feat: implement core agent architecture including LLM integration, Redis-backed memory, tool definitions, and comprehensive test suite.

agent.py

@@ -14,6 +14,7 @@ if project_root not in sys.path:
 import logging
 from typing import Any
 from langchain_core.messages import HumanMessage, AIMessage, ToolMessage, SystemMessage
+from datetime import datetime
 from src.llm import llm
 from src.config import DEFAULT_MODEL, LOG_LEVEL
 from src.tools import get_tool_schemas, execute_tool
@@ -21,9 +22,26 @@ from src.tools import get_tool_schemas, execute_tool
 logging.basicConfig(level=LOG_LEVEL, format="%(asctime)s [%(levelname)s] %(message)s")
 logger = logging.getLogger(__name__)
 
-SYSTEM_PROMPT = """You are an intelligent AI assistant.
-You can use the provided tools to complete tasks.
-Think step by step and use tools when necessary."""
+SYSTEM_PROMPT = """You are an intelligent AI Assistant designed to be a "Second Brain" for the user.
+Your primary goal is to help the user manage their life while respecting UNIQUE constraints.
+
+CALENDAR-FIRST PRIORITY RULES:
+1.  **Event > Habit**: A specific calendar event (e.g., "Family Anniversary", "OOO", "Anniversary", "Off-grid") on a specific date ALWAYS overrules a general habit (e.g., "Monday Deep Work", "17h Swimming").
+2.  **No Exceptions for OOO**: If the user is OOO or "Off-grid" on a day, you MUST NOT suggest any work or meetings for that day. A "Sếp đòi phương án sáng mai" request must be resolved by proposing action **BEFORE** the OOO starts (e.g., tonight) or **DELEGATING** completely to an internal staff (e.g., Anh Hoàng).
+3.  **Validate Every Date**: Always check the specific date/day for a request (e.g., "Sếp đòi sáng mai" when today is Wednesday means Thursday). Cross-reference this date with your schedule results.
+
+SEARCH & REASONING RULES:
+1.  **BROAD SEARCH**: Call `get_memories(limit=100)` with NO query for audits.
+2.  **SINGLE-WORD KEYWORDS**: Only use single-word keywords for search (e.g., "azure", "ghét").
+3.  **RANGE SEARCH**: Call `get_schedule(date_str="next 2 weeks")`.
+4.  **Parallel Context**: Call all 3 context tools TOGETHER in your first turn.
+5.  **Strict Taboos**: Strictly reject any tech/vendor the user has an aversion to (Azure) and any work practices they hate (Outsourcing).
+
+THINK STEP-BY-STEP:
+1. Call search tools in parallel.
+2. Cross-reference all chat proposals against specific calendar events (First Priority).
+3. Apply habits and taboos (Second Priority).
+4. Synthesize the final plan."""
 
 
 def create_agent():
@@ -35,8 +53,12 @@ def run_agent_loop(client: Any, user_input: str, max_turns: int = 10) -> str:
     """
     Run the agent loop: send message -> receive response -> call tool -> repeat.
     """
+    # Dynamic Date Injection
+    today = datetime.now()
+    time_context = f"\n[CURRENT TIME CONTEXT]\nToday is {today.strftime('%A, %B %d, %Y')}.\n"
+    
     messages = [
-        SystemMessage(content=SYSTEM_PROMPT),
+        SystemMessage(content=SYSTEM_PROMPT + time_context),
         HumanMessage(content=user_input)
     ]
 

config.py

@@ -18,6 +18,11 @@ QWEN_API_KEY = os.getenv("QWEN_API_KEY", "")
 QWEN_BASE_URL = os.getenv("QWEN_BASE_URL", "https://dashscope.aliyuncs.com/compatible-mode/v1")
 QWEN_MODEL = os.getenv("QWEN_MODEL", "qwen-plus")
 
+# OpenRouter
+OPENROUTER_API_KEY = os.getenv("OPENROUTER_API_KEY", "")
+OPENROUTER_BASE_URL = os.getenv("OPENROUTER_BASE_URL", "https://openrouter.ai/api/v1")
+OPENROUTER_MODEL = os.getenv("OPENROUTER_MODEL", "google/gemma-4-26b-a4b-it")
+
 # Local LLM config
 USE_LOCAL_LLM = os.getenv("USE_LOCAL_LLM", "false").lower() == "true"
 LOCAL_MODEL_ID = os.getenv("LOCAL_MODEL_ID", "Qwen/Qwen2.5-0.5B-Instruct")

llm.py

@@ -1,17 +1,35 @@
 from langchain_google_genai import ChatGoogleGenerativeAI
-from src.config import GEMINI_API_KEY, DEFAULT_MODEL
+from langchain_openai import ChatOpenAI
+from src.config import GEMINI_API_KEY, DEFAULT_MODEL, OPENROUTER_API_KEY, OPENROUTER_BASE_URL, OPENROUTER_MODEL
 
-llm = ChatGoogleGenerativeAI(
-    model=DEFAULT_MODEL,
-    temperature=0,
-    top_p=1,
-    top_k=1,
-    max_tokens=None,
-    timeout=None,
-    max_retries=2,
-    google_api_key=GEMINI_API_KEY
-)
+def get_agent_llm():
+    """Returns the LLM instance based on availability: OpenRouter > Gemini."""
+    if OPENROUTER_API_KEY and not OPENROUTER_API_KEY.startswith("your-"):
+        return ChatOpenAI(
+            model=OPENROUTER_MODEL,
+            api_key=OPENROUTER_API_KEY,
+            base_url=OPENROUTER_BASE_URL,
+            temperature=0,
+            max_tokens=4096,
+            default_headers={
+                "HTTP-Referer": "https://github.com/a20-ai-thuc-chien", # Optional for OpenRouter
+                "X-Title": "A20 AI Assistant",
+            }
+        )
+    
+    # Fallback to Gemini
+    return ChatGoogleGenerativeAI(
+        model=DEFAULT_MODEL,
+        temperature=0,
+        google_api_key=GEMINI_API_KEY
+    )
+
+llm = get_agent_llm()
 
 if __name__ == "__main__":
-    response = llm.invoke("Hello World là gì?").content
-    print(response)
+    # Test
+    try:
+        response = llm.invoke("Hello, who are you?").content
+        print(f"LLM Response: {response}")
+    except Exception as e:
+        print(f"Error testing LLM: {e}")

redis_client.py

@@ -12,6 +12,7 @@ from typing import Optional, Any
 
 import redis
 import time
+from datetime import datetime
 
 # Thêm path để load config nếu cần
 project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), '..'))
@@ -123,7 +124,6 @@ class RedisClient:
                 return True
 
             # ISO timestamp -> unix timestamp logic
-            from datetime import datetime
             ts = int(datetime.now().timestamp() * 1000)
             if "time" in event_data:
                 try:
@@ -144,7 +144,22 @@ class RedisClient:
         try:
             if self._use_local:
                 db = self._load_local()
-                return list(db["events"].values())
+                events = list(db["events"].values())
+                filtered = []
+                for ev in events:
+                    try:
+                        # Parse time to check against range
+                        dt = datetime.fromisoformat(ev["time"])
+                        ts = int(dt.timestamp() * 1000)
+                        if start_ts <= ts <= end_ts:
+                            filtered.append(ev)
+                        else:
+                            logger.info(f"Event {ev.get('name')} ({ev.get('time')}) excluded: ts {ts} outside {start_ts}-{end_ts}")
+                    except (ValueError, KeyError, TypeError):
+                        # Fallback: if no time, only include if full range
+                        if start_ts == 0 and end_ts >= 3000000000000:
+                            filtered.append(ev)
+                return filtered
 
             index_key = self._key("evt", "index")
             event_ids = self._client.zrangebyscore(index_key, start_ts, end_ts)

tools/base.py

@@ -11,9 +11,17 @@ from langchain_openai import ChatOpenAI
 from langchain_huggingface import HuggingFacePipeline
 from langchain_google_genai import ChatGoogleGenerativeAI
 try:
-    from ..config import QWEN_API_KEY, QWEN_BASE_URL, QWEN_MODEL, LOG_LEVEL, USE_LOCAL_LLM, LOCAL_MODEL_ID, GEMINI_API_KEY
+    from ..config import (
+        QWEN_API_KEY, QWEN_BASE_URL, QWEN_MODEL, 
+        LOG_LEVEL, USE_LOCAL_LLM, LOCAL_MODEL_ID, 
+        GEMINI_API_KEY, OPENROUTER_API_KEY, OPENROUTER_BASE_URL, OPENROUTER_MODEL
+    )
 except (ImportError, ValueError):
-    from config import QWEN_API_KEY, QWEN_BASE_URL, QWEN_MODEL, LOG_LEVEL, USE_LOCAL_LLM, LOCAL_MODEL_ID, GEMINI_API_KEY
+    from config import (
+        QWEN_API_KEY, QWEN_BASE_URL, QWEN_MODEL, 
+        LOG_LEVEL, USE_LOCAL_LLM, LOCAL_MODEL_ID, 
+        GEMINI_API_KEY, OPENROUTER_API_KEY, OPENROUTER_BASE_URL, OPENROUTER_MODEL
+    )
 
 logger = logging.getLogger(__name__)
 
@@ -38,11 +46,22 @@ def get_llm():
     """
     Initialize and return the LLM based on configuration (Gemini > Local > Cloud Qwen).
     """
-    # 1. Prioritize Gemini if API key is present
+    # 1. Prioritize OpenRouter
+    if OPENROUTER_API_KEY and not OPENROUTER_API_KEY.startswith("your-"):
+        logger.info(f"Initializing OpenRouter LLM ({OPENROUTER_MODEL})...")
+        return ChatOpenAI(
+            model=OPENROUTER_MODEL,
+            api_key=OPENROUTER_API_KEY,
+            base_url=OPENROUTER_BASE_URL,
+            temperature=0.1,
+            max_tokens=4096,
+        )
+
+    # 2. Fallback to Gemini
     if GEMINI_API_KEY and not GEMINI_API_KEY.startswith("your-"):
         logger.info("Initializing Google Gemini LLM...")
         return ChatGoogleGenerativeAI(
-            model="gemini-2.0-flash", # or from config
+            model="gemini-2.0-flash", 
             google_api_key=GEMINI_API_KEY,
             temperature=0.1,
         )

tools/memory.py

@@ -50,11 +50,11 @@ def tool_save_memory(content: str, category: str = "general") -> dict:
     name="get_memories",
     description="Tìm kiếm và truy xuất các thông tin đã ghi nhớ trước đây dựa trên từ khóa.",
     parameters=[
-        {"name": "query", "type": "string", "description": "Từ khóa tìm kiếm thông tin trong bộ nhớ.", "required": False},
+        {"name": "query", "type": "string", "description": "TỪ KHÓA DUY NHẤT (VD: 'azure', 'ghét'). Để trống nếu muốn lấy toàn bộ 100 ghi nhớ mới nhất.", "required": False},
         {"name": "limit", "type": "integer", "description": "Số lượng kết quả tối đa.", "required": False}
     ]
 )
-def tool_get_memories(query: str = None, limit: int = 10) -> dict:
+def tool_get_memories(query: str = None, limit: int = 100) -> dict:
     """
     Retrieves memories from Redis.
     """

tools/scheduler.py

@@ -15,7 +15,7 @@ logger = logging.getLogger(__name__)
         {
             "name": "query", 
             "type": "string", 
-            "description": "Từ khóa tìm kiếm trong tên hoặc mô tả sự kiện. Ví dụ: 'họp nhóm', 'báo cáo'", 
+            "description": "TỪ KHÓA DUY NHẤT (VD: 'họp'). Để trống nếu muốn xem toàn bộ lịch trình.", 
             "required": False
         },
         {
@@ -49,10 +49,17 @@ def tool_get_schedule(query: str = "", date_str: str = "", room_id: str = None)
             start_ts = int(day_start.timestamp() * 1000)
             end_ts = int(day_end.timestamp() * 1000)
         else:
-            return {
-                "status": "error",
-                "message": f"Không thể hiểu được khoảng thời gian: '{date_str}'."
-            }
+            # Fallback for range-like strings (e.g., "next 2 weeks", "tuần tới")
+            range_keywords = ["tuần", "tháng", "next", "week", "month", "khoảng", "tới"]
+            if any(k in date_str.lower() for k in range_keywords):
+                logger.info(f"Date string '{date_str}' looks like a range. Returning all future events (30 days).")
+                start_ts = int(datetime.now().timestamp() * 1000)
+                end_ts = start_ts + (30 * 24 * 60 * 60 * 1000) # 30 days
+            else:
+                return {
+                    "status": "error",
+                    "message": f"Không thể hiểu được khoảng thời gian: '{date_str}'."
+                }
 
     # Retrieve from Redis
     events = redis_client.list_events(start_ts, end_ts)
@@ -68,6 +75,21 @@ def tool_get_schedule(query: str = "", date_str: str = "", room_id: str = None)
         if match:
             results.append(event)
 
+    # Robustness Fallback: 
+    # If results are empty and there was a query but no date_str, 
+    # check if the query was meant to be a date (e.g., "tối nay").
+    # We only fallback if the query contains common time-related keywords to avoid false positives (e.g., "ăn").
+    time_keywords = ["nay", "mai", "mốt", "hôm", "tối", "sáng", "chiều", "trưa", "ngày", "lịch", "tuần", "tháng"]
+    is_time_query = any(k in query.lower() for k in time_keywords) or any(char.isdigit() for char in query)
+    
+    if not results and query and not date_str and is_time_query:
+        fallback_date = dateparser.parse(query, settings={'PREFER_DATES_FROM': 'future'})
+        if fallback_date:
+            # Check if parsing was actually meaningful (not just a random number or word parsed as current year)
+            logger.info(f"Query '{query}' looks like a date. Retrying search with date filtering.")
+            # Recursive call with query moved to date_str
+            return tool_get_schedule(query="", date_str=query, room_id=room_id)
+
     # 2. Fetch Chat Context (Hybrid Memory)
     chat_context = []
     if room_id:

tools/summarizer.py

@@ -16,6 +16,8 @@ try:
 except (ImportError, ValueError):
     from redis_client import redis_client
 
+logger = logging.getLogger(__name__)
+
 # --- Pydantic Schemas ---
 class ThreadSummary(BaseModel):
     """Schema cho tóm tắt của một thread."""