Update app.py

app.py

@@ -3,51 +3,33 @@ import pandas as pd
 from transformers import pipeline
 
 # ------------------------------------------------
-# Load Qwen-3B (polishing only)
+# Load Qwen 3B
 # ------------------------------------------------
 generator = pipeline(
-    task="text-generation",
+    "text-generation",
     model="Qwen/Qwen2.5-3B-Instruct",
     device_map="auto",
     trust_remote_code=True
 )
 
 # ------------------------------------------------
-# Generic helpers (pure statistics → language)
+# Quantization helpers (FACTS only)
 # ------------------------------------------------
-def trend_word(diff):
-    return "increased" if diff > 0 else "decreased"
-
-def magnitude_word(abs_diff):
-    if abs_diff < 0.05:
-        return "marginally"
-    elif abs_diff < 0.2:
-        return "slightly"
+def magnitude_bucket(x):
+    if x < 0.05:
+        return "low"
+    elif x < 0.2:
+        return "medium"
     else:
-        return "noticeably"
+        return "high"
 
-def unit_word(kpi):
-    return "percentage points" if "%" in kpi else ""
-
-def summarize_secondary(diffs):
-    avg_change = diffs.abs().mean()
-    inc_ratio = (diffs > 0).mean()
-
-    if avg_change < 0.05:
-        magnitude = "minor variation"
-    elif avg_change < 0.2:
-        magnitude = "moderate movement"
-    else:
-        magnitude = "notable movement"
-
-    if inc_ratio > 0.7:
-        direction = "mostly increased"
-    elif inc_ratio < 0.3:
-        direction = "mostly decreased"
+def direction_bucket(diff):
+    if diff > 0:
+        return "increase"
+    elif diff < 0:
+        return "decrease"
     else:
-        direction = "mixed movement"
-
-    return f"Remaining indicators showed {magnitude} with {direction}"
+        return "no_change"
 
 # ------------------------------------------------
 # Core logic
@@ -55,69 +37,79 @@ def summarize_secondary(diffs):
 def analyze_kpi(csv_file, top_n):
     df = pd.read_csv(csv_file.name)
 
-    date_cols = df.columns[1:]
-    prev_date = date_cols[-2]
-    curr_date = date_cols[-1]
+    dates = df.columns[1:]
+    prev_date, curr_date = dates[-2], dates[-1]
 
-    df["Diff"] = df[curr_date] - df[prev_date]
-    df["Change"] = df["Diff"].abs()
+    df["diff"] = df[curr_date] - df[prev_date]
+    df["abs_diff"] = df["diff"].abs()
 
-    df_sorted = df.sort_values("Change", ascending=False)
-    top_kpis = df_sorted.head(top_n)
+    ranked = df.sort_values("abs_diff", ascending=False).head(top_n)
 
     # -------------------------------
-    # Primary KPI sentence
+    # Primary KPI facts
     # -------------------------------
-    primary = top_kpis.iloc[0]
-    abs_diff = abs(primary["Diff"])
-
-    primary_sentence = (
-        f"{primary['Kpi']} "
-        f"{trend_word(primary['Diff'])} "
-        f"{magnitude_word(abs_diff)} "
-        f"by {abs_diff:.2f} {unit_word(primary['Kpi'])}".strip()
-    )
+    primary = ranked.iloc[0]
+
+    primary_facts = {
+        "PRIMARY_KPI": primary["Kpi"],
+        "PRIMARY_DIRECTION": direction_bucket(primary["diff"]),
+        "PRIMARY_CHANGE": round(primary["abs_diff"], 2),
+        "PRIMARY_MAGNITUDE": magnitude_bucket(primary["abs_diff"]),
+        "PRIMARY_UNIT": "percentage points" if "%" in primary["Kpi"] else "units"
+    }
 
     # -------------------------------
-    # Secondary KPI sentence
+    # Secondary KPI facts
     # -------------------------------
-    secondary_diffs = top_kpis.iloc[1:]["Diff"]
-    secondary_sentence = summarize_secondary(secondary_diffs)
+    secondary = ranked.iloc[1:]
+
+    secondary_facts = {
+        "SECONDARY_COUNT": len(secondary),
+        "SECONDARY_AVG_CHANGE": round(secondary["abs_diff"].mean(), 2),
+        "SECONDARY_MAGNITUDE": magnitude_bucket(secondary["abs_diff"].mean()),
+        "SECONDARY_DIRECTION_BALANCE": (
+            "mostly_increase" if (secondary["diff"] > 0).mean() > 0.7
+            else "mostly_decrease" if (secondary["diff"] > 0).mean() < 0.3
+            else "mixed"
+        )
+    }
 
-    # Model input (already well-formed)
-    model_input = primary_sentence + ". " + secondary_sentence + "."
+    # -------------------------------
+    # Model input = FACT BLOCK
+    # -------------------------------
+    model_input = (
+        "Generate a short operational summary from the following facts.\n\n"
+        f"{primary_facts}\n"
+        f"{secondary_facts}"
+    )
 
-    # Qwen generation (polishing only)
     output = generator(
         model_input,
-        max_new_tokens=60,
+        max_new_tokens=80,
         do_sample=False
     )[0]["generated_text"]
 
-    return top_kpis[["Kpi", "Change"]], output
+    return ranked[["Kpi", "abs_diff"]], output
 
 # ------------------------------------------------
-# Gradio UI (HF Space)
+# UI
 # ------------------------------------------------
-with gr.Blocks(title="KPI Change Newsletter") as demo:
-    gr.Markdown("## 📰 KPI Change Newsletter Generator")
-    gr.Markdown(
-        "Upload a KPI CSV file to rank changes and generate "
-        "a short, data-driven summary."
-    )
+with gr.Blocks(title="KPI Summary Generator") as demo:
+    gr.Markdown("## KPI Change Summary")
+    gr.Markdown("Upload CSV. Summary is generated strictly from data-derived facts.")
 
-    csv_input = gr.File(label="Upload CSV", file_types=[".csv"])
-    top_n_input = gr.Slider(3, 5, value=3, step=1, label="Top KPIs")
+    csv_input = gr.File(file_types=[".csv"])
+    top_n_input = gr.Slider(3, 5, value=3, step=1)
 
-    generate_btn = gr.Button("Generate Summary")
+    btn = gr.Button("Generate")
 
-    table_output = gr.Dataframe(label="Top KPI Changes")
-    summary_output = gr.Textbox(label="Newsletter Summary", lines=2)
+    table = gr.Dataframe()
+    summary = gr.Textbox(lines=3)
 
-    generate_btn.click(
+    btn.click(
         analyze_kpi,
         inputs=[csv_input, top_n_input],
-        outputs=[table_output, summary_output]
+        outputs=[table, summary]
     )
 
 demo.launch()