Update app.py

app.py

@@ -160,7 +160,7 @@ def stream_save_and_hash(upload_file: UploadFile, tmp_path: str, size_limit: Opt
 
 # ---------- AI interaction (blocking) ----------
 def generate_summary_blocking(meta, fiverow) -> str:
-    api_key = os.getenv("GEMINI_API_KEY")
+    api_key = "AIzaSyB1jgGCuzg7ELPwNEEwaluQZoZhxhgLmAs"
     if not api_key:
         raise RuntimeError("GEMINI_API_KEY not set")
     client = genai.Client(api_key=api_key)
@@ -171,9 +171,39 @@ Input contains:
 - meta: dataframe metadata
 - fiverow: first 5 records of dataframe
 You must output JSON with the following structure:
-{ "summary": "<short natural language overview>", "recommended_charts": [ ... ] }
-Always produce syntactically valid JSON ONLY.
-"""
+{
+  "summary": "<short natural language overview of dataset>",
+  "recommended_charts": [
+    {
+      "type": "<one of: bar, pie, timeseries, histogram, scatter, multiple_columns, stacked_bar, heatmap>",
+      "title": "<short title for chart>",
+      "columns": ["<col1>", "<col2>", "..."],
+      "python_code": "<full runnable Python code using seaborn/matplotlib that produces the chart>"
+    },
+    ...
+  ]
+}
+Mandatory rules:
+- Always produce syntactically valid JSON ONLY. No text outside the JSON object.
+- Provide at least these chart types somewhere in recommended_charts: bar, pie, timeseries, histogram, scatter, multiple_columns, stacked_bar, heatmap.
+- Use only column names that appear in meta['column_names'].
+- The python_code string must be self-contained and runnable assuming a variable `df` exists containing the full cleaned DataFrame. Start the code with imports:
+    import pandas as pd
+    import seaborn as sns
+    import matplotlib.pyplot as plt
+  and include any necessary preprocessing steps (e.g., parsing dates).
+- For timeseries charts ensure the datetime column is parsed (`pd.to_datetime`) before plotting.
+- For multiple_columns provide a pairplot or facetgrid example that uses up to 4 numeric columns or sensible categorical splits.
+- For stacked_bar, show aggregation code (groupby + unstack) and plotting with df.plot(kind='bar', stacked=True).
+- For heatmap, compute correlation matrix and plot sns.heatmap with annotations.
+- For pie charts, ensure grouping/aggregation when there are >20 unique categories (group small categories into 'Other').
+- For histogram and scatter include axis labels and tight_layout; include plt.show() at the end.
+- Keep code minimal but complete so a user can copy-paste and run (assume seaborn, matplotlib, pandas installed).
+- For each chart add a sensible "columns" list showing which columns the code uses.
+- Do not include examples using columns not present in meta.
+- Do not include more than 10 recommended_charts.
+- Ensure strings inside the JSON are escaped properly so the JSON parses.
+Produce concise natural-language one-line summary in "summary". Ensure JSON is parseable by json.loads in Python."""
     user_prompt = {"meta": meta, "fiverow": fiverow}
     contents = [
         types.Content(