fixes

conf/gaia_demo.yaml

@@ -2,7 +2,7 @@ defaults:
   - llm: gpt4o
   - _self_
 
-exp_name: gpt4o_demo2
+exp_name: gpt4o_demo_hotel2
 
 exp_path: outputs/gaia/runs/${exp_name}
 split: validation

conf/llm/gpt4o.yaml

@@ -3,4 +3,4 @@ model_name: chatgpt-4o-latest
 use_cache: false
 context_size: 128000
 parameters:
-  temperature: 0.0
+  temperature: 0.2

examples/gaia_agent/prompts.py

@@ -14,7 +14,7 @@
 Check if the action lead to desired outcome or not. Then explain its effect on the task and the plan.
 After that propose the best next step to do, according to the plan. Do not forget to mention the reasoning behind the next step.
 If you see the cookie consent form, accept it first.
-Quote the relevant part of the observation if possible.
+Quote the relevant part of the observation if the action depends on it, for example when interacting with the page.
 {FORMAT}"""
 
 VERIFY = f"""

examples/gaia_agent/scripts/ui.py

@@ -120,6 +120,8 @@ def render_step(step: Step) -> str:
             msg = f"No answer found:\n{step.overview}"
     elif step.kind in ["python_code_action", "search_action", "watch_video_action"]:
         msg = to_pretty_str(step.llm_dict())
+    elif step and step.kind == "scroll_action":
+        msg = "Scrolling..."
     elif isinstance(step, Action):
         msg = "Interacting with the browser..."
     elif step and step.kind == "search_results_observation":

tapeagents/tools/browser.py

@@ -174,7 +174,9 @@ class Browser(Multitool):
     observations: tuple[type[Observation], ...] = (PageObservation,)
     tab_actions: list[type[Action]] = [CloseTabAction, NewTabAction, TabFocusAction]
     axtree: bool = True
-    viewport_size: int = 64000
+    viewport_chars: int = 64000
+    vieport_height: int = 768
+    viewport_width: int = 1024
     timeout_ms: int = 30000
     headless: bool = True
     save_video: bool = False
@@ -226,7 +228,7 @@ def model_post_init(self, __context: Any):
             record_video_dir=self._record_video_dir if self.save_video else None,
             action_mapping=HighLevelActionSet(demo_mode="default").to_python_code,
             timeout=self.timeout_ms,
-            viewport={"width": 1024, "height": 768},
+            viewport={"width": self.viewport_width, "height": self.vieport_height},
             task_kwargs={"start_url": "about:blank"},
             **self.gym_kwargs,
         )  # type: ignore
@@ -325,10 +327,12 @@ def format_error(self, err: str) -> str:
     def scroll(self, direction: str) -> PageObservation:
         if direction == "down" and self._current_viewport < self._n_viewports:
             self._current_viewport += 1
+            self.run_browser_action(f"scroll(0, {self.vieport_height})")
         elif direction == "up" and self._current_viewport > 1:
             self._current_viewport -= 1
+            self.run_browser_action(f"scroll(0, -{self.vieport_height})")
         page = self._current_page[
-            self.viewport_size * (self._current_viewport - 1) : self.viewport_size * self._current_viewport
+            self.viewport_chars * (self._current_viewport - 1) : self.viewport_chars * self._current_viewport
         ]
         return PageObservation(text=page, current_page=self._current_viewport, total_pages=self._n_viewports)
 
@@ -394,10 +398,10 @@ def next_page(self) -> PageObservation:
 
     def get_viewport(self, content: str) -> str:
         self._current_page = content
-        self._n_viewports = len(self._current_page) // self.viewport_size + 1
+        self._n_viewports = len(self._current_page) // self.viewport_chars + 1
         self._current_viewport = 1
         return self._current_page[
-            self.viewport_size * (self._current_viewport - 1) : self.viewport_size * self._current_viewport
+            self.viewport_chars * (self._current_viewport - 1) : self.viewport_chars * self._current_viewport
         ]
 
 
@@ -415,7 +419,7 @@ def flatten_axtree(
     coord_decimals: int = 0,
     ignored_properties=IGNORED_AXTREE_PROPERTIES,
     ignore_navigation: bool = False,
-    hide_bid_if_invisible: bool = False,
+    hide_bid_if_invisible: bool = True,
     hide_all_children: bool = False,
     nodes_with_bid: list[str] = NODES_WITH_BID,
 ) -> str: