Dynamic chart

evalbench/util/sessionmgr.py

@@ -41,11 +41,28 @@ def __init__(
         self.sessions = {}
         self.ttl = 10800
         self.lock = RWLock()
+        self.load_sessions_from_disk()
         logging.debug("Starting reaper...")
         reaper = Thread(target=self.reaper, args=[])
         reaper.daemon = True
         reaper.start()
 
+    def load_sessions_from_disk(self):
+        try:
+            if not os.path.exists(SESSION_RESOURCES_PATH):
+                return
+            for sid in os.listdir(SESSION_RESOURCES_PATH):
+                dir_path = os.path.join(SESSION_RESOURCES_PATH, sid)
+                if os.path.isdir(dir_path):
+                    mtime = os.path.getmtime(dir_path)
+                    logging.info(f"Loading session {sid} from disk with mtime {mtime}.")
+                    self.sessions[sid] = {
+                        "create_ts": mtime,
+                        "session_id": sid,
+                    }
+        except Exception as e:
+            logging.error(f"Error loading sessions from disk: {e}")
+
     def set_ttl(self, ttl):
         self.ttl = ttl
 
@@ -78,7 +95,10 @@ def prune_resource_files(self, session_id):
                 os.remove(file_path)
             for dir in dirs:
                 dir_path = os.path.join(root, dir)
-                os.rmdir(dir_path)
+                if os.path.islink(dir_path):
+                    os.unlink(dir_path)
+                else:
+                    os.rmdir(dir_path)
         os.rmdir(path)
 
     def create_session(self, session_id):

evalbench_service/k8s/evalbench.yaml

@@ -37,11 +37,11 @@ spec:
         name: evalbench-eval
         resources:
           requests:
-            cpu: "42"
-            memory: "168Gi"
+            cpu: "20"
+            memory: "80Gi"
           limits:
-            cpu: "42"
-            memory: "168Gi"
+            cpu: "20"
+            memory: "80Gi"
         securityContext:
           allowPrivilegeEscalation: true
           capabilities:

viewer/main.py

@@ -237,6 +237,19 @@ def on_load(e: me.LoadEvent):
     if job_id and job_id in directories:
         state.selected_directory = job_id
 
+    tab = me.query_params.get("tab")
+    eval1 = me.query_params.get("eval1")
+    eval2 = me.query_params.get("eval2")
+
+    if tab == "compare" and eval1 and eval2:
+        state.selected_main_tab = "Compare"
+        state.compare_tab_visible = True
+        state.compare_evals = json.dumps([eval1, eval2])
+        # Trigger the AI comparison
+        state.ai_comparison = compare_evals(eval1, eval2)
+
+
+
 
 
 def status_component():
@@ -2281,36 +2294,25 @@ def on_next_conversation(e: me.ClickEvent):
                             from trends import trends_component
                             state = me.state(State)
 
+                            def on_main_tab_change(e: me.ButtonToggleChangeEvent):
+                                st = me.state(State)
+                                st.selected_main_tab = e.value
+                                logging.info(f"Main tab changed to: {e.value}")
+
                             with me.box(style=me.Style(margin=me.Margin(bottom="12px"))):
-                                tabs = ["Status", "List", "Charts"]
+                                buttons = [
+                                    me.ButtonToggleButton(label="Status", value="Status"),
+                                    me.ButtonToggleButton(label="List", value="List"),
+                                    me.ButtonToggleButton(label="Charts", value="Charts"),
+                                ]
                                 if state.compare_tab_visible:
-                                    tabs.append("Compare")
-                                for tab in tabs:
-                                    is_active = state.selected_main_tab == tab
-                                    tab_text = tab
-                                    if tab == "Compare" and state.ai_comparison == "Comparing...":
-                                        tab_text += " (Loading...)"
-
-                                    click_handler = None
-                                    if tab == "Status": click_handler = on_status_tab_click
-                                    elif tab == "List": click_handler = on_list_tab_click
-                                    elif tab == "Charts": click_handler = on_charts_tab_click
-                                    elif tab == "Compare": click_handler = on_compare_tab_click
+                                    buttons.append(me.ButtonToggleButton(label="Compare", value="Compare"))
 
-                                    me.button(
-                                        tab_text,
-                                        on_click=click_handler,
-                                        style=me.Style(
-                                            padding=me.Padding.symmetric(vertical="6px", horizontal="12px"),
-                                            background="#1e293b" if is_active else "#f1f5f9",
-                                            color="#ffffff" if is_active else "#475569",
-                                            border_radius="4px",
-                                            cursor="pointer",
-                                            font_weight="600" if is_active else "500",
-                                            font_size="14px",
-                                            margin=me.Margin(right="8px")
-                                        ),
-                                    )
+                                me.button_toggle(
+                                    value=state.selected_main_tab,
+                                    buttons=buttons,
+                                    on_change=on_main_tab_change,
+                                )
 
                             if state.selected_main_tab == "List":
                                 try: