Update app.py

app.py

@@ -452,87 +452,112 @@ weather_api_key = os.environ.get("WEATHER_API_KEY")
 
 def weather_agent_tool(query: str) -> str:
     try:
-        from dateutil import parser as date_parser
-        from zoneinfo import ZoneInfo
-        from geopy.geocoders import Nominatim
-        from timezonefinder import TimezoneFinder
-
         weather_api_key = os.environ.get("WEATHER_API_KEY")
         if not weather_api_key:
             return "Weather API key not found. Please set WEATHER_API_KEY env variable."
 
-        # Step 1: Use LLM to extract location (Few-Shot examples included)
+        # Step 1: Extract location
         location_prompt = f"""
 You are a location extractor. Given a user's query about weather, extract the location mentioned in it.
-If no location is found, return "London".
+If not found, return "London".
 
 Examples:
-- "What's the weather in Tokyo now?" → Tokyo
-- "Will it rain tomorrow in New York?" → New York
-- "今天台北會下雨嗎？" → Taipei
-- "How's the weather?" → London
-- "Is it hot in LDN later?" → London
+- "Is it gonna rain in Tokyo?" → Tokyo
+- "Will it be hot in New York later?" → New York
+- "明天下午高雄會不會下雨？" → Kaohsiung
+- "How’s the weather?" → London
 
-Now extract the location from: "{query}"
+Query: "{query}"
 """
-        location_response = llm_gpt4.invoke(location_prompt)
-        location = location_response.content.strip() if isinstance(location_response, AIMessage) else str(location_response).strip()
+        location_resp = llm_gpt4.invoke(location_prompt)
+        location = location_resp.content.strip() if isinstance(location_resp, AIMessage) else str(location_resp).strip()
 
-        # Step 2: Convert location to local time
-        geo = Nominatim(user_agent="weather_agent_demo")
+        # Step 2: Get timezone
+        geo = Nominatim(user_agent="weather_agent")
         loc = geo.geocode(location)
+        if not loc:
+            return f"Unable to resolve location: {location}"
         tf = TimezoneFinder()
-        tz_str = tf.timezone_at(lat=loc.latitude, lng=loc.longitude) if loc else "Europe/London"
+        tz_str = tf.timezone_at(lat=loc.latitude, lng=loc.longitude) or "Europe/London"
         local_now = datetime.now(ZoneInfo(tz_str))
-        local_now_str = local_now.strftime("%Y-%m-%d %H:%M:%S")
 
-        # Step 3: Use LLM to extract intended date (Few-Shot style prompt)
+        # Step 3: Extract intended datetime
         time_prompt = f"""
-Today is {local_now_str} in {location}.
+Today is {local_now.strftime('%Y-%m-%d %H:%M:%S')} in {location}.
+
+You are a time extractor. Given a user's question about weather, extract the local date and hour it refers to.
+Return format: YYYY-MM-DD HH:00 (24h format)
 
-You are a time extractor. Given a weather-related question, convert the user's time reference into a local date.
-Use format: YYYY-MM-DD.
 Examples:
-- "What's the weather in New York now?" → {local_now.strftime('%Y-%m-%d')}
-- "Will it rain in Paris tomorrow?" → {(local_now + timedelta(days=1)).strftime('%Y-%m-%d')}
-- "How's the weather in Tokyo in 2 hours?" → {local_now.strftime('%Y-%m-%d')}
-- "昨天倫敦的天氣如何？" → {(local_now - timedelta(days=1)).strftime('%Y-%m-%d')}
-- "Weather in Taipei this weekend?" → {local_now.strftime('%Y-%m-%d')} (or closest matching day)
-If unclear, assume today.
+- "Is it gonna rain in Paris in 2 hours?" → {(local_now + timedelta(hours=2)).strftime('%Y-%m-%d %H:00')}
+- "Will it be cold tomorrow?" → {(local_now + timedelta(days=1)).strftime('%Y-%m-%d 12:00')}
+- "今天下午高雄會熱嗎？" → {(local_now.replace(hour=15)).strftime('%Y-%m-%d %H:00')}
+- "What's the weather?" → {local_now.strftime('%Y-%m-%d %H:00')}
+
 Query: "{query}"
-Return only the date like 2025-04-04.
 """
-        date_response = llm_gpt4.invoke(time_prompt)
-        date_str = date_response.content.strip() if isinstance(date_response, AIMessage) else str(date_response).strip()
-        query_date = date_parser.parse(date_str).date()
-        today = local_now.date()
-
-        # Step 4: Query WeatherAPI with appropriate endpoint
-        if query_date == today:
-            url = f"http://api.weatherapi.com/v1/current.json?key={weather_api_key}&q={location}&aqi=no"
-            data = requests.get(url).json()
-            cur = data["current"]
-            return f"The current weather in {location.title()} is {cur['condition']['text']} with {cur['temp_c']}°C (feels like {cur['feelslike_c']}°C), humidity {cur['humidity']}%, wind {cur['wind_kph']} kph."
-
-        elif query_date > today:
-            url = f"http://api.weatherapi.com/v1/forecast.json?key={weather_api_key}&q={location}&days=3"
-            data = requests.get(url).json()
-            forecast_days = data.get("forecast", {}).get("forecastday", [])
-            for day in forecast_days:
-                if day["date"] == date_str:
-                    d = day["day"]
-                    condition = d['condition']['text'].lower()
-                    rain_comment = "Yes, it might rain." if "rain" in condition else "No rain expected."
-                    return f"{rain_comment} Forecast for {location.title()} on {query_date.strftime('%A, %B %d')}: {d['condition']['text']}, avg {d['avgtemp_c']}°C, humidity {d['avghumidity']}%, max {d['maxtemp_c']}°C, min {d['mintemp_c']}°C."
-            return f"Forecast unavailable for {query_date} (WeatherAPI supports up to 3 days ahead)."
-
+        time_resp = llm_gpt4.invoke(time_prompt)
+        time_str = time_resp.content.strip() if isinstance(time_resp, AIMessage) else str(time_resp).strip()
+        target_dt = date_parser.parse(time_str)
+
+        # Step 4: Call WeatherAPI
+        if target_dt.date() < local_now.date() - timedelta(days=7):
+            return "Only supports up to 7 days of historical data."
+        elif target_dt.date() > local_now.date() + timedelta(days=2):
+            return "Only supports up to 3 days of forecast."
+
+        if target_dt.date() < local_now.date():
+            # Use history API
+            url = f"http://api.weatherapi.com/v1/history.json?key={weather_api_key}&q={location}&dt={target_dt.strftime('%Y-%m-%d')}"
         else:
-            url = f"http://api.weatherapi.com/v1/history.json?key={weather_api_key}&q={location}&dt={date_str}"
-            data = requests.get(url).json()
-            hist = data.get("forecast", {}).get("forecastday", [{}])[0].get("day", {})
-            if not hist:
-                return f"Historical weather not found for {location.title()} on {query_date}."
-            return f"The weather in {location.title()} on {query_date.strftime('%A, %B %d')}: {hist['condition']['text']}, avg {hist['avgtemp_c']}°C, humidity {hist['avghumidity']}%, max {hist['maxtemp_c']}°C, min {hist['mintemp_c']}°C."
+            # Use forecast API (for today or future)
+            url = f"http://api.weatherapi.com/v1/forecast.json?key={weather_api_key}&q={location}&days=3&aqi=no&alerts=no"
+
+        data = requests.get(url).json()
+        forecast_hours = []
+        if "forecast" in data:
+            for day in data["forecast"]["forecastday"]:
+                for hour in day["hour"]:
+                    forecast_hours.append(hour)
+        elif "forecastday" in data:
+            forecast_hours = data["forecastday"][0]["hour"]
+        else:
+            return "No forecast data available."
+
+        # Step 5: Find closest hour
+        min_diff = float("inf")
+        closest_hour = None
+        for hour_data in forecast_hours:
+            hour_dt = date_parser.parse(hour_data["time"]).astimezone(ZoneInfo(tz_str))
+            diff = abs((hour_dt - target_dt).total_seconds())
+            if diff < min_diff:
+                min_diff = diff
+                closest_hour = hour_data
+
+        if not closest_hour:
+            return f"No hourly data found for {target_dt.strftime('%Y-%m-%d %H:%M')}."
+
+        # Step 6: Generate summary
+        condition = closest_hour["condition"]["text"]
+        temp = closest_hour["temp_c"]
+        feels = closest_hour["feelslike_c"]
+        humidity = closest_hour["humidity"]
+        chance_rain = closest_hour.get("chance_of_rain", 0)
+        hour_str = closest_hour["time"].split(" ")[1]
+
+        summary_prompt = f"""
+Summarise this weather forecast naturally:
+Location: {location}
+Time: {target_dt.strftime('%Y-%m-%d')} at {hour_str}
+Condition: {condition}
+Temp: {temp}°C (Feels like {feels}°C)
+Humidity: {humidity}%
+Chance of rain: {chance_rain}%
+
+Make it short, friendly, and human-style.
+"""
+        response = llm_gpt4.invoke(summary_prompt)
+        return response.content.strip() if isinstance(response, AIMessage) else str(response)
 
     except Exception as e:
         return f"Weather Agent Error: {e}"
@@ -540,89 +565,114 @@ Return only the date like 2025-04-04.
 
 @tool("weather")
 def weather_tool(query: str) -> str:
-    """Weather Agent: Provide real-time, future, or historical weather info (max 3 days forecast, 7 days history)."""
+    """Weather Agent: Return current, hourly, or historical weather info using WeatherAPI."""
     try:
-        from dateutil import parser as date_parser
-        from zoneinfo import ZoneInfo
-        from geopy.geocoders import Nominatim
-        from timezonefinder import TimezoneFinder
-
         weather_api_key = os.environ.get("WEATHER_API_KEY")
         if not weather_api_key:
             return "Weather API key not found. Please set WEATHER_API_KEY env variable."
 
-        # Step 1: Use LLM to extract location (Few-Shot examples included)
+        # Step 1: Extract location
         location_prompt = f"""
 You are a location extractor. Given a user's query about weather, extract the location mentioned in it.
-If no location is found, return "London".
+If not found, return "London".
 
 Examples:
-- "What's the weather in Tokyo now?" → Tokyo
-- "Will it rain tomorrow in New York?" → New York
-- "今天台北會下雨嗎？" → Taipei
-- "How's the weather?" → London
-- "Is it hot in LDN later?" → London
+- "Is it gonna rain in Tokyo?" → Tokyo
+- "Will it be hot in New York later?" → New York
+- "明天��午高雄會不會下雨？" → Kaohsiung
+- "How’s the weather?" → London
 
-Now extract the location from: "{query}"
+Query: "{query}"
 """
-        location_response = llm_gpt4.invoke(location_prompt)
-        location = location_response.content.strip() if isinstance(location_response, AIMessage) else str(location_response).strip()
+        location_resp = llm_gpt4.invoke(location_prompt)
+        location = location_resp.content.strip() if isinstance(location_resp, AIMessage) else str(location_resp).strip()
 
-        # Step 2: Convert location to local time
-        geo = Nominatim(user_agent="weather_agent_demo")
+        # Step 2: Get timezone
+        geo = Nominatim(user_agent="weather_agent")
         loc = geo.geocode(location)
+        if not loc:
+            return f"Unable to resolve location: {location}"
         tf = TimezoneFinder()
-        tz_str = tf.timezone_at(lat=loc.latitude, lng=loc.longitude) if loc else "Europe/London"
+        tz_str = tf.timezone_at(lat=loc.latitude, lng=loc.longitude) or "Europe/London"
         local_now = datetime.now(ZoneInfo(tz_str))
-        local_now_str = local_now.strftime("%Y-%m-%d %H:%M:%S")
 
-        # Step 3: Use LLM to extract intended date (Few-Shot style prompt)
+        # Step 3: Extract intended datetime
         time_prompt = f"""
-Today is {local_now_str} in {location}.
+Today is {local_now.strftime('%Y-%m-%d %H:%M:%S')} in {location}.
+
+You are a time extractor. Given a user's question about weather, extract the local date and hour it refers to.
+Return format: YYYY-MM-DD HH:00 (24h format)
 
-You are a time extractor. Given a weather-related question, convert the user's time reference into a local date.
-Use format: YYYY-MM-DD.
 Examples:
-- "What's the weather in New York now?" → {local_now.strftime('%Y-%m-%d')}
-- "Will it rain in Paris tomorrow?" → {(local_now + timedelta(days=1)).strftime('%Y-%m-%d')}
-- "How's the weather in Tokyo in 2 hours?" → {local_now.strftime('%Y-%m-%d')}
-- "昨天倫敦的天氣如何？" → {(local_now - timedelta(days=1)).strftime('%Y-%m-%d')}
-- "Weather in Taipei this weekend?" → {local_now.strftime('%Y-%m-%d')} (or closest matching day)
-If unclear, assume today.
+- "Is it gonna rain in Paris in 2 hours?" → {(local_now + timedelta(hours=2)).strftime('%Y-%m-%d %H:00')}
+- "Will it be cold tomorrow?" → {(local_now + timedelta(days=1)).strftime('%Y-%m-%d 12:00')}
+- "今天下午高雄會熱嗎？" → {(local_now.replace(hour=15)).strftime('%Y-%m-%d %H:00')}
+- "What's the weather?" → {local_now.strftime('%Y-%m-%d %H:00')}
+
 Query: "{query}"
-Return only the date like 2025-04-04.
 """
-        date_response = llm_gpt4.invoke(time_prompt)
-        date_str = date_response.content.strip() if isinstance(date_response, AIMessage) else str(date_response).strip()
-        query_date = date_parser.parse(date_str).date()
-        today = local_now.date()
-
-        # Step 4: Query WeatherAPI with appropriate endpoint
-        if query_date == today:
-            url = f"http://api.weatherapi.com/v1/current.json?key={weather_api_key}&q={location}&aqi=no"
-            data = requests.get(url).json()
-            cur = data["current"]
-            return f"The current weather in {location.title()} is {cur['condition']['text']} with {cur['temp_c']}°C (feels like {cur['feelslike_c']}°C), humidity {cur['humidity']}%, wind {cur['wind_kph']} kph."
-
-        elif query_date > today:
-            url = f"http://api.weatherapi.com/v1/forecast.json?key={weather_api_key}&q={location}&days=3"
-            data = requests.get(url).json()
-            forecast_days = data.get("forecast", {}).get("forecastday", [])
-            for day in forecast_days:
-                if day["date"] == date_str:
-                    d = day["day"]
-                    condition = d['condition']['text'].lower()
-                    rain_comment = "Yes, it might rain." if "rain" in condition else "No rain expected."
-                    return f"{rain_comment} Forecast for {location.title()} on {query_date.strftime('%A, %B %d')}: {d['condition']['text']}, avg {d['avgtemp_c']}°C, humidity {d['avghumidity']}%, max {d['maxtemp_c']}°C, min {d['mintemp_c']}°C."
-            return f"Forecast unavailable for {query_date} (WeatherAPI supports up to 3 days ahead)."
-
+        time_resp = llm_gpt4.invoke(time_prompt)
+        time_str = time_resp.content.strip() if isinstance(time_resp, AIMessage) else str(time_resp).strip()
+        target_dt = date_parser.parse(time_str)
+
+        # Step 4: Call WeatherAPI
+        if target_dt.date() < local_now.date() - timedelta(days=7):
+            return "Only supports up to 7 days of historical data."
+        elif target_dt.date() > local_now.date() + timedelta(days=2):
+            return "Only supports up to 3 days of forecast."
+
+        if target_dt.date() < local_now.date():
+            # Use history API
+            url = f"http://api.weatherapi.com/v1/history.json?key={weather_api_key}&q={location}&dt={target_dt.strftime('%Y-%m-%d')}"
         else:
-            url = f"http://api.weatherapi.com/v1/history.json?key={weather_api_key}&q={location}&dt={date_str}"
-            data = requests.get(url).json()
-            hist = data.get("forecast", {}).get("forecastday", [{}])[0].get("day", {})
-            if not hist:
-                return f"Historical weather not found for {location.title()} on {query_date}."
-            return f"The weather in {location.title()} on {query_date.strftime('%A, %B %d')}: {hist['condition']['text']}, avg {hist['avgtemp_c']}°C, humidity {hist['avghumidity']}%, max {hist['maxtemp_c']}°C, min {hist['mintemp_c']}°C."
+            # Use forecast API (for today or future)
+            url = f"http://api.weatherapi.com/v1/forecast.json?key={weather_api_key}&q={location}&days=3&aqi=no&alerts=no"
+
+        data = requests.get(url).json()
+        forecast_hours = []
+        if "forecast" in data:
+            for day in data["forecast"]["forecastday"]:
+                for hour in day["hour"]:
+                    forecast_hours.append(hour)
+        elif "forecastday" in data:
+            forecast_hours = data["forecastday"][0]["hour"]
+        else:
+            return "No forecast data available."
+
+        # Step 5: Find closest hour
+        min_diff = float("inf")
+        closest_hour = None
+        for hour_data in forecast_hours:
+            hour_dt = date_parser.parse(hour_data["time"]).astimezone(ZoneInfo(tz_str))
+            diff = abs((hour_dt - target_dt).total_seconds())
+            if diff < min_diff:
+                min_diff = diff
+                closest_hour = hour_data
+
+        if not closest_hour:
+            return f"No hourly data found for {target_dt.strftime('%Y-%m-%d %H:%M')}."
+
+        # Step 6: Generate summary
+        condition = closest_hour["condition"]["text"]
+        temp = closest_hour["temp_c"]
+        feels = closest_hour["feelslike_c"]
+        humidity = closest_hour["humidity"]
+        chance_rain = closest_hour.get("chance_of_rain", 0)
+        hour_str = closest_hour["time"].split(" ")[1]
+
+        summary_prompt = f"""
+Summarise this weather forecast naturally:
+Location: {location}
+Time: {target_dt.strftime('%Y-%m-%d')} at {hour_str}
+Condition: {condition}
+Temp: {temp}°C (Feels like {feels}°C)
+Humidity: {humidity}%
+Chance of rain: {chance_rain}%
+
+Make it short, friendly, and human-style.
+"""
+        response = llm_gpt4.invoke(summary_prompt)
+        return response.content.strip() if isinstance(response, AIMessage) else str(response)
 
     except Exception as e:
         return f"Weather Agent Error: {e}"