Prompts and card logics improved

app/ads1/budget_optimizer.py

@@ -36,16 +36,16 @@ def build_budget_optimizer_context(dfs: dict, campaign_name: str | None = None):
     total_conversions = dfs["campaigns"]["conversions"].fillna(0).sum()
     account_cpa = dfs["campaigns"]["cost"].fillna(0).sum() / total_conversions if total_conversions else 0
 
-    def action_for(row):
+    def action_type_for(row):
         if row["conversions"] == 0:
-            return "Reduce or cap budget until conversion tracking and intent quality are fixed."
+            return "reduce"
         if account_cpa and row["cpa"] <= account_cpa * 0.8:
-            return "Increase budget cautiously because CPA is better than the account average."
+            return "increase"
         if account_cpa and row["cpa"] >= account_cpa * 1.25:
-            return "Reduce budget or tighten bids because CPA is worse than the account average."
-        return "Hold budget and monitor because efficiency is near the account average."
+            return "reduce"
+        return "hold"
 
-    df["budget_action"] = df.apply(action_for, axis=1)
+    df["budget_action"] = df.apply(action_type_for, axis=1)
     df["segment"] = "Campaign"
     df = df.sort_values(["conv_per_cost", "conversions"], ascending=[False, False]).head(8)
 
@@ -64,7 +64,7 @@ def build_budget_optimizer_context(dfs: dict, campaign_name: str | None = None):
             kw = build_budget_features(kw)
             kw["name"] = "Keyword: " + kw["keyword"].astype(str)
             kw["segment"] = "Keyword"
-            kw["budget_action"] = kw.apply(action_for, axis=1)
+            kw["budget_action"] = kw.apply(action_type_for, axis=1)
             kw = kw.sort_values(["conv_per_cost", "conversions", "cost"], ascending=[False, False, False]).head(6)
             kw_cols = [col for col in keep_cols if col in kw.columns]
             action_rows = pd.concat([action_rows, kw[kw_cols]], ignore_index=True)
@@ -81,10 +81,40 @@ def build_budget_optimizer_prompt(context: dict) -> str:
 
     return (
         f"Write 3 to 5 bullet points of actionable budget optimization insights for {name}.\n"
-        "Use the budget_actions list only. Say where to increase, reduce, hold, or protect budget using cost, conversions, CPA, CPC, and conversion efficiency.\n"
-        "Use simple language. One clear budget action per bullet. Start each line with '- '. No intro sentence.\n\n"
+        "Use the budget_actions list only. Each bullet must mention the campaign or keyword name, the budget action, and the evidence.\n"
+        "Use simple language. One self-contained budget action per bullet. Start each line with '- '. No intro sentence. Do not quote JSON values by themselves.\n\n"
         f"Data (JSON):\n{payload}"
     )
+
+
+def rule_based_budget_actions(context: dict) -> str:
+    rows = context.get("budget_actions", [])
+    if not rows:
+        return "- Hold budget for this campaign because no usable budget rows were found; verify campaign and keyword data before changing spend."
+
+    bullets = []
+    for row in rows[:5]:
+        name = row.get("name", "this segment")
+        action = row.get("budget_action", "hold")
+        cost = row.get("cost", 0)
+        conversions = row.get("conversions", 0)
+        cpa = row.get("cpa", 0)
+        cpc = row.get("cpc", 0)
+        conv_per_cost = row.get("conv_per_cost", 0)
+
+        if action == "increase":
+            bullets.append(
+                f"- Increase budget cautiously on {name} because it has {conversions} conversions at CPA {cpa:.2f}, CPC {cpc:.2f}, and conversion efficiency {conv_per_cost:.3f} on {cost:.2f} spend."
+            )
+        elif action == "reduce":
+            bullets.append(
+                f"- Reduce or cap budget on {name} because it has {conversions} conversions at CPA {cpa:.2f} after {cost:.2f} spend, making it a weaker use of budget."
+            )
+        else:
+            bullets.append(
+                f"- Hold budget on {name} because performance is near benchmark with {conversions} conversions, CPA {cpa:.2f}, and CPC {cpc:.2f}."
+            )
+    return "\n\n".join(bullets)
     
 
 def run_budget_optimizer(dfs: dict, campaign_name: str | None = None) -> str:
@@ -103,12 +133,9 @@ def run_budget_optimizer(dfs: dict, campaign_name: str | None = None) -> str:
 
     result = generate_explanation(prompt)
 
-    if is_bad_llm_output(result):
+    if is_bad_llm_output(result) or not result.strip().startswith("-") or result.count('"') >= 4:
         print("⚠️ [budget_optimizer] fallback triggered", flush=True)
-        return (
-            "- Unable to generate budget recommendations right now.\n"
-            "- Try again or check campaign data quality."
-        )
+        return rule_based_budget_actions(context)
 
     print("📤 [budget_optimizer] result received", flush=True)
     return result

app/ads1/growth_finder.py

@@ -46,13 +46,7 @@ def build_growth_finder_context(dfs: dict, campaign_name: str | None = None):
         ascending=[False, False, False],
     ).head(8)
 
-    df["growth_action"] = df.apply(
-        lambda row: (
-            f"Scale '{row['keyword']}' with a cautious bid or budget increase because it has "
-            f"{int(row['conversions'])} conversions at CPA {row['cpa']:.2f}."
-        ),
-        axis=1,
-    )
+    df["growth_action"] = "scale"
 
     keep_cols = [
         col
@@ -73,11 +67,34 @@ def build_growth_finder_prompt(context: dict) -> str:
     return (
         f"Write 3 to 5 bullet points of actionable growth opportunities for {name}.\n"
         "Use the growth_candidates list only. Suggest ways to scale winners, expand related intent, or increase budget on efficient areas.\n"
-        "Do not list weak keywords. Do not diagnose poor performance. Every bullet must include a growth action.\n"
-        "Use simple language. One growth opportunity per bullet. Start each line with '- '. No intro sentence.\n\n"
+        "Each bullet must mention the keyword, the growth action, and the evidence. Do not list weak keywords or diagnose poor performance.\n"
+        "Use simple language. One self-contained growth opportunity per bullet. Start each line with '- '. No intro sentence. Do not quote JSON values by themselves.\n\n"
         f"Data (JSON):\n{payload}"
     )
 
+
+def rule_based_growth_actions(context: dict) -> str:
+    rows = context.get("growth_candidates", [])
+    if not rows:
+        account_cpa = context.get("account_average_cpa", 0)
+        campaign_name = context.get("campaign_name", "this campaign")
+        return (
+            f"- No safe scale-up candidate found for {campaign_name} because its converting keywords are not efficient enough versus the account average CPA of {account_cpa}.\n"
+            "- Improve growth readiness first by cutting weak intent, tightening match types, and waiting for lower CPA before increasing budget."
+        )
+
+    bullets = []
+    for row in rows[:5]:
+        keyword = row.get("keyword", "this keyword")
+        conversions = row.get("conversions", 0)
+        cpa = row.get("cpa", 0)
+        cvr = row.get("cvr", 0)
+        ctr = row.get("ctr", 0)
+        bullets.append(
+            f"- Scale '{keyword}' because it has {conversions} conversions at CPA {cpa:.2f}, CVR {cvr:.2f}%, and CTR {ctr:.2f}%; test a cautious bid or budget increase."
+        )
+    return "\n\n".join(bullets)
+
 def run_growth_finder(dfs: dict, campaign_name: str | None = None) -> str:
     print("\n🚀 [growth_finder] STARTED", flush=True)
 
@@ -86,11 +103,7 @@ def run_growth_finder(dfs: dict, campaign_name: str | None = None) -> str:
 
     context = build_growth_finder_context(dfs, campaign_name)
     if not context.get("growth_candidates"):
-        account_cpa = context.get("account_average_cpa", 0)
-        return (
-            f"- No safe scale-up candidate found for {campaign_name} because its converting keywords are not efficient enough versus the account average CPA of {account_cpa}.\n"
-            "- Improve growth readiness first by cutting weak intent, tightening match types, and waiting for lower CPA before increasing budget."
-        )
+        return rule_based_growth_actions(context)
 
     print("🧠 [growth_finder] context built", flush=True)
 
@@ -100,12 +113,9 @@ def run_growth_finder(dfs: dict, campaign_name: str | None = None) -> str:
 
     result = generate_explanation(prompt)
 
-    if is_bad_llm_output(result):
+    if is_bad_llm_output(result) or not result.strip().startswith("-") or result.count('"') >= 4:
         print("⚠️ [growth_finder] fallback triggered", flush=True)
-        return (
-            "- Unable to generate scaling opportunities right now.\n"
-            "- Try again or check data quality."
-        )
+        return rule_based_growth_actions(context)
 
     print("📤 [growth_finder] result received", flush=True)
     return result

app/ads1/search_term_optimizer.py

@@ -34,7 +34,9 @@ def prepare_context_df(df: pd.DataFrame, max_rows: int = 200) -> pd.DataFrame:
     Instead of semantic filtering, we just cap size for token control.
     Keeps high-variance distribution intact.
     """
-    return df.sort_values("cost", ascending=False).head(max_rows)
+    waste = df[df["conversions"] == 0].sort_values("cost", ascending=False).head(10)
+    converters = df[df["conversions"] > 0].sort_values(["conversions", "cpa"], ascending=[False, True]).head(10)
+    return pd.concat([waste, converters], ignore_index=True).drop_duplicates(subset=["search_term"]).head(max_rows)
 
 # -------------------------
 # Prompt (LLM owns all reasoning now)
@@ -65,8 +67,8 @@ def build_search_optimizer_prompt(context: dict) -> str:
 
     return (
         f"Write 3 to 5 bullet points of actionable search term cleanup insights for {name}.\n"
-        "Identify terms to pause, add as negatives, add as keywords, scale, or investigate using cost, clicks, conversions, CPA, and CVR.\n"
-        "Use simple language. One search term action per bullet. Start each line with '- '. No intro sentence.\n\n"
+        "Use the search_terms list only. total_cost is total spend for that search term; cpc is cost per click.\n"
+        "Each bullet must mention the search term, the action, and the evidence. Use simple language. Start each line with '- '. No intro sentence.\n\n"
         f"Data (JSON):\n{payload}"
     )
 
@@ -82,13 +84,47 @@ def build_search_optimizer_context(dfs: dict, campaign_name: str | None = None):
 
     df = build_search_term_features(df)
     df = prepare_context_df(df)
+    df["action_type"] = df.apply(
+        lambda row: "add as keyword or scale" if row["conversions"] > 0 else "pause or add as negative",
+        axis=1,
+    )
+    df = df.rename(columns={"cost": "total_cost"})
+    keep_cols = [
+        col
+        for col in ["search_term", "action_type", "total_cost", "clicks", "impressions", "conversions", "ctr", "cvr", "cpc", "cpa"]
+        if col in df.columns
+    ]
 
     return {
         "campaign_name": campaign_name,
-        "search_terms": df.to_dict("records")  # FULL dataset context (bounded)
+        "search_terms": df[keep_cols].round(2).to_dict("records")
     }
 
 
+def rule_based_search_actions(context: dict) -> str:
+    rows = context.get("search_terms", [])
+    if not rows:
+        return "- No search term cleanup action found because no usable search term rows were available."
+
+    bullets = []
+    for row in rows[:5]:
+        term = row.get("search_term", "this search term")
+        total_cost = row.get("total_cost", 0)
+        clicks = row.get("clicks", 0)
+        conversions = row.get("conversions", 0)
+        cpa = row.get("cpa", 0)
+        cvr = row.get("cvr", 0)
+        if conversions > 0:
+            bullets.append(
+                f"- Add or scale '{term}' because it produced {conversions} conversions from {clicks} clicks at CPA {cpa:.2f} and CVR {cvr:.2f}% on {total_cost:.2f} total spend."
+            )
+        else:
+            bullets.append(
+                f"- Pause or add '{term}' as a negative because it spent {total_cost:.2f} across {clicks} clicks with 0 conversions."
+            )
+    return "\n\n".join(bullets)
+
+
 # -------------------------
 # Runner
 # -------------------------
@@ -108,12 +144,9 @@ def run_search_term_optimizer(dfs: dict, campaign_name: str | None = None) -> st
 
     result = generate_explanation(prompt)
 
-    if is_bad_llm_output(result):
+    if is_bad_llm_output(result) or not result.strip().startswith("-") or "cost is" in result.lower():
         print("⚠️ [search_term_optimizer] LLM fallback triggered", flush=True)
-        return (
-            "- Unable to generate insights right now.\n"
-            "- Try again or check data quality."
-        )
+        return rule_based_search_actions(context)
 
     print("📤 [search_term_optimizer] result received", flush=True)
     return result