refactor: consolidate dataset scenarios by updating bq-search-and-insight and adding bq-insight-and-forecast

Author: omkargaikwad23

evals/dataset.json

@@ -1,30 +1,32 @@
 {
   "scenarios": [
     {
-      "id": "bq-search-catalog",
-      "starting_prompt": "Search for tables related to sales in project ext-test-bigquery-analytics.",
-      "conversation_plan": "Ask the agent to search for tables with 'sales' in the prompt.",
+      "id": "bq-search-and-insight",
+      "starting_prompt": "Find tables related to sales in project ext-test-bigquery-analytics.",
+      "conversation_plan": "First, ask the agent to find tables related to sales. Once it lists the tables (which should include 'sales_data' in 'evalbench_ci'), ask it to identify the top product by sales in that table.",
       "expected_trajectory": [
-        "search_catalog"
+        "search_catalog",
+        "ask_data_insights"
       ],
       "env": {
         "GOOGLE_CLOUD_PROJECT": "ext-test-bigquery-analytics"
       },
       "kind": "tools",
-      "max_turns": 3
+      "max_turns": 4
     },
     {
-      "id": "bq-ask-insights",
-      "starting_prompt": "What are the top products by sales in the table 'sales_data' in dataset 'sales' and project 'ext-test-bigquery-analytics'?",
-      "conversation_plan": "Ask the agent to get insights about the sales table.",
+      "id": "bq-insight-and-forecast",
+      "starting_prompt": "What are the top products by sales in the table 'sales_data' in dataset 'evalbench_ci' and project 'ext-test-bigquery-analytics'?",
+      "conversation_plan": "First, ask the agent to find the top products by sales in the sales_data table. After it identifies the top products, ask it to forecast the sales for the top product for the next 5 steps.",
       "expected_trajectory": [
-        "ask_data_insights"
+        "ask_data_insights",
+        "forecast"
       ],
       "env": {
         "GOOGLE_CLOUD_PROJECT": "ext-test-bigquery-analytics"
       },
       "kind": "tools",
-      "max_turns": 3
+      "max_turns": 4
     }
   ]
 }