 from common.llm_factory import create_llm
 
 from .models import AgentConfig, AgentState, get_config
-from .prompt import get_last_modified, get_system_prompt, get_system_prompt_template
+from .prompt_utils import get_system_prompt_template
 from .tools.get_tools import get_all_tools, get_collection_tools
 
 logger = logging.getLogger(__name__)
@@ -48,11 +48,13 @@ class CANIFAGraph:
         self.retrieval_tools = self.all_tools
 
         self.llm_with_tools = self.llm.bind_tools(self.all_tools, strict=True)
-        # Lưu template với {date_str} placeholder → inject dynamic mỗi request
-        self.system_prompt_template = get_system_prompt_template()
-        self.prompt_template = ChatPromptTemplate.from_messages(
+        self.cache = InMemoryCache()
+
+    def _build_chain(self, system_prompt_template: str):
+        """Build chain with dynamic system prompt (fetched from Langfuse per request)."""
+        prompt_template = ChatPromptTemplate.from_messages(
             [
-                ("system", self.system_prompt_template),
+                ("system", system_prompt_template),
                 (
                     "system",
                     "===== USER INSIGHT (TỪ TURN TRƯỚC) =====\n⚡ BẮT BUỘC: Đọc [NEXT] bên dưới và THỰC HIỆN chiến lược đã lên kế hoạch!\n\n{user_insight}\n=====================================",
@@ -64,9 +66,7 @@ class CANIFAGraph:
                 MessagesPlaceholder(variable_name="messages"),
             ]
         )
-
-        self.chain = self.prompt_template | self.llm_with_tools
-        self.cache = InMemoryCache()
+        return prompt_template | self.llm_with_tools
 
     async def _agent_node(self, state: AgentState, config: RunnableConfig) -> dict:
         """Agent node - Chỉ việc đổ dữ liệu riêng vào khuôn đã có sẵn."""
@@ -104,10 +104,12 @@ class CANIFAGraph:
             or "⚠️ TRẠNG THÁI KHỞI TẠO: Chưa có User Insight từ lịch sử. Hãy bắt đầu thu thập thông tin mới (Nếu thiếu thông tin thì ghi 'Chưa rõ')."
         )
 
-        # Inject date_str ĐỘNG mỗi request (không cache từ __init__)
+        # Fetch prompt from Langfuse DYNAMICALLY per request (cache_ttl=0 → 30ms)
         current_date_str = datetime.now().strftime("%d/%m/%Y")
+        system_prompt_template = get_system_prompt_template()
+        chain = self._build_chain(system_prompt_template)
 
-        response = await self.chain.ainvoke(
+        response = await chain.ainvoke(
             {
                 "date_str": current_date_str,
                 "history": history,
@@ -181,34 +183,23 @@ def build_graph(config: AgentConfig | None = None, llm: BaseChatModel | None = N
 def get_graph_manager(
     config: AgentConfig | None = None, llm: BaseChatModel | None = None, tools: list | None = None
 ) -> CANIFAGraph:
-    """Get CANIFAGraph instance (Auto-rebuild if model config changes OR prompt version changed)."""
-
-    current_prompt_version = get_last_modified()
+    """Get CANIFAGraph instance (Auto-rebuild if model config changes).
     
+    Prompt is now fetched dynamically per request from Langfuse,
+    so no need to rebuild graph when prompt changes.
+    """
     # 1. New Instance if Empty
     if _instance[0] is None:
         _instance[0] = CANIFAGraph(config, llm, tools)
-        _instance[0].prompt_version = current_prompt_version
-        logger.info(f"✨ Graph Created: {_instance[0].config.model_name}, prompt_version={current_prompt_version}")
+        logger.info(f"✨ Graph Created: {_instance[0].config.model_name} (prompts from Langfuse)")
         return _instance[0]
 
-    # 2. Check for Config Changes (Model Switch OR Prompt Version Change)
+    # 2. Check for Model Config Changes only
     is_model_changed = config and config.model_name != _instance[0].config.model_name
-    is_prompt_changed = current_prompt_version != getattr(_instance[0], "prompt_version", 0)
 
-    if is_model_changed or is_prompt_changed:
-        change_reason = []
     if is_model_changed:
-            change_reason.append(f"Model ({_instance[0].config.model_name}->{config.model_name})")
-        if is_prompt_changed:
-            change_reason.append(
-                f"Prompt Version ({getattr(_instance[0], 'prompt_version', 0)}->{current_prompt_version})"
-            )
-
-        logger.info(f"🔄 Rebuilding Graph due to: {', '.join(change_reason)}")
-
+        logger.info(f"🔄 Rebuilding Graph: Model ({_instance[0].config.model_name}->{config.model_name})")
         _instance[0] = CANIFAGraph(config, llm, tools)
-        _instance[0].prompt_version = current_prompt_version
         return _instance[0]
 
     return _instance[0]

backend/agent/prompt.py

-
-"""
-CiCi Fashion Consultant - System Prompt
-Tư vấn thời trang CANIFA chuyên nghiệp
-Version 3.4 - Simplified Summary History
-Last updated: 2026-01-29 11:27
-"""
-
-import os
-from datetime import datetime
-
-PROMPT_FILE_PATH = os.path.join(os.path.dirname(__file__), "system_prompt.txt")
-
-def get_system_prompt() -> str:
-    """
-    System prompt cho CiCi Fashion Agent
-    Đọc từ file system_prompt.txt để có thể update dynamic.
-    
-    Returns:
-        str: System prompt với ngày hiện tại
-    """
-    now = datetime.now()
-    date_str = now.strftime("%d/%m/%Y")
-
-    try:
-        if os.path.exists(PROMPT_FILE_PATH):
-            with open(PROMPT_FILE_PATH, "r", encoding="utf-8") as f:
-                prompt_template = f.read()
-                return prompt_template.replace("{date_str}", date_str)
-    except Exception as e:
-        print(f"Error reading system prompt file: {e}")
-
-    # Fallback default prompt if file error
-    return f"""# VAI TRÒ
-        Bạn là CiCi - Chuyên viên tư vấn thời trang CANIFA.
-        Hôm nay: {date_str}
-
-        KHÔNG BAO GIỜ BỊA ĐẶT. TRẢ LỜI NGẮN GỌN.
-        """
-
-
-def get_system_prompt_template() -> str:
-    """
-    Trả về system prompt template CHƯA replace {date_str}.
-    Dùng cho ChatPromptTemplate để inject ngày động mỗi request.
-    
-    Returns:
-        str: System prompt template với placeholder {date_str}
-    """
-    try:
-        if os.path.exists(PROMPT_FILE_PATH):
-            with open(PROMPT_FILE_PATH, "r", encoding="utf-8") as f:
-                return f.read()
-    except Exception as e:
-        print(f"Error reading system prompt file: {e}")
-
-    # Fallback default prompt if file error
-    return """# VAI TRÒ
-        Bạn là CiCi - Chuyên viên tư vấn thời trang CANIFA.
-        Hôm nay: {date_str}
-
-        KHÔNG BAO GIỜ BỊA ĐẶT. TRẢ LỜI NGẮN GỌN.
-        """
-
-
-def get_last_modified() -> float:
-    """Trả về timestamp lần sửa cuối cùng của file system_prompt.txt."""
-    try:
-        if os.path.exists(PROMPT_FILE_PATH):
-            return os.path.getmtime(PROMPT_FILE_PATH)
-    except Exception:
-        pass
-    return 0.0
\ No newline at end of file

backend/agent/prompt_utils.py

+"""
+Prompt Utilities — ALL prompts from Langfuse.
+System prompt + Tool prompts, single source of truth.
+"""
 
-import os
 import logging
+from datetime import datetime
+
+from langfuse import Langfuse
 
 logger = logging.getLogger(__name__)
 
-# Directory name for tool prompts
-PROMPTS_DIR_NAME = "tool_prompts"
-PROMPTS_DIR = os.path.join(os.path.dirname(__file__), PROMPTS_DIR_NAME)
+LANGFUSE_SYSTEM_PROMPT_NAME = "canifa-stylist-system-prompt"
 
-def get_tool_prompt_path(filename: str) -> str:
-    """Get absolute path for a tool prompt file."""
-    if not filename.endswith(".txt"):
-        filename += ".txt"
-    return os.path.join(PROMPTS_DIR, filename)
+LANGFUSE_TOOL_PROMPT_MAP = {
+    "brand_knowledge_tool": "canifa-tool-brand-knowledge",
+    "check_is_stock": "canifa-tool-check-stock",
+    "data_retrieval_tool": "canifa-tool-data-retrieval",
+    "promotion_canifa_tool": "canifa-tool-promotion",
+}
+
+_langfuse_client: Langfuse | None = None
+
+def _get_langfuse() -> Langfuse:
+    global _langfuse_client
+    if _langfuse_client is None:
+        _langfuse_client = Langfuse()
+    return _langfuse_client
+
+
+def get_system_prompt() -> str:
+    """System prompt với ngày hiện tại đã inject."""
+    lf = _get_langfuse()
+    prompt = lf.get_prompt(LANGFUSE_SYSTEM_PROMPT_NAME, label="production", cache_ttl_seconds=0)
+    return prompt.compile(date_str=datetime.now().strftime("%d/%m/%Y"))
+
+
+def get_system_prompt_template() -> str:
+    """Template chưa replace date_str — dùng cho ChatPromptTemplate."""
+    lf = _get_langfuse()
+    prompt = lf.get_prompt(LANGFUSE_SYSTEM_PROMPT_NAME, label="production", cache_ttl_seconds=0)
+    return prompt.prompt.replace("{{date_str}}", "{date_str}")
 
-def read_tool_prompt(filename: str, default_prompt: str = "") -> str:
-    """
-    Read tool prompt from file. 
-    Returns default_prompt if file not found or empty.
-    """
-    file_path = get_tool_prompt_path(filename)
-    try:
-        if os.path.exists(file_path):
-            with open(file_path, "r", encoding="utf-8") as f:
-                content = f.read().strip()
-                if content:
-                    return content
-    except Exception as e:
-        logger.error(f"Error reading tool prompt {filename}: {e}")
 
+def read_tool_prompt(filename: str, default_prompt: str = "") -> str:
+    """Read tool prompt from Langfuse."""
+    name_key = filename.replace(".txt", "")
+    langfuse_name = LANGFUSE_TOOL_PROMPT_MAP.get(name_key)
+    if not langfuse_name:
+        logger.warning(f"⚠️ No Langfuse mapping for tool prompt '{name_key}'")
         return default_prompt
 
+    lf = _get_langfuse()
+    prompt = lf.get_prompt(langfuse_name, label="production", cache_ttl_seconds=0)
+    return prompt.prompt
+
+
 def write_tool_prompt(filename: str, content: str) -> bool:
-    """Write content to tool prompt file."""
-    file_path = get_tool_prompt_path(filename)
-    try:
-        os.makedirs(os.path.dirname(file_path), exist_ok=True)
-        with open(file_path, "w", encoding="utf-8") as f:
-            f.write(content)
-        return True
-    except Exception as e:
-        logger.error(f"Error writing tool prompt {filename}: {e}")
+    """Push tool prompt to Langfuse as new version."""
+    name_key = filename.replace(".txt", "")
+    langfuse_name = LANGFUSE_TOOL_PROMPT_MAP.get(name_key)
+    if not langfuse_name:
+        logger.error(f"No Langfuse mapping for '{name_key}'")
         return False
 
+    lf = _get_langfuse()
+    lf.create_prompt(
+        name=langfuse_name,
+        prompt=content,
+        labels=["production"],
+        tags=["canifa", "tool-prompt"],
+        type="text",
+    )
+    logger.info(f"✅ Tool prompt '{name_key}' pushed to Langfuse")
+    return True
+
+
 def list_tool_prompts() -> list[str]:
-    """List all available tool prompt files."""
-    try:
-        if not os.path.exists(PROMPTS_DIR):
-            return []
-        
-        files = [f for f in os.listdir(PROMPTS_DIR) if f.endswith(".txt")]
-        return sorted(files)
-    except Exception as e:
-        logger.error(f"Error listing tool prompts: {e}")
-        return []
+    """List available tool prompt names."""
+    return sorted(LANGFUSE_TOOL_PROMPT_MAP.keys())

backend/agent/tool_prompts/check_is_stock.txt

@@ -2,6 +2,7 @@ Công cụ KIỂM TRA TỒN KHO sản phẩm CANIFA theo mã sản phẩm.
 
 KHI NÀO GỌI TOOL NÀY:
 - Khách hỏi "còn hàng không?", "còn size không?", "check tồn kho"
+- Khách hỏi "có những size nào?", "còn những size nào?", "size chart", "số lượng từng size"
 - Khách hỏi về MÃ SẢN PHẨM CỤ THỂ kèm từ khóa tồn kho (vd: "8TS24W001 còn size L không?")
 - Khách muốn biết số lượng tồn kho của một hoặc nhiều mã sản phẩm
 
@@ -30,7 +31,6 @@ CASE 3: KIỂM TRA MÃ KÈM SIZE
 User: "Mã 6ST25W005-SE091 còn size M và L không?"
 -> Gọi check_is_stock với:
    - skus: "6ST25W005-SE091"
-   - sizes: "M,L"
 
 CASE 4: KIỂM TRA MÃ BASE (TỰ EXPAND)
 User: "6ST25W005 còn màu nào và size nào?"
@@ -38,8 +38,23 @@ User: "6ST25W005 còn màu nào và size nào?"
    - skus: "6ST25W005"
    (Tool sẽ tự động expand ra tất cả các biến thể từ DB)
 
-CÁCH ĐỌC KẾT QUẢ:
+CASE 5: HỎI CÓ NHỮNG SIZE NÀO / CÒN SIZE NÀO
+User: "Áo này có những size nào?" hoặc "Sản phẩm này còn size gì?"
+-> Gọi check_is_stock với mã sản phẩm (lấy từ context cuộc hội thoại)
+   - skus: "6ST25W005-SE091"
+
+CÁCH ĐỌC VÀ TRÌNH BÀY KẾT QUẢ:
 - stock_responses: Danh sách tồn kho từng SKU
 - is_in_stock: true/false - còn hàng hay không
 - qty: số lượng tồn kho
-- Nếu hết hàng -> Gợi ý size/màu khác còn hàng
+
+CÁCH TRÌNH BÀY CHO KHÁCH:
+1. Liệt kê RÕ RÀNG từng size kèm tình trạng:
+   - Size S: Còn hàng (12 sản phẩm)
+   - Size M: Còn hàng (8 sản phẩm)
+   - Size L: Hết hàng ❌
+   - Size XL: Còn hàng (3 sản phẩm)
+2. Nếu khách hỏi "có những size nào" → Liệt kê TẤT CẢ size có trong kết quả, kèm số lượng
+3. Nếu khách hỏi "còn size nào" → Chỉ liệt kê size CÒN HÀNG (qty > 0)
+4. Nếu hết hàng size khách muốn → Gợi ý size/màu khác còn hàng
+5. Nếu sản phẩm hoàn toàn hết hàng → Thông báo rõ và gợi ý sản phẩm tương tự

backend/agent/tools/data_retrieval_tool.py

 """
 CANIFA Data Retrieval Tool - Tối giản cho Agentic Workflow.
 Hỗ trợ Hybrid Search: Semantic (Vector) + Metadata Filter.
+
+Tracing:
+- LangChain CallbackHandler auto-traces tool call (name, input, output)
+- Manual span 'db-query-search' in _execute_single_search for SQL timing & product count
+  (nested under parent LangChain trace via start_as_current_observation)
 """
 
 import asyncio
@@ -8,7 +13,8 @@ import json
 import logging
 import time
 
-from langfuse import observe
+
+from langfuse import get_client as get_langfuse
 
 from langchain_core.tools import tool
 from pydantic import BaseModel, Field
@@ -103,12 +109,12 @@ class MultiSearchParams(BaseModel):
     searches: list[SearchItem] = Field(description="Danh sách các truy vấn tìm kiếm")
 
 
-@observe(name="db-query-search")
 async def _execute_single_search(
     db, item: SearchItem, query_vector: list[float] | None = None
 ) -> tuple[list[dict], dict]:
     """
     Thực thi một search query đơn lẻ (Async).
+    Creates a manual 'db-query-search' span under the current Langfuse context.
 
     Returns:
         Tuple of (products, filter_info)
@@ -130,14 +136,15 @@ async def _execute_single_search(
         logger.debug("SQL built, length=%s, build_time_ms=%.2f", len(sql), query_build_time)
 
         if not sql:
-            return []
+            return [], {"fallback_used": False}
 
         # Timer: execute DB query
         db_start = time.time()
         products = await db.execute_query_async(sql, params=params)
         db_time = (time.time() - db_start) * 1000  # Convert to ms
         logger.info(
-            "_execute_single_search done, products=%s, build_ms=%.2f, db_ms=%.2f, total_ms=%.2f",
+            "🔍 _execute_single_search done: query=%r, products=%s, build_ms=%.2f, db_ms=%.2f, total_ms=%.2f",
+            short_query,
             len(products),
             query_build_time,
             db_time,
@@ -159,26 +166,44 @@ async def _execute_single_search(
                 (first_p.get("description_text") or "")[:100],
             )
 
-        from langfuse import get_client as get_langfuse
+        # ✅ Nest 'db-query-search' span dưới chat-request trace hiện tại trên Langfuse
         try:
             lf = get_langfuse()
-            lf.update_current_observation(
+            if lf:
+                with lf.start_as_current_observation(
+                    as_type="span",
+                    name="db-query-search",
+                ) as span:
+                    span.update(
                         input={
                             "description": item.description[:200] if item.description else None,
                             "product_name": item.product_name,
                             "gender": item.gender_by_product,
                             "age": item.age_by_product,
+                            "color": item.master_color,
+                            "price_range": f"{item.price_min or ''}-{item.price_max or ''}",
                             "product_line_vn": item.product_line_vn,
                             "magento_ref_code": item.magento_ref_code,
+                            "discovery_mode": item.discovery_mode,
                         },
                         output={
-                    "raw_count": len(products),
+                            "product_count": len(products),
                             "build_ms": round(query_build_time, 2),
                             "db_ms": round(db_time, 2),
+                            "total_ms": round(query_build_time + db_time, 2),
+                            "sample_products": [
+                                {
+                                    "name": p.get("product_name", "?"),
+                                    "code": p.get("magento_ref_code", "?"),
+                                    "price": float(p["sale_price"]) if p.get("sale_price") else None,
+                                }
+                                for p in products[:3]
+                            ] if products else [],
                         },
+                        metadata={"sql_length": len(sql)},
                     )
-        except Exception:
-            pass  # Don't break if Langfuse not available
+        except Exception as trace_err:
+            logger.warning("⚠️ Langfuse db-query-search span failed: %s", trace_err, exc_info=True)
 
         return format_product_results(products), {"fallback_used": False}
     except Exception as e:
@@ -192,8 +217,26 @@ async def data_retrieval_tool(searches: list[SearchItem]) -> str:
     Công cụ tìm kiếm sản phẩm CANIFA.
     Hỗ trợ tìm kiếm Semantic và lọc theo Metadata.
     """
+    tool_start = time.time()
     logger.info("🔧 data_retrieval_tool called with %d items", len(searches))
 
+    # Log search params cho debugging
+    for i, s in enumerate(searches):
+        logger.info(
+            "🔧 Search[%d]: desc=%r, name=%r, gender=%s, age=%s, color=%s, "
+            "price=%s-%s, line=%s, code=%s, mode=%s",
+            i,
+            (s.description[:80] + "...") if s.description and len(s.description) > 80 else s.description,
+            s.product_name,
+            s.gender_by_product,
+            s.age_by_product,
+            s.master_color,
+            s.price_min, s.price_max,
+            s.product_line_vn,
+            s.magento_ref_code,
+            s.discovery_mode,
+        )
+
     # Get DB Connection
     db = get_db_connection()
     if not db:
@@ -213,8 +256,6 @@ async def data_retrieval_tool(searches: list[SearchItem]) -> str:
         if filter_info:
             all_filter_infos.append(filter_info)
 
-
-
     # Aggregate filter info from first result for simplicity in response
     final_info = all_filter_infos[0] if all_filter_infos else {}
 
@@ -237,16 +278,18 @@ async def data_retrieval_tool(searches: list[SearchItem]) -> str:
         "filter_info": final_info,
     }
 
+    total_ms = (time.time() - tool_start) * 1000
+
     logger.info(
-        "🎁 Final result: %d products. Fallback used: %s",
+        "🎁 Final result: %d products, total_ms=%.2f. Fallback used: %s",
         len(combined_results),
+        total_ms,
         final_info.get("fallback_used", False),
     )
 
     return json.dumps(output, ensure_ascii=False, default=str)
 
 
-# Load dynamic docstring
 # Load dynamic docstring
 dynamic_prompt = read_tool_prompt("data_retrieval_tool")
 if dynamic_prompt:

backend/agent/tools/product_search_helpers.py

@@ -2,7 +2,7 @@ import logging
 import os
 import time
 
-from langfuse import observe
+# Note: tracing handled by parent data_retrieval_tool via context manager
 
 from common.embedding_service import create_embedding_async
 
@@ -103,7 +103,6 @@ def _get_metadata_clauses(params, sql_params: list) -> list[str]:
     return clauses
 
 
-@observe(name="build-starrocks-query")
 async def build_starrocks_query(params, query_vector: list[float] | None = None) -> tuple[str, list]:
     """
     Build SQL query với Parameterized Query để tránh SQL Injection.

backend/api/prompt_route.py

 from fastapi import APIRouter, HTTPException, Request
 from pydantic import BaseModel
-import os
-import re
-from agent.graph import reset_graph
-from common.cache import bump_prompt_version
 
-router = APIRouter()
+from agent.prompt_utils import get_system_prompt_template, LANGFUSE_SYSTEM_PROMPT_NAME, _get_langfuse
 
-PROMPT_FILE_PATH = os.path.join(os.path.dirname(__file__), "../agent/system_prompt.txt")
+router = APIRouter()
 
-# Allowed variables in prompt (single braces OK for these)
-ALLOWED_VARIABLES = {"date_str"}
 
 class PromptUpdateRequest(BaseModel):
     content: str
 
-def validate_prompt_braces(content: str) -> tuple[bool, list[str]]:
-    """
-    Validate that all braces in prompt are properly escaped.
-    Returns (is_valid, list of problematic patterns)
-    """
-    # Find all {word} patterns
-    single_brace_pattern = re.findall(r'\{([^{}]+)\}', content)
-    
-    # Filter out allowed variables
-    problematic = [
-        var for var in single_brace_pattern 
-        if var.strip() not in ALLOWED_VARIABLES and not var.startswith('{')
-    ]
-    
-    return len(problematic) == 0, problematic
 
 from common.rate_limit import rate_limit_service
 
+
 @router.get("/api/agent/system-prompt")
 async def get_system_prompt_content(request: Request):
-    """Get current system prompt content"""
+    """Get current system prompt content from Langfuse."""
     try:
-        if os.path.exists(PROMPT_FILE_PATH):
-            with open(PROMPT_FILE_PATH, "r", encoding="utf-8") as f:
-                content = f.read()
-            return {"status": "success", "content": content}
-        else:
-            return {"status": "error", "message": "Prompt file not found"}
+        return {"status": "success", "content": get_system_prompt_template()}
     except Exception as e:
         raise HTTPException(status_code=500, detail=str(e))
 
@@ -50,41 +25,24 @@ async def get_system_prompt_content(request: Request):
 @router.post("/api/agent/system-prompt")
 @rate_limit_service.limiter.limit("10/minute")
 async def update_system_prompt_content(request: Request, body: PromptUpdateRequest):
-    """Update system prompt content"""
+    """Create new system prompt version in Langfuse."""
     try:
-        # Validate braces
-        is_valid, problematic = validate_prompt_braces(body.content)
-        
-        if not is_valid:
-            # Return warning but still allow save
-            warning = (
-                f"⚠️ Phát hiện {{...}} chưa escape: {problematic[:3]}... "
-                f"Nếu đây là JSON, hãy dùng {{{{ }}}} thay vì {{ }}. "
-                f"Prompt vẫn được lưu nhưng có thể gây lỗi khi chat."
+        lf = _get_langfuse()
+        content = body.content
+        if "{date_str}" in content and "{{date_str}}" not in content:
+            content = content.replace("{date_str}", "{{date_str}}")
+
+        prompt = lf.create_prompt(
+            name=LANGFUSE_SYSTEM_PROMPT_NAME,
+            prompt=content,
+            labels=["production"],
+            tags=["canifa", "system-prompt"],
+            type="text",
         )
-        else:
-            warning = None
-            
-        # 1. Update file
-        with open(PROMPT_FILE_PATH, "w", encoding="utf-8") as f:
-            f.write(body.content)
-            
-        # 2. Bump prompt version in Redis (ALL workers will detect this)
-        new_version = await bump_prompt_version()
-        
-        # 3. Reset local worker's Graph Singleton (immediate effect for this worker)
-        reset_graph()
-        
-        response = {
+        return {
             "status": "success",
-            "message": f"System prompt updated. Version: {new_version}. All workers will reload on next request.",
-            "prompt_version": new_version
+            "message": f"System prompt updated. Version: {prompt.version}",
+            "prompt_version": prompt.version,
         }
-        if warning:
-            response["warning"] = warning
-            
-        return response
     except Exception as e:
         raise HTTPException(status_code=500, detail=str(e))
-
-

backend/api/tool_prompt_route.py

@@ -23,7 +23,7 @@ async def get_tool_prompts_list(request: Request):
 
 @router.get("/api/agent/tool-prompts/{filename}")
 async def get_tool_prompt_content(filename: str, request: Request):
-    """Get content of a specific tool prompt file."""
+    """Get content of a specific tool prompt (from Langfuse, fallback to file)."""
     try:
         content = read_tool_prompt(filename)
         if not content:
@@ -38,7 +38,7 @@ async def get_tool_prompt_content(filename: str, request: Request):
 
 @router.post("/api/agent/tool-prompts/{filename}")
 async def update_tool_prompt_content(filename: str, request: Request, body: ToolPromptUpdateRequest):
-    """Update content of a tool prompt file and reset graph."""
+    """Update tool prompt (writes to file + syncs to Langfuse) and reset graph."""
     try:
         # Ensure filename is safe (basic check)
         if ".." in filename or "/" in filename or "\\" in filename:
@@ -51,7 +51,7 @@ async def update_tool_prompt_content(filename: str, request: Request, body: Tool
         # Reset Graph to reload tools with new prompts
         reset_graph()
 
-        return {"status": "success", "message": f"Tool prompt {filename} updated successfully. Graph reloaded."}
+        return {"status": "success", "message": f"Tool prompt {filename} updated (file + Langfuse). Graph reloaded."}
     except HTTPException:
         raise
     except Exception as e:

backend/tests/test_model.py

+"""Quick test: Gemini 2.0 Flash via Google GenAI."""
+
+from google import genai
+
+API_KEY = "AIzaSyCLAk9LsvS7RtxuXJLf5uqx4sI034Hc3Xg"
+
+client = genai.Client(api_key=API_KEY)
+
+response = client.models.generate_content(
+    model="gemini-3-flash-preview",
+    contents="Xin chào! Bạn là ai? Trả lời ngắn gọn bằng tiếng Việt.",
+)
+
+print("=" * 60)
+print("Model: gemini--flash")
+print("=" * 60)
+print(response.text)
+print("=" * 60)