geoai-py 0.15.0__py2.py3-none-any.whl → 0.18.0__py2.py3-none-any.whl

geoai/agents/geo_agents.py

@@ -1,11 +1,10 @@
 from __future__ import annotations
 
-import asyncio
+import json
 import os
 import uuid
-from concurrent.futures import ThreadPoolExecutor
 from types import SimpleNamespace
-from typing import Any, Callable, Optional
+from typing import Any, Callable, Dict, Optional
 
 import boto3
 import ipywidgets as widgets
@@ -19,14 +18,47 @@ from strands.models.anthropic import AnthropicModel
 from strands.models.ollama import OllamaModel as _OllamaModel
 from strands.models.openai import OpenAIModel
 
+from .catalog_tools import CatalogTools
 from .map_tools import MapSession, MapTools
+from .stac_tools import STACTools
 
-try:
-    import nest_asyncio
 
-    nest_asyncio.apply()
-except Exception:
-    pass
+class UICallbackHandler:
+    """Callback handler that updates UI status widget with agent progress.
+
+    This handler intercepts tool calls and progress events to provide
+    real-time feedback without overwhelming the user.
+    """
+
+    def __init__(self, status_widget=None):
+        """Initialize the callback handler.
+
+        Args:
+            status_widget: Optional ipywidgets.HTML widget to update with status.
+        """
+        self.status_widget = status_widget
+        self.current_tool = None
+
+    def __call__(self, **kwargs):
+        """Handle callback events from the agent.
+
+        Args:
+            **kwargs: Event data from the agent execution.
+        """
+        # Track when tools are being called
+        if "current_tool_use" in kwargs and kwargs["current_tool_use"].get("name"):
+            tool_name = kwargs["current_tool_use"]["name"]
+            self.current_tool = tool_name
+
+            # Update status widget if available
+            if self.status_widget is not None:
+                # Make tool names more user-friendly
+                friendly_name = tool_name.replace("_", " ").title()
+                self.status_widget.value = (
+                    f"<span style='color:#0a7'>"
+                    f"<i class='fas fa-spinner fa-spin' style='font-size:1.2em'></i> "
+                    f"{friendly_name}...</span>"
+                )
 
 
 class OllamaModel(_OllamaModel):
@@ -170,18 +202,6 @@ def create_bedrock_model(
     )
 
 
-def _ensure_loop() -> asyncio.AbstractEventLoop:
-    try:
-        loop = asyncio.get_event_loop()
-    except RuntimeError:
-        loop = asyncio.new_event_loop()
-        asyncio.set_event_loop(loop)
-    if loop.is_closed():
-        loop = asyncio.new_event_loop()
-        asyncio.set_event_loop(loop)
-    return loop
-
-
 class GeoAgent(Agent):
     """Geospatial AI agent with interactive mapping capabilities."""
 
@@ -311,30 +331,14 @@ class GeoAgent(Agent):
     def ask(self, prompt: str) -> str:
         """Send a single-turn prompt to the agent.
 
-        Runs entirely on the same thread/event loop as the Agent
-        to avoid cross-loop asyncio object issues.
-
         Args:
             prompt: The text prompt to send to the agent.
 
         Returns:
             The agent's response as a string.
         """
-        # Ensure there's an event loop bound to this thread (Jupyter-safe)
-        loop = _ensure_loop()
-
-        # Preserve existing conversation messages
-        existing_messages = self.messages.copy()
-
-        # Create a fresh model but keep conversation history
-        self.model = self._model_factory()
-
-        # Restore the conversation messages
-        self.messages = existing_messages
-
-        # Execute the prompt using the Agent's async API on this loop
-        # Avoid Agent.__call__ since it spins a new thread+loop
-        result = loop.run_until_complete(self.invoke_async(prompt))
+        # Use strands' built-in __call__ method which now supports multiple calls
+        result = self(prompt)
         return getattr(result, "final_text", str(result))
 
     def show_ui(self, *, height: int = 700) -> None:
@@ -349,7 +353,10 @@ class GeoAgent(Agent):
             m.create_container()
 
         map_panel = widgets.VBox(
-            [widgets.HTML("<h3 style='margin:0 0 8px 0'>Map</h3>"), m.container],
+            [
+                widgets.HTML("<h3 style='margin:0 0 8px 0'>Map</h3>"),
+                m.floating_sidebar_widget,
+            ],
             layout=widgets.Layout(
                 flex="2 1 0%",
                 min_width="520px",
@@ -391,7 +398,10 @@ class GeoAgent(Agent):
         btn_clear = widgets.Button(
             description="Clear", icon="trash", layout=widgets.Layout(width="120px")
         )
-        status = widgets.HTML("<span style='color:#666'>Ready.</span>")
+        status = widgets.HTML(
+            "<link rel='stylesheet' href='https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.4.0/css/all.min.css'>"
+            "<span style='color:#666'>Ready.</span>"
+        )
 
         examples = widgets.Dropdown(
             options=[
@@ -427,7 +437,6 @@ class GeoAgent(Agent):
             examples=examples,
         )
         self._pending = {"fut": None}
-        self._executor = ThreadPoolExecutor(max_workers=1)
 
         def _esc(s: str) -> str:
             """Escape HTML characters in a string.
@@ -494,7 +503,14 @@ class GeoAgent(Agent):
             _lock(True)
             self._ui.status.value = "<span style='color:#0a7'>Running…</span>"
             try:
-                out = self.ask(text)  # fresh Agent/model per call, silent
+                # Create a callback handler that updates the status widget
+                callback_handler = UICallbackHandler(status_widget=self._ui.status)
+
+                # Temporarily set callback_handler for this call
+                old_callback = self.callback_handler
+                self.callback_handler = callback_handler
+
+                out = self.ask(text)  # fresh Agent/model per call, with callback
                 _append("assistant", out)
                 self._ui.status.value = "<span style='color:#0a7'>Done.</span>"
             except Exception as e:
@@ -503,6 +519,8 @@ class GeoAgent(Agent):
                     "<span style='color:#c00'>Finished with an issue.</span>"
                 )
             finally:
+                # Restore old callback handler
+                self.callback_handler = old_callback
                 self._ui.inp.value = ""
                 _lock(False)
 
@@ -593,3 +611,877 @@ class GeoAgent(Agent):
             [map_panel, right], layout=widgets.Layout(width="100%", gap="8px")
         )
         display(root)
+
+
+class STACAgent(Agent):
+    """AI agent for searching and interacting with STAC catalogs."""
+
+    def __init__(
+        self,
+        *,
+        model: str = "llama3.1",
+        system_prompt: str = "default",
+        endpoint: str = "https://planetarycomputer.microsoft.com/api/stac/v1",
+        model_args: dict = None,
+        map_instance: Optional[leafmap.Map] = None,
+        **kwargs: Any,
+    ) -> None:
+        """Initialize the STAC Agent.
+
+        Args:
+            model: Model identifier (default: "llama3.1").
+            system_prompt: System prompt for the agent (default: "default").
+            endpoint: STAC API endpoint URL. Defaults to Microsoft Planetary Computer.
+            model_args: Additional keyword arguments for the model.
+            map_instance: Optional leafmap.Map instance for visualization. If None, creates a new one.
+            **kwargs: Additional keyword arguments for the Agent.
+        """
+        self.stac_tools: STACTools = STACTools(endpoint=endpoint)
+        self.map_instance = map_instance if map_instance is not None else leafmap.Map()
+
+        if model_args is None:
+            model_args = {}
+
+        # --- save a model factory we can call each turn ---
+        if model == "llama3.1":
+            self._model_factory: Callable[[], OllamaModel] = (
+                lambda: create_ollama_model(
+                    host="http://localhost:11434", model_id=model, **model_args
+                )
+            )
+        elif isinstance(model, str):
+            self._model_factory: Callable[[], BedrockModel] = (
+                lambda: create_bedrock_model(model_id=model, **model_args)
+            )
+        elif isinstance(model, OllamaModel):
+            # Extract configuration from existing OllamaModel and create new instances
+            model_id = model.config["model_id"]
+            host = model.host
+            client_args = model.client_args
+            self._model_factory: Callable[[], OllamaModel] = (
+                lambda: create_ollama_model(
+                    host=host, model_id=model_id, client_args=client_args, **model_args
+                )
+            )
+        elif isinstance(model, OpenAIModel):
+            # Extract configuration from existing OpenAIModel and create new instances
+            model_id = model.config["model_id"]
+            client_args = model.client_args.copy()
+            self._model_factory: Callable[[], OpenAIModel] = (
+                lambda mid=model_id, client_args=client_args: create_openai_model(
+                    model_id=mid, client_args=client_args, **model_args
+                )
+            )
+        elif isinstance(model, AnthropicModel):
+            # Extract configuration from existing AnthropicModel and create new instances
+            model_id = model.config["model_id"]
+            client_args = model.client_args.copy()
+            self._model_factory: Callable[[], AnthropicModel] = (
+                lambda mid=model_id, client_args=client_args: create_anthropic_model(
+                    model_id=mid, client_args=client_args, **model_args
+                )
+            )
+        else:
+            raise ValueError(f"Invalid model: {model}")
+
+        # build initial model (first turn)
+        model = self._model_factory()
+
+        if system_prompt == "default":
+            system_prompt = """You are a STAC search agent. Follow these steps EXACTLY:
+
+1. Determine collection ID based on data type:
+   - "sentinel-2-l2a" for Sentinel-2 or optical satellite imagery
+   - "landsat-c2-l2" for Landsat
+   - "naip" for NAIP or aerial imagery (USA only)
+   - "sentinel-1-grd" for Sentinel-1 or SAR/radar
+   - "cop-dem-glo-30" for DEM, elevation, or terrain data
+   - "aster-l1t" for ASTER
+   For other data (e.g., MODIS, land cover): call list_collections(filter_keyword="<keyword>")
+
+2. If location mentioned:
+   - Call geocode_location("<name>") FIRST
+   - WAIT for the response
+   - Extract the "bbox" array from the JSON response
+   - This bbox is [west, south, east, north] format
+
+3. Call search_items():
+   - collection: REQUIRED
+   - bbox: Use the EXACT bbox array from geocode_location (REQUIRED if location mentioned)
+   - time_range: "YYYY-MM-DD/YYYY-MM-DD" format if dates mentioned
+   - query: Use for cloud cover filtering (see examples)
+   - max_items: 1
+
+Cloud cover filtering:
+   - "<10% cloud": query={"eo:cloud_cover": {"lt": 10}}
+   - "<20% cloud": query={"eo:cloud_cover": {"lt": 20}}
+   - "<5% cloud": query={"eo:cloud_cover": {"lt": 5}}
+
+Examples:
+1. "Find Landsat over Paris from June to July 2023"
+   geocode_location("Paris") → {"bbox": [2.224, 48.815, 2.469, 48.902], ...}
+   search_items(collection="landsat-c2-l2", bbox=[2.224, 48.815, 2.469, 48.902], time_range="2023-06-01/2023-07-31")
+
+2. "Find Landsat with <10% cloud cover over Paris"
+   geocode_location("Paris") → {"bbox": [2.224, 48.815, 2.469, 48.902], ...}
+   search_items(collection="landsat-c2-l2", bbox=[2.224, 48.815, 2.469, 48.902], query={"eo:cloud_cover": {"lt": 10}})
+
+4. Return first item as JSON:
+{"id": "...", "collection": "...", "datetime": "...", "bbox": [...], "assets": [...]}
+
+ERROR HANDLING:
+- If no items found: {"error": "No items found"}
+- If tool result too large: {"error": "Result too large, try narrower search"}
+- If tool error: {"error": "Search failed: <error message>"}
+
+CRITICAL: Return ONLY JSON. NO explanatory text, NO made-up data."""
+
+        super().__init__(
+            name="STAC Search Agent",
+            model=model,
+            tools=[
+                self.stac_tools.list_collections,
+                self.stac_tools.search_items,
+                self.stac_tools.get_item_info,
+                self.stac_tools.geocode_location,
+                self.stac_tools.get_common_collections,
+            ],
+            system_prompt=system_prompt,
+            callback_handler=None,
+        )
+
+    def _extract_search_items_payload(self, result: Any) -> Optional[Dict[str, Any]]:
+        """Return the parsed payload from the search_items tool, if available."""
+        # Try to get tool_results from the result object
+        tool_results = getattr(result, "tool_results", None)
+        if tool_results:
+            for tool_result in tool_results:
+                if getattr(tool_result, "tool_name", "") != "search_items":
+                    continue
+
+                payload = getattr(tool_result, "result", None)
+                if payload is None:
+                    continue
+
+                if isinstance(payload, str):
+                    try:
+                        payload = json.loads(payload)
+                    except json.JSONDecodeError:
+                        continue
+
+                if isinstance(payload, dict):
+                    return payload
+
+        # Alternative: check messages for tool results
+        messages = getattr(self, "messages", [])
+        for msg in reversed(messages):  # Check recent messages first
+            # Handle dict-style messages
+            if isinstance(msg, dict):
+                role = msg.get("role")
+                content = msg.get("content", [])
+
+                # Look for tool results in user messages (strands pattern)
+                if role == "user" and isinstance(content, list):
+                    for item in content:
+                        if isinstance(item, dict) and "toolResult" in item:
+                            tool_result = item["toolResult"]
+                            # Check if this is a search_items result
+                            # We need to look at the preceding assistant message to identify the tool
+                            if tool_result.get("status") == "success":
+                                result_content = tool_result.get("content", [])
+                                if isinstance(result_content, list) and result_content:
+                                    text_content = result_content[0].get("text", "")
+                                    try:
+                                        payload = json.loads(text_content)
+                                        # Return ANY search_items payload, even if items is empty
+                                        # This is identified by having "query" and "collection" fields
+                                        if (
+                                            "query" in payload
+                                            and "collection" in payload
+                                            and "items" in payload
+                                        ):
+                                            return payload
+                                    except json.JSONDecodeError:
+                                        continue
+
+        return None
+
+    def ask(self, prompt: str) -> str:
+        """Send a single-turn prompt to the agent.
+
+        Args:
+            prompt: The text prompt to send to the agent.
+
+        Returns:
+            The agent's response as a string (JSON format for search queries).
+        """
+        # Use strands' built-in __call__ method which now supports multiple calls
+        result = self(prompt)
+
+        search_payload = self._extract_search_items_payload(result)
+        if search_payload is not None:
+            if "error" in search_payload:
+                return json.dumps({"error": search_payload["error"]}, indent=2)
+
+            items = search_payload.get("items") or []
+            if items:
+                return json.dumps(items[0], indent=2)
+
+            return json.dumps({"error": "No items found"}, indent=2)
+
+        return getattr(result, "final_text", str(result))
+
+    def search_and_get_first_item(self, prompt: str) -> Optional[Dict[str, Any]]:
+        """Search for imagery and return the first item as a structured dict.
+
+        This method sends a search query to the agent, extracts the search results
+        directly from the tool calls, and returns the first item as a STACItemInfo-compatible
+        dictionary.
+
+        Note: This method uses LLM inference which adds ~5-10 seconds overhead.
+        For faster searches, use STACTools directly:
+            >>> from geoai.agents import STACTools
+            >>> tools = STACTools()
+            >>> result = tools.search_items(
+            ...     collection="sentinel-2-l2a",
+            ...     bbox=[-122.5, 37.7, -122.4, 37.8],
+            ...     time_range="2024-08-01/2024-08-31"
+            ... )
+
+        Args:
+            prompt: Natural language search query (e.g., "Find Sentinel-2 imagery
+                    over San Francisco in September 2024").
+
+        Returns:
+            Dictionary containing STACItemInfo fields (id, collection, datetime,
+            bbox, assets, properties), or None if no results found.
+
+        Example:
+            >>> agent = STACAgent()
+            >>> item = agent.search_and_get_first_item(
+            ...     "Find Sentinel-2 imagery over Paris in summer 2023"
+            ... )
+            >>> print(item['id'])
+            >>> print(item['assets'][0]['key'])    # or 'title'
+        """
+        # Use strands' built-in __call__ method which now supports multiple calls
+        result = self(prompt)
+
+        search_payload = self._extract_search_items_payload(result)
+        if search_payload is not None:
+            if "error" in search_payload:
+                print(f"Search error: {search_payload['error']}")
+                return None
+
+            items = search_payload.get("items") or []
+            if items:
+                return items[0]
+
+            print("No items found in search results")
+            return None
+
+        # Fallback: try to parse the final text response
+        response = getattr(result, "final_text", str(result))
+
+        try:
+            item_data = json.loads(response)
+
+            if "error" in item_data:
+                print(f"Search error: {item_data['error']}")
+                return None
+
+            if not all(k in item_data for k in ["id", "collection"]):
+                print("Response missing required fields (id, collection)")
+                return None
+
+            return item_data
+
+        except json.JSONDecodeError:
+            print("Could not extract item data from agent response")
+            return None
+
+    def _visualize_stac_item(self, item: Dict[str, Any]) -> None:
+        """Visualize a STAC item on the map.
+
+        Args:
+            item: STAC item dictionary with id, collection, assets, etc.
+        """
+        if not item or "id" not in item or "collection" not in item:
+            return
+
+        # Get the collection and item ID
+        collection = item.get("collection")
+        item_id = item.get("id")
+
+        kwargs = {}
+
+        # Determine which assets to display based on collection
+        assets = None
+        if collection == "sentinel-2-l2a":
+            assets = ["B04", "B03", "B02"]  # True color RGB
+        elif collection == "landsat-c2-l2":
+            assets = ["red", "green", "blue"]  # Landsat RGB
+        elif collection == "naip":
+            assets = ["image"]  # NAIP 4-band imagery
+        elif "sentinel-1" in collection:
+            assets = ["vv"]  # Sentinel-1 VV polarization
+        elif collection == "cop-dem-glo-30":
+            assets = ["data"]
+            kwargs["colormap_name"] = "terrain"
+        elif collection == "aster-l1t":
+            assets = ["VNIR"]  # ASTER L1T imagery
+        elif collection == "3dep-lidar-hag":
+            assets = ["data"]
+            kwargs["colormap_name"] = "terrain"
+        else:
+            # Try to find common asset names
+            if "assets" in item:
+                asset_keys = [
+                    a.get("key") for a in item["assets"] if isinstance(a, dict)
+                ]
+                # Look for visual or RGB assets
+                for possible in ["visual", "image", "data"]:
+                    if possible in asset_keys:
+                        assets = [possible]
+                        break
+                # If still no assets, use first few assets
+                if not assets and asset_keys:
+                    assets = asset_keys[:1]
+
+        if not assets:
+            return None
+        try:
+            # Add the STAC layer to the map
+            layer_name = f"{collection[:20]}_{item_id[:15]}"
+            self.map_instance.add_stac_layer(
+                collection=collection,
+                item=item_id,
+                assets=assets,
+                name=layer_name,
+                before_id=self.map_instance.first_symbol_layer_id,
+                **kwargs,
+            )
+            return assets  # Return the assets that were visualized
+        except Exception as e:
+            print(f"Could not visualize item on map: {e}")
+            return None
+
+    def show_ui(self, *, height: int = 700) -> None:
+        """Display an interactive UI with map and chat interface for STAC searches.
+
+        Args:
+            height: Height of the UI in pixels (default: 700).
+        """
+        m = self.map_instance
+        if not hasattr(m, "container") or m.container is None:
+            m.create_container()
+
+        map_panel = widgets.VBox(
+            [
+                widgets.HTML("<h3 style='margin:0 0 8px 0'>Map</h3>"),
+                m.floating_sidebar_widget,
+            ],
+            layout=widgets.Layout(
+                flex="2 1 0%",
+                min_width="520px",
+                border="1px solid #ddd",
+                padding="8px",
+                height=f"{height}px",
+                overflow="hidden",
+            ),
+        )
+
+        # ----- chat widgets -----
+        session_id = str(uuid.uuid4())[:8]
+        title = widgets.HTML(
+            f"<h3 style='margin:0'>STAC Search Agent</h3>"
+            f"<p style='margin:4px 0 8px;color:#666'>Session: {session_id}</p>"
+        )
+        log = widgets.HTML(
+            value="<div style='color:#777'>No messages yet. Try searching for satellite imagery!</div>",
+            layout=widgets.Layout(
+                border="1px solid #ddd",
+                padding="8px",
+                height="520px",
+                overflow_y="auto",
+            ),
+        )
+        inp = widgets.Textarea(
+            placeholder="Search for satellite/aerial imagery (e.g., 'Find Sentinel-2 imagery over Paris in summer 2024')",
+            layout=widgets.Layout(width="99%", height="90px"),
+        )
+        btn_send = widgets.Button(
+            description="Search",
+            button_style="primary",
+            icon="search",
+            layout=widgets.Layout(width="120px"),
+        )
+        btn_stop = widgets.Button(
+            description="Stop", icon="stop", layout=widgets.Layout(width="120px")
+        )
+        btn_clear = widgets.Button(
+            description="Clear", icon="trash", layout=widgets.Layout(width="120px")
+        )
+        status = widgets.HTML(
+            "<link rel='stylesheet' href='https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.4.0/css/all.min.css'>"
+            "<span style='color:#666'>Ready to search.</span>"
+        )
+
+        examples = widgets.Dropdown(
+            options=[
+                ("— Example Searches —", ""),
+                (
+                    "Sentinel-2 over Las Vegas",
+                    "Find Sentinel-2 imagery over Las Vegas in August 2025",
+                ),
+                (
+                    "Landsat over Paris",
+                    "Find Landsat imagery over Paris from June to July 2025",
+                ),
+                (
+                    "Landsat with <10% cloud cover",
+                    "Find Landsat imagery over Paris with <10% cloud cover in June 2025",
+                ),
+                ("NAIP for NYC", "Show me NAIP aerial imagery for New York City"),
+                ("DEM for Seattle", "Show me DEM data for Seattle"),
+                (
+                    "3DEP Lidar HAG",
+                    "Show me data over Austin from 3dep-lidar-hag collection",
+                ),
+                ("ASTER for Tokyo", "Show me ASTER imagery for Tokyo"),
+            ],
+            value="",
+            layout=widgets.Layout(width="auto"),
+        )
+
+        # --- state kept on self so it persists ---
+        self._ui = SimpleNamespace(
+            messages=[],
+            map_panel=map_panel,
+            title=title,
+            log=log,
+            inp=inp,
+            btn_send=btn_send,
+            btn_stop=btn_stop,
+            btn_clear=btn_clear,
+            status=status,
+            examples=examples,
+        )
+        self._pending = {"fut": None}
+
+        def _esc(s: str) -> str:
+            """Escape HTML characters in a string."""
+            return (
+                s.replace("&", "&amp;")
+                .replace("<", "&lt;")
+                .replace(">", "&gt;")
+                .replace("\n", "<br/>")
+            )
+
+        def _append(role: str, msg: str) -> None:
+            """Append a message to the chat log."""
+            self._ui.messages.append((role, msg))
+            parts = []
+            for r, mm in self._ui.messages:
+                if r == "user":
+                    parts.append(
+                        f"<div style='margin:6px 0;padding:6px 8px;border-radius:8px;background:#eef;'><b>You</b>: {_esc(mm)}</div>"
+                    )
+                else:
+                    parts.append(
+                        f"<div style='margin:6px 0;padding:6px 8px;border-radius:8px;background:#f7f7f7;'><b>Agent</b>: {_esc(mm)}</div>"
+                    )
+            self._ui.log.value = (
+                "<div>"
+                + (
+                    "".join(parts)
+                    if parts
+                    else "<div style='color:#777'>No messages yet.</div>"
+                )
+                + "</div>"
+            )
+
+        def _lock(lock: bool) -> None:
+            """Lock or unlock UI controls."""
+            self._ui.btn_send.disabled = lock
+            self._ui.btn_stop.disabled = not lock
+            self._ui.btn_clear.disabled = lock
+            self._ui.inp.disabled = lock
+            self._ui.examples.disabled = lock
+
+        def _on_send(_: Any = None) -> None:
+            """Handle send button click or Enter key press."""
+            text = self._ui.inp.value.strip()
+            if not text:
+                return
+            _append("user", text)
+            _lock(True)
+            self._ui.status.value = "<span style='color:#0a7'>Searching…</span>"
+            try:
+                # Create a callback handler that updates the status widget
+                callback_handler = UICallbackHandler(status_widget=self._ui.status)
+
+                # Temporarily set callback_handler for this call
+                old_callback = self.callback_handler
+                self.callback_handler = callback_handler
+
+                # Get the structured search result directly (will show progress via callback)
+                item_data = self.search_and_get_first_item(text)
+
+                if item_data is not None:
+                    # Update status for visualization step
+                    self._ui.status.value = (
+                        "<span style='color:#0a7'>Adding to map...</span>"
+                    )
+
+                    # Visualize on map
+                    visualized_assets = self._visualize_stac_item(item_data)
+
+                    # Format response for display
+                    formatted_response = (
+                        f"Found item: {item_data['id']}\n"
+                        f"Collection: {item_data['collection']}\n"
+                        f"Date: {item_data.get('datetime', 'N/A')}\n"
+                    )
+
+                    if visualized_assets:
+                        assets_str = ", ".join(visualized_assets)
+                        formatted_response += f"✓ Added to map (assets: {assets_str})"
+                    else:
+                        formatted_response += "✓ Added to map"
+
+                    _append("assistant", formatted_response)
+                else:
+                    _append(
+                        "assistant",
+                        "No items found. Try adjusting your search query or date range.",
+                    )
+
+                self._ui.status.value = "<span style='color:#0a7'>Done.</span>"
+            except Exception as e:
+                _append("assistant", f"[error] {type(e).__name__}: {e}")
+                self._ui.status.value = (
+                    "<span style='color:#c00'>Finished with an issue.</span>"
+                )
+            finally:
+                # Restore old callback handler
+                self.callback_handler = old_callback
+                self._ui.inp.value = ""
+                _lock(False)
+
+        def _on_stop(_: Any = None) -> None:
+            """Handle stop button click."""
+            fut = self._pending.get("fut")
+            if fut and not fut.done():
+                self._pending["fut"] = None
+                self._ui.status.value = (
+                    "<span style='color:#c00'>Stop requested.</span>"
+                )
+                _lock(False)
+
+        def _on_clear(_: Any = None) -> None:
+            """Handle clear button click."""
+            self._ui.messages.clear()
+            self._ui.log.value = "<div style='color:#777'>No messages yet.</div>"
+            self._ui.status.value = "<span style='color:#666'>Cleared.</span>"
+
+        def _on_example_change(change: dict[str, Any]) -> None:
+            """Handle example dropdown selection change."""
+            if change["name"] == "value" and change["new"]:
+                self._ui.inp.value = change["new"]
+                self._ui.examples.value = ""
+                self._ui.inp.send({"method": "focus"})
+
+        # keep handler refs
+        self._handlers = SimpleNamespace(
+            on_send=_on_send,
+            on_stop=_on_stop,
+            on_clear=_on_clear,
+            on_example_change=_on_example_change,
+        )
+
+        # wire events
+        self._ui.btn_send.on_click(self._handlers.on_send)
+        self._ui.btn_stop.on_click(self._handlers.on_stop)
+        self._ui.btn_clear.on_click(self._handlers.on_clear)
+        self._ui.examples.observe(self._handlers.on_example_change, names="value")
+
+        # Ctrl+Enter on textarea
+        self._keyev = Event(
+            source=self._ui.inp, watched_events=["keyup"], prevent_default_action=False
+        )
+
+        def _on_key(e: dict[str, Any]) -> None:
+            """Handle keyboard events on the input textarea."""
+            if (
+                e.get("type") == "keyup"
+                and e.get("key") == "Enter"
+                and e.get("ctrlKey")
+            ):
+                if self._ui.inp.value.endswith("\n"):
+                    self._ui.inp.value = self._ui.inp.value[:-1]
+                self._handlers.on_send()
+
+        # store callback too
+        self._on_key_cb: Callable[[dict[str, Any]], None] = _on_key
+        self._keyev.on_dom_event(self._on_key_cb)
+
+        buttons = widgets.HBox(
+            [
+                self._ui.btn_send,
+                self._ui.btn_stop,
+                self._ui.btn_clear,
+                widgets.Box(
+                    [self._ui.examples], layout=widgets.Layout(margin="0 0 0 auto")
+                ),
+            ]
+        )
+        right = widgets.VBox(
+            [
+                self._ui.title,
+                self._ui.log,
+                self._ui.inp,
+                buttons,
+                self._ui.status,
+            ],
+            layout=widgets.Layout(flex="1 1 0%", min_width="360px"),
+        )
+        root = widgets.HBox(
+            [map_panel, right], layout=widgets.Layout(width="100%", gap="8px")
+        )
+        display(root)
+
+
+class CatalogAgent(Agent):
+    """AI agent for searching data catalogs with natural language queries."""
+
+    def __init__(
+        self,
+        *,
+        model: str = "llama3.1",
+        system_prompt: str = "default",
+        catalog_url: Optional[str] = None,
+        catalog_df: Optional[Any] = None,
+        model_args: dict = None,
+        **kwargs: Any,
+    ) -> None:
+        """Initialize the Catalog Agent.
+
+        Args:
+            model: Model identifier (default: "llama3.1").
+            system_prompt: System prompt for the agent (default: "default").
+            catalog_url: URL to a catalog file (TSV, CSV, or JSON). Use JSON format for spatial search support.
+                Example: "https://raw.githubusercontent.com/opengeos/Earth-Engine-Catalog/refs/heads/master/gee_catalog.json"
+            catalog_df: Pre-loaded catalog as a pandas DataFrame.
+            model_args: Additional keyword arguments for the model.
+            **kwargs: Additional keyword arguments for the Agent.
+        """
+        self.catalog_tools: CatalogTools = CatalogTools(
+            catalog_url=catalog_url, catalog_df=catalog_df
+        )
+
+        if model_args is None:
+            model_args = {}
+
+        # --- save a model factory we can call each turn ---
+        if model == "llama3.1":
+            self._model_factory: Callable[[], OllamaModel] = (
+                lambda: create_ollama_model(
+                    host="http://localhost:11434", model_id=model, **model_args
+                )
+            )
+        elif isinstance(model, str):
+            self._model_factory: Callable[[], BedrockModel] = (
+                lambda: create_bedrock_model(model_id=model, **model_args)
+            )
+        elif isinstance(model, OllamaModel):
+            # Extract configuration from existing OllamaModel and create new instances
+            model_id = model.config["model_id"]
+            host = model.host
+            client_args = model.client_args
+            self._model_factory: Callable[[], OllamaModel] = (
+                lambda: create_ollama_model(
+                    host=host, model_id=model_id, client_args=client_args, **model_args
+                )
+            )
+        elif isinstance(model, OpenAIModel):
+            # Extract configuration from existing OpenAIModel and create new instances
+            model_id = model.config["model_id"]
+            client_args = model.client_args.copy()
+            self._model_factory: Callable[[], OpenAIModel] = (
+                lambda mid=model_id, client_args=client_args: create_openai_model(
+                    model_id=mid, client_args=client_args, **model_args
+                )
+            )
+        elif isinstance(model, AnthropicModel):
+            # Extract configuration from existing AnthropicModel and create new instances
+            model_id = model.config["model_id"]
+            client_args = model.client_args.copy()
+            self._model_factory: Callable[[], AnthropicModel] = (
+                lambda mid=model_id, client_args=client_args: create_anthropic_model(
+                    model_id=mid, client_args=client_args, **model_args
+                )
+            )
+        else:
+            raise ValueError(f"Invalid model: {model}")
+
+        # build initial model (first turn)
+        model = self._model_factory()
+
+        if system_prompt == "default":
+            system_prompt = """You are a data catalog search agent. Your job is to help users find datasets from a data catalog.
+
+IMPORTANT: Follow these steps EXACTLY:
+
+1. Understand the user's query:
+   - What type of data are they looking for? (e.g., landcover, elevation, imagery)
+   - Are they searching for a specific geographic region? (e.g., California, San Francisco, bounding box)
+   - Are they filtering by time period? (e.g., "from 2020", "between 2015-2020", "recent data")
+   - Are they filtering by provider? (e.g., NASA, USGS)
+   - Are they filtering by dataset type? (e.g., image, image_collection, table)
+
+2. Use the appropriate tool:
+   - search_by_region: PREFERRED for spatial queries - search datasets covering a geographic region
+     * Use location parameter for place names (e.g., "California", "San Francisco")
+     * Use bbox parameter for coordinates [west, south, east, north]
+     * Can combine with keywords, dataset_type, provider, start_date, end_date filters
+   - search_datasets: For keyword-only searches without spatial filter
+     * Can filter by start_date and end_date for temporal queries
+   - geocode_location: Convert location names to coordinates (called automatically by search_by_region)
+   - get_dataset_info: Get details about a specific dataset by ID
+   - list_dataset_types: Show available dataset types
+   - list_providers: Show available data providers
+   - get_catalog_stats: Get overall catalog statistics
+
+3. Search strategy:
+   - SPATIAL QUERIES: If user mentions ANY location or region, IMMEDIATELY use search_by_region
+     * Pass location names directly to the location parameter - DO NOT ask user for bbox coordinates
+     * Examples of locations: California, San Francisco, New York, Paris, any city/state/country name
+     * search_by_region will automatically geocode location names - you don't need to call geocode_location separately
+   - TEMPORAL QUERIES: If user mentions ANY time period, ALWAYS add start_date/end_date parameters
+     * "from 2022" or "since 2022" or "2022 onwards" → start_date="2022-01-01"
+     * "until 2023" or "before 2023" → end_date="2023-12-31"
+     * "between 2020 and 2023" → start_date="2020-01-01", end_date="2023-12-31"
+     * "recent" or "latest" → start_date="2020-01-01"
+     * Time indicators: from, since, after, before, until, between, onwards, recent, latest
+   - KEYWORD QUERIES: If no location mentioned, use search_datasets
+   - Extract key search terms from the user's query
+   - Use keywords parameter for the main search terms
+   - Use dataset_type parameter if user specifies type (image, table, etc.)
+   - Use provider parameter if user specifies provider (NASA, USGS, etc.)
+   - Default max_results is 10, but can be adjusted
+
+CRITICAL RULES:
+1. NEVER ask the user to provide bbox coordinates. If they mention a location name, pass it directly to search_by_region(location="name")
+2. ALWAYS add start_date or end_date when user mentions ANY time period (from, since, onwards, recent, etc.)
+3. Convert years to YYYY-MM-DD format: 2022 → "2022-01-01"
+
+4. Examples:
+   - "Find landcover datasets covering California" → search_by_region(location="California", keywords="landcover")
+   - "Show elevation data for San Francisco" → search_by_region(location="San Francisco", keywords="elevation")
+   - "Find datasets in bbox [-122, 37, -121, 38]" → search_by_region(bbox=[-122, 37, -121, 38])
+   - "Find landcover datasets from NASA" → search_datasets(keywords="landcover", provider="NASA")
+   - "Show me elevation data" → search_datasets(keywords="elevation")
+   - "What types of datasets are available?" → list_dataset_types()
+   - "Find image collections about forests" → search_datasets(keywords="forest", dataset_type="image_collection")
+   - "Find landcover data from 2020 onwards" → search_datasets(keywords="landcover", start_date="2020-01-01")
+   - "Show California datasets between 2015 and 2020" → search_by_region(location="California", start_date="2015-01-01", end_date="2020-12-31")
+   - "Find recent elevation data" → search_datasets(keywords="elevation", start_date="2020-01-01")
+
+5. Return results clearly:
+   - Summarize the number of results found
+   - List the top results with their EXACT IDs and titles FROM THE TOOL RESPONSE
+   - Mention key information like provider, type, geographic coverage, and date range if available
+   - For spatial searches, mention the search region
+
+ERROR HANDLING:
+- If no results found: Suggest trying different keywords, broader region, or removing filters
+- If location not found: Suggest alternative spellings or try a broader region
+- If tool error: Explain the error and suggest alternatives
+
+CRITICAL RULES - MUST FOLLOW:
+1. NEVER make up or hallucinate dataset IDs, titles, or any other information
+2. ONLY report datasets that appear in the actual tool response
+3. Copy dataset IDs and titles EXACTLY as they appear in the tool response
+4. If a field is null/None in the tool response, say "N/A" or omit it - DO NOT guess
+5. DO NOT use your training data knowledge about Earth Engine datasets
+6. DO NOT fill in missing information from your knowledge
+7. If unsure, say "Information not available in results"
+
+Example of CORRECT behavior:
+Tool returns: {"id": "AAFC/ACI", "title": "Canada AAFC Annual Crop Inventory"}
+Your response: "Found dataset: AAFC/ACI - Canada AAFC Annual Crop Inventory"
+
+Example of INCORRECT behavior (DO NOT DO THIS):
+Tool returns: {"id": "AAFC/ACI", "title": "Canada AAFC Annual Crop Inventory"}
+Your response: "Found dataset: USGS/NED - USGS Elevation Data"  ← WRONG! This ID wasn't in the tool response!"""
+
+        super().__init__(
+            name="Catalog Search Agent",
+            model=model,
+            tools=[
+                self.catalog_tools.search_datasets,
+                self.catalog_tools.search_by_region,
+                self.catalog_tools.get_dataset_info,
+                self.catalog_tools.geocode_location,
+                self.catalog_tools.list_dataset_types,
+                self.catalog_tools.list_providers,
+                self.catalog_tools.get_catalog_stats,
+            ],
+            system_prompt=system_prompt,
+            callback_handler=None,
+        )
+
+    def ask(self, prompt: str) -> str:
+        """Send a single-turn prompt to the agent.
+
+        Args:
+            prompt: The text prompt to send to the agent.
+
+        Returns:
+            The agent's response as a string.
+        """
+        # Use strands' built-in __call__ method which now supports multiple calls
+        result = self(prompt)
+        return getattr(result, "final_text", str(result))
+
+    def search_datasets(
+        self,
+        keywords: Optional[str] = None,
+        dataset_type: Optional[str] = None,
+        provider: Optional[str] = None,
+        max_results: int = 10,
+    ) -> List[Dict[str, Any]]:
+        """Search for datasets and return structured results.
+
+        This method directly uses the CatalogTools without LLM inference for faster searches.
+
+        Args:
+            keywords: Keywords to search for.
+            dataset_type: Filter by dataset type.
+            provider: Filter by provider.
+            max_results: Maximum number of results to return.
+
+        Returns:
+            List of dataset dictionaries.
+
+        Example:
+            >>> agent = CatalogAgent(catalog_url="...")
+            >>> datasets = agent.search_datasets(keywords="landcover", provider="NASA")
+            >>> for ds in datasets:
+            ...     print(ds['id'], ds['title'])
+        """
+        result_json = self.catalog_tools.search_datasets(
+            keywords=keywords,
+            dataset_type=dataset_type,
+            provider=provider,
+            max_results=max_results,
+        )
+
+        result = json.loads(result_json)
+
+        if "error" in result:
+            print(f"Search error: {result['error']}")
+            return []
+
+        return result.get("datasets", [])