agg narative country list

dags/etl_food_security.py

@@ -1,8 +1,57 @@
+"""
+AIRFLOW DAG — ETL Food Security BigQuery
+Kimball Data Warehouse Architecture
+
+Schedule : Setiap 3 hari sekali (timedelta(days=3))
+Catchup  : False
+
+Kimball ETL Flow:
+    ┌──────────────────────────────────────────────────────────────────────────┐
+    │  BRONZE (Raw)       SILVER (Staging→Cleaned)   GOLD (DW → Analytical)   │
+    │                                                                          │
+    │  raw_fao       ─┐                              dim_country               │
+    │  raw_worldbank ─┼→ staging_integrated          dim_indicator             │
+    │  raw_unicef    ─┘        ↓                     dim_time                  │
+    │                   cleaned_integrated  ───────→ dim_source                │
+    │                                                dim_pillar                │
+    │                                                fact_food_security        │
+    │                                                      ↓                   │
+    │                                                analytical_food_security  │
+    │                                                      ↓                   │
+    │                                                agg_pillar_composite      │
+    │                                                agg_pillar_by_country     │
+    │                                                agg_framework_by_country  │
+    │                                                agg_framework_asean       │
+    │                                                                          │
+    │  AUDIT : etl_logs, etl_metadata (setiap layer)                           │
+    └──────────────────────────────────────────────────────────────────────────┘
+
+Task Order:
+    verify_bigquery_connection
+        → load_fao_to_bronze
+        → load_worldbank_to_bronze
+        → load_unicef_to_bronze
+        → staging_integration_to_silver
+        → cleaned_integration_to_silver
+        → dimensional_model_to_gold
+        → analytical_layer_to_gold
+        → aggregation_to_gold
+
+Scripts folder harus berisi:
+    - bigquery_raw_layer.py             (run_verify_connection, run_load_fao, ...)
+    - bigquery_cleaned_layer.py         (run_cleaned_integration)
+    - bigquery_dimensional_model.py     (run_dimensional_model)
+    - bigquery_analytical_layer.py      (run_analytical_layer)
+    - bigquery_analysis_aggregation.py  (run_aggregation)
+    - bigquery_config.py
+    - bigquery_helpers.py
+    - bigquery_datasource.py
+"""
+
 from airflow import DAG
 from airflow.operators.python import PythonOperator
-from datetime import datetime, timedelta
+from datetime import datetime
 
-# Import fungsi dari folder scripts
 from scripts.bigquery_raw_layer import (
     run_verify_connection,
     run_load_fao,
@@ -10,32 +59,36 @@ from scripts.bigquery_raw_layer import (
     run_load_unicef,
     run_staging_integration,
 )
-
 from scripts.bigquery_cleaned_layer import (
     run_cleaned_integration,
 )
-
-from scripts.bigquery_dimesional_model import (
+from scripts.bigquery_dimensional_model import (
     run_dimensional_model,
 )
-
 from scripts.bigquery_analytical_layer import (
     run_analytical_layer,
 )
-
 from scripts.bigquery_aggregate_layer import (
     run_aggregation,
 )
 
+# DEFAULT ARGS
+
+default_args = {
+    'owner': 'data-engineering',
+    'email': ['d1041221004@student.untan.ac.id'],
+}
+
+# DAG DEFINITION
 
 with DAG(
     dag_id            = "etl_food_security_bigquery",
-    description       = "Kimball ETL: FAO, World Bank, UNICEF to BigQuery (Bronze to Silver to Gold)",
+    description       = "Kimball ETL: FAO, World Bank, UNICEF → BigQuery (Bronze → Silver → Gold)",
+    default_args      = default_args,
     start_date        = datetime(2026, 3, 1),
-    schedule_interval = "@daily",
-    schedule_interval = timedelta(days=3),
+    schedule_interval = "0 0 */3 * *",
     catchup           = False,
-    tags              = ["food-security", "bigquery", "kimball"]
+    tags              = ["food-security", "bigquery", "kimball"],
 ) as dag:
 
     task_verify = PythonOperator(
@@ -62,26 +115,26 @@ with DAG(
         task_id         = "staging_integration_to_silver",
         python_callable = run_staging_integration
     )
-    
+
     task_cleaned = PythonOperator(
         task_id         = "cleaned_integration_to_silver",
         python_callable = run_cleaned_integration
     )
-    
+
     task_dimensional = PythonOperator(
         task_id         = "dimensional_model_to_gold",
         python_callable = run_dimensional_model
     )
-    
+
     task_analytical = PythonOperator(
         task_id         = "analytical_layer_to_gold",
         python_callable = run_analytical_layer
     )
-    
+
     task_aggregation = PythonOperator(
         task_id         = "aggregation_to_gold",
         python_callable = run_aggregation
     )
-    
 
+    
     task_verify >> task_fao >> task_worldbank >> task_unicef >> task_staging >> task_cleaned >> task_dimensional >> task_analytical >> task_aggregation

scripts/bigquery_aggregate_layer.py

@@ -1,11 +1,13 @@
 """
 BIGQUERY ANALYSIS LAYER - FOOD SECURITY AGGREGATION
 Semua agregasi pakai norm_value dari _get_norm_value_df()
-FIXED: Hanya simpan 4 tabel ke fs_asean_gold (layer='gold'):
+UPDATED: Simpan 6 tabel ke fs_asean_gold (layer='gold'):
   - agg_pillar_composite
   - agg_pillar_by_country
   - agg_framework_by_country
   - agg_framework_asean
+  - agg_narrative_overview
+  - agg_narrative_pillar
 """
 
 import pandas as pd
@@ -24,7 +26,6 @@ from scripts.bigquery_helpers import (
     save_etl_metadata,
 )
 from google.cloud import bigquery
-from sklearn.preprocessing import MinMaxScaler
 
 
 # =============================================================================
@@ -87,12 +88,10 @@ def global_minmax(series: pd.Series, lo: float = 1.0, hi: float = 100.0) -> pd.S
     v_min, v_max = values.min(), values.max()
     if v_min == v_max:
         return pd.Series((lo + hi) / 2.0, index=series.index)
-    scaler = MinMaxScaler(feature_range=(lo, hi))
-    result = np.full(len(series), np.nan)
+    result  = np.full(len(series), np.nan)
     not_nan = series.notna()
-    result[not_nan.values] = scaler.fit_transform(
-        series[not_nan].values.reshape(-1, 1)
-    ).flatten()
+    raw     = series[not_nan].values
+    result[not_nan.values] = lo + (raw - v_min) / (v_max - v_min) * (hi - lo)
     return pd.Series(result, index=series.index)
 
 
@@ -132,6 +131,260 @@ def check_and_dedup(df: pd.DataFrame, key_cols: list, context: str = "", logger=
     return df
 
 
+# =============================================================================
+# NARRATIVE BUILDER FUNCTIONS (pure functions, easy to unit-test)
+# =============================================================================
+
+def _fmt_score(score) -> str:
+    """Format score to 2 decimal places."""
+    if score is None or (isinstance(score, float) and np.isnan(score)):
+        return "N/A"
+    return f"{score:.2f}"
+
+
+def _fmt_delta(delta) -> str:
+    """Format YoY delta with sign and 2 decimal places."""
+    if delta is None or (isinstance(delta, float) and np.isnan(delta)):
+        return "N/A"
+    sign = "+" if delta >= 0 else ""
+    return f"{sign}{delta:.2f}"
+
+
+def _build_overview_narrative(
+    year: int,
+    n_mdg: int,
+    n_sdg: int,
+    n_total_ind: int,
+    score: float,
+    yoy_val,
+    yoy_pct,
+    prev_year: int,
+    prev_score,
+    ranking_list: list,
+    most_improved_country,
+    most_improved_delta,
+    most_declined_country,
+    most_declined_delta,
+) -> str:
+    """
+    Compose a full English prose narrative for the Overview tab.
+
+    Narrative structure
+    -------------------
+    1. Indicator composition  (MDGs first, then SDGs)
+    2. ASEAN score + YoY
+    3. Country ranking
+    4. Most improved / declined country
+    """
+
+    # -- Sentence 1: indicator composition ----------------------------------
+    parts_ind = []
+    if n_mdg > 0:
+        parts_ind.append(f"{n_mdg} MDG indicator{'s' if n_mdg > 1 else ''}")
+    if n_sdg > 0:
+        parts_ind.append(f"{n_sdg} SDG indicator{'s' if n_sdg > 1 else ''}")
+
+    if parts_ind:
+        ind_detail = " and ".join(parts_ind)
+        sent1 = (
+            f"In {year}, the ASEAN food security assessment incorporated a total of "
+            f"{n_total_ind} indicator{'s' if n_total_ind != 1 else ''}, "
+            f"consisting of {ind_detail}."
+        )
+    else:
+        sent1 = (
+            f"In {year}, the ASEAN food security assessment incorporated "
+            f"{n_total_ind} indicator{'s' if n_total_ind != 1 else ''}."
+        )
+
+    # -- Sentence 2: ASEAN score + YoY -------------------------------------
+    if yoy_val is not None and prev_score is not None:
+        direction_word = "increasing" if yoy_val >= 0 else "decreasing"
+        pct_clause     = ""
+        if yoy_pct is not None:
+            abs_pct    = abs(yoy_pct)
+            trend_word = "improvement" if yoy_val >= 0 else "decline"
+            pct_clause = f", which represents a {abs_pct:.2f}% {trend_word} year-over-year"
+        sent2 = (
+            f"The ASEAN overall score (Total framework) reached {_fmt_score(score)}, "
+            f"{direction_word} by {abs(yoy_val):.2f} points compared to the previous year "
+            f"({_fmt_score(prev_score)} in {prev_year}){pct_clause}."
+        )
+    else:
+        sent2 = (
+            f"The ASEAN overall score (Total framework) reached {_fmt_score(score)} in {year}; "
+            f"no prior-year data is available for year-over-year comparison."
+        )
+
+    # -- Sentence 3: country ranking ----------------------------
+    sent3 = ""
+    if ranking_list:
+        first  = ranking_list[0]
+        last   = ranking_list[-1]
+        middle = ranking_list[1:-1]
+
+        if len(ranking_list) == 1:
+            sent3 = (
+                f"In terms of country performance, {first['country_name']} was the only "
+                f"country assessed, scoring {_fmt_score(first['score'])} in {year}."
+            )
+        elif len(ranking_list) == 2:
+            sent3 = (
+                f"In terms of country performance, {first['country_name']} led the region "
+                f"with a score of {_fmt_score(first['score'])}, while "
+                f"{last['country_name']} recorded the lowest score of "
+                f"{_fmt_score(last['score'])} in {year}."
+            )
+        else:
+            # Susun semua negara di tengah: "B (xx.xx), C (xx.xx), ..., and Y (xx.xx)"
+            middle_parts = [
+                f"{c['country_name']} ({_fmt_score(c['score'])})"
+                for c in middle
+            ]
+            if len(middle_parts) == 1:
+                middle_str = middle_parts[0]
+            else:
+                middle_str = ", ".join(middle_parts[:-1]) + f", and {middle_parts[-1]}"
+
+            sent3 = (
+                f"In terms of country performance, {first['country_name']} led the region "
+                f"with a score of {_fmt_score(first['score'])}, followed by {middle_str}. "
+                f"At the other end, {last['country_name']} recorded the lowest score "
+                f"of {_fmt_score(last['score'])} in {year}."
+            )
+
+    # -- Sentence 4: most improved / declined ------------------------------
+    sent4_parts = []
+    if most_improved_country and most_improved_delta is not None:
+        sent4_parts.append(
+            f"the most notable improvement was seen in {most_improved_country}, "
+            f"which gained {_fmt_delta(most_improved_delta)} points from the previous year"
+        )
+    if most_declined_country and most_declined_delta is not None:
+        if most_declined_delta < 0:
+            sent4_parts.append(
+                f"while {most_declined_country} experienced the largest decline "
+                f"of {_fmt_delta(most_declined_delta)} points"
+            )
+        else:
+            sent4_parts.append(
+                f"while {most_declined_country} recorded the smallest gain "
+                f"of {_fmt_delta(most_declined_delta)} points"
+            )
+
+    sent4 = ""
+    if sent4_parts:
+        sent4 = ", ".join(sent4_parts) + "."
+        sent4 = sent4[0].upper() + sent4[1:]
+
+    # -- Assemble ----------------------------------------------------------
+    return " ".join(s for s in [sent1, sent2, sent3, sent4] if s)
+
+
+def _build_pillar_narrative(
+    year: int,
+    pillar_name: str,
+    pillar_score: float,
+    rank_in_year: int,
+    n_pillars: int,
+    yoy_val,
+    top_country,
+    top_country_score,
+    bot_country,
+    bot_country_score,
+    strongest_pillar,
+    strongest_score,
+    weakest_pillar,
+    weakest_score,
+    most_improved_pillar,
+    most_improved_delta,
+    most_declined_pillar,
+    most_declined_delta,
+) -> str:
+    """
+    Compose a full English prose narrative for a single pillar in a given year.
+
+    Narrative structure
+    -------------------
+    1. Pillar score and rank
+    2. Strongest / weakest pillar context
+    3. Top / bottom country within this pillar
+    4. YoY movement for this pillar + biggest mover across all pillars
+    """
+
+    # -- Sentence 1: pillar overview ----------------------------------------
+    rank_suffix = {1: "st", 2: "nd", 3: "rd"}.get(rank_in_year, "th")
+    sent1 = (
+        f"In {year}, the {pillar_name} pillar scored {_fmt_score(pillar_score)}, "
+        f"ranking {rank_in_year}{rank_suffix} out of {n_pillars} pillars assessed across ASEAN."
+    )
+
+    # -- Sentence 2: strongest / weakest context ----------------------------
+    sent2 = ""
+    if strongest_pillar and weakest_pillar:
+        if strongest_pillar == pillar_name:
+            sent2 = (
+                f"This made {pillar_name} the strongest performing pillar in {year}, "
+                f"compared to the weakest pillar, {weakest_pillar}, "
+                f"which scored {_fmt_score(weakest_score)}."
+            )
+        elif weakest_pillar == pillar_name:
+            sent2 = (
+                f"This made {pillar_name} the weakest performing pillar in {year}, "
+                f"compared to the strongest pillar, {strongest_pillar}, "
+                f"which scored {_fmt_score(strongest_score)}."
+            )
+        else:
+            sent2 = (
+                f"Across all pillars in {year}, {strongest_pillar} was the strongest "
+                f"(score: {_fmt_score(strongest_score)}), while {weakest_pillar} "
+                f"was the weakest (score: {_fmt_score(weakest_score)})."
+            )
+
+    # -- Sentence 3: country top / bottom within this pillar ---------------
+    sent3 = ""
+    if top_country and bot_country:
+        if top_country != bot_country:
+            sent3 = (
+                f"Within the {pillar_name} pillar, {top_country} led with a score of "
+                f"{_fmt_score(top_country_score)}, while {bot_country} recorded the lowest "
+                f"score of {_fmt_score(bot_country_score)}."
+            )
+        else:
+            sent3 = (
+                f"Within the {pillar_name} pillar, {top_country} was the only country "
+                f"with available data, scoring {_fmt_score(top_country_score)}."
+            )
+
+    # -- Sentence 4: YoY movement -------------------------------------------
+    if yoy_val is not None:
+        direction_word = "improved" if yoy_val >= 0 else "declined"
+        sent4 = (
+            f"Compared to the previous year, the {pillar_name} pillar "
+            f"{direction_word} by {abs(yoy_val):.2f} points"
+        )
+    else:
+        sent4 = (
+            f"No prior-year data is available to calculate year-over-year change "
+            f"for the {pillar_name} pillar in {year}"
+        )
+
+    if most_improved_pillar and most_improved_delta is not None \
+            and most_declined_pillar and most_declined_delta is not None \
+            and most_improved_pillar != most_declined_pillar:
+        sent4 += (
+            f". Across all pillars, {most_improved_pillar} showed the greatest improvement "
+            f"({_fmt_delta(most_improved_delta)} pts), while {most_declined_pillar} "
+            f"recorded the largest decline ({_fmt_delta(most_declined_delta)} pts)"
+        )
+
+    sent4 += "."
+    sent4 = sent4[0].upper() + sent4[1:]
+
+    # -- Assemble ----------------------------------------------------------
+    return " ".join(s for s in [sent1, sent2, sent3, sent4] if s)
+
+
 # =============================================================================
 # MAIN CLASS
 # =============================================================================
@@ -148,6 +401,8 @@ class FoodSecurityAggregator:
             "agg_pillar_by_country":    {"rows_loaded": 0, "status": "pending", "start_time": None, "end_time": None},
             "agg_framework_by_country": {"rows_loaded": 0, "status": "pending", "start_time": None, "end_time": None},
             "agg_framework_asean":      {"rows_loaded": 0, "status": "pending", "start_time": None, "end_time": None},
+            "agg_narrative_overview":   {"rows_loaded": 0, "status": "pending", "start_time": None, "end_time": None},
+            "agg_narrative_pillar":     {"rows_loaded": 0, "status": "pending", "start_time": None, "end_time": None},
         }
 
         self.df   = None
@@ -274,11 +529,13 @@ class FoodSecurityAggregator:
                 norm_parts.append(grp)
                 continue
 
-            scaler = MinMaxScaler(feature_range=(0, 1))
+            raw    = grp.loc[valid_mask, "value"].values
+            v_min, v_max = raw.min(), raw.max()
             normed = np.full(len(grp), np.nan)
-            normed[valid_mask.values] = scaler.fit_transform(
-                grp.loc[valid_mask, ["value"]]
-            ).flatten()
+            if v_min == v_max:
+                normed[valid_mask.values] = 0.5
+            else:
+                normed[valid_mask.values] = (raw - v_min) / (v_max - v_min)
 
             if do_invert:
                 normed = np.where(np.isnan(normed), np.nan, 1.0 - normed)
@@ -664,6 +921,308 @@ class FoodSecurityAggregator:
         self._finalize(table_name, rows)
         return df
 
+    # =========================================================================
+    # STEP 6: agg_narrative_overview -> Gold  (NEW)
+    #
+    # Sumber data : df_framework_asean (framework='Total') + df_framework_by_country
+    # Granularity : 1 row per year
+    # Columns     : year, n_mdg_indicators, n_sdg_indicators, n_total_indicators,
+    #               asean_total_score, yoy_change, yoy_change_pct,
+    #               country_ranking_json, most_improved_country, most_improved_delta,
+    #               most_declined_country, most_declined_delta, narrative_overview
+    # =========================================================================
+
+    def calc_narrative_overview(
+        self,
+        df_framework_asean:      pd.DataFrame,
+        df_framework_by_country: pd.DataFrame,
+    ) -> pd.DataFrame:
+        table_name = "agg_narrative_overview"
+        self.load_metadata[table_name]["start_time"] = datetime.now()
+        self.logger.info("\n" + "=" * 70)
+        self.logger.info(f"STEP 6: {table_name} -> [Gold] fs_asean_gold")
+        self.logger.info("=" * 70)
+
+        # ASEAN-level Total framework rows only, sorted by year
+        # PENTING: filter framework='Total' dulu sebelum apapun
+        asean_total = (
+            df_framework_asean[df_framework_asean["framework"] == "Total"]
+            .sort_values("year")
+            .reset_index(drop=True)
+        )
+
+        # Buat lookup score per tahun untuk ambil prev_score yang akurat
+        # Tidak mengandalkan score - yoy_val karena floating point bisa drift
+        score_by_year = dict(zip(
+            asean_total["year"].astype(int),
+            asean_total["framework_score_1_100"].astype(float),
+        ))
+
+        # Country-level Total framework rows (ranking + YoY per country)
+        country_total = (
+            df_framework_by_country[df_framework_by_country["framework"] == "Total"]
+            .copy()
+        )
+
+        # Indicator counts per year per framework (self.df already has 'framework' column)
+        ind_year = self.df.drop_duplicates(subset=["indicator_id", "year", "framework"])
+
+        records = []
+
+        for _, row in asean_total.iterrows():
+            yr      = int(row["year"])
+            score   = float(row["framework_score_1_100"])
+            yoy     = row["year_over_year_change"]
+            yoy_val = float(yoy) if pd.notna(yoy) else None
+
+            # -- Indicator counts per framework for this year ---------------
+            yr_ind      = ind_year[ind_year["year"] == yr]
+            n_mdg       = int(yr_ind[yr_ind["framework"] == "MDGs"]["indicator_id"].nunique())
+            n_sdg       = int(yr_ind[yr_ind["framework"] == "SDGs"]["indicator_id"].nunique())
+            n_total_ind = int(yr_ind["indicator_id"].nunique())
+
+            # -- prev_score diambil langsung dari lookup, bukan score - yoy_val
+            # Ini memastikan nilai konsisten 100% dengan tabel agg_framework_asean
+            prev_score = score_by_year.get(yr - 1, None)
+
+            # -- YoY % -----------------------------------------------------
+            yoy_pct = (
+                (yoy_val / prev_score * 100)
+                if (yoy_val is not None and prev_score is not None and prev_score != 0)
+                else None
+            )
+
+            # -- Country ranking for this year -----------------------------
+            yr_country = (
+                country_total[country_total["year"] == yr]
+                .sort_values("rank_in_framework_year")
+                .reset_index(drop=True)
+            )
+
+            ranking_list = []
+            for _, cr in yr_country.iterrows():
+                cr_yoy = cr.get("year_over_year_change", None)
+                ranking_list.append({
+                    "rank":         int(cr["rank_in_framework_year"]),
+                    "country_name": str(cr["country_name"]),
+                    "score":        round(float(cr["framework_score_1_100"]), 2),
+                    "yoy_change":   round(float(cr_yoy), 2) if pd.notna(cr_yoy) else None,
+                })
+            country_ranking_json = json.dumps(ranking_list, ensure_ascii=False)
+
+            # -- Most improved / declined country --------------------------
+            yr_country_yoy = yr_country.dropna(subset=["year_over_year_change"])
+            if not yr_country_yoy.empty:
+                best_idx  = yr_country_yoy["year_over_year_change"].idxmax()
+                worst_idx = yr_country_yoy["year_over_year_change"].idxmin()
+                most_improved_country = str(yr_country_yoy.loc[best_idx,  "country_name"])
+                most_improved_delta   = round(float(yr_country_yoy.loc[best_idx,  "year_over_year_change"]), 2)
+                most_declined_country = str(yr_country_yoy.loc[worst_idx, "country_name"])
+                most_declined_delta   = round(float(yr_country_yoy.loc[worst_idx, "year_over_year_change"]), 2)
+            else:
+                most_improved_country = most_declined_country = None
+                most_improved_delta   = most_declined_delta   = None
+
+            # -- Build narrative -------------------------------------------
+            narrative = _build_overview_narrative(
+                year                  = yr,
+                n_mdg                 = n_mdg,
+                n_sdg                 = n_sdg,
+                n_total_ind           = n_total_ind,
+                score                 = score,
+                yoy_val               = yoy_val,
+                yoy_pct               = yoy_pct,
+                prev_year             = yr - 1,
+                prev_score            = prev_score,
+                ranking_list          = ranking_list,
+                most_improved_country = most_improved_country,
+                most_improved_delta   = most_improved_delta,
+                most_declined_country = most_declined_country,
+                most_declined_delta   = most_declined_delta,
+            )
+
+            records.append({
+                "year":                  yr,
+                "n_mdg_indicators":      n_mdg,
+                "n_sdg_indicators":      n_sdg,
+                "n_total_indicators":    n_total_ind,
+                "asean_total_score":     round(score, 2),
+                "yoy_change":            yoy_val,
+                "yoy_change_pct":        round(yoy_pct, 2) if yoy_pct is not None else None,
+                "country_ranking_json":  country_ranking_json,
+                "most_improved_country": most_improved_country,
+                "most_improved_delta":   most_improved_delta,
+                "most_declined_country": most_declined_country,
+                "most_declined_delta":   most_declined_delta,
+                "narrative_overview":    narrative,
+            })
+
+        df = pd.DataFrame(records)
+        df["year"]               = df["year"].astype(int)
+        df["n_mdg_indicators"]   = df["n_mdg_indicators"].astype(int)
+        df["n_sdg_indicators"]   = df["n_sdg_indicators"].astype(int)
+        df["n_total_indicators"] = df["n_total_indicators"].astype(int)
+        df["asean_total_score"]  = df["asean_total_score"].astype(float)
+        for col in ["yoy_change", "yoy_change_pct", "most_improved_delta", "most_declined_delta"]:
+            df[col] = pd.to_numeric(df[col], errors="coerce").astype(float)
+
+        schema = [
+            bigquery.SchemaField("year",                  "INTEGER", mode="REQUIRED"),
+            bigquery.SchemaField("n_mdg_indicators",      "INTEGER", mode="REQUIRED"),
+            bigquery.SchemaField("n_sdg_indicators",      "INTEGER", mode="REQUIRED"),
+            bigquery.SchemaField("n_total_indicators",    "INTEGER", mode="REQUIRED"),
+            bigquery.SchemaField("asean_total_score",     "FLOAT",   mode="REQUIRED"),
+            bigquery.SchemaField("yoy_change",            "FLOAT",   mode="NULLABLE"),
+            bigquery.SchemaField("yoy_change_pct",        "FLOAT",   mode="NULLABLE"),
+            bigquery.SchemaField("country_ranking_json",  "STRING",  mode="REQUIRED"),
+            bigquery.SchemaField("most_improved_country", "STRING",  mode="NULLABLE"),
+            bigquery.SchemaField("most_improved_delta",   "FLOAT",   mode="NULLABLE"),
+            bigquery.SchemaField("most_declined_country", "STRING",  mode="NULLABLE"),
+            bigquery.SchemaField("most_declined_delta",   "FLOAT",   mode="NULLABLE"),
+            bigquery.SchemaField("narrative_overview",    "STRING",  mode="REQUIRED"),
+        ]
+        rows = load_to_bigquery(
+            self.client, df, table_name, layer='gold',
+            write_disposition="WRITE_TRUNCATE", schema=schema,
+        )
+        self._finalize(table_name, rows)
+        return df
+
+    # =========================================================================
+    # STEP 7: agg_narrative_pillar -> Gold  (NEW)
+    #
+    # Sumber data : df_pillar_composite + df_pillar_by_country
+    # Granularity : 1 row per (year, pillar_id)
+    # Columns     : year, pillar_id, pillar_name, pillar_score, rank_in_year,
+    #               yoy_change, top_country, top_country_score,
+    #               bottom_country, bottom_country_score, narrative_pillar
+    # =========================================================================
+
+    def calc_narrative_pillar(
+        self,
+        df_pillar_composite:  pd.DataFrame,
+        df_pillar_by_country: pd.DataFrame,
+    ) -> pd.DataFrame:
+        table_name = "agg_narrative_pillar"
+        self.load_metadata[table_name]["start_time"] = datetime.now()
+        self.logger.info("\n" + "=" * 70)
+        self.logger.info(f"STEP 7: {table_name} -> [Gold] fs_asean_gold")
+        self.logger.info("=" * 70)
+
+        records = []
+        years   = sorted(df_pillar_composite["year"].unique())
+
+        for yr in years:
+            yr_pillars = (
+                df_pillar_composite[df_pillar_composite["year"] == yr]
+                .sort_values("rank_in_year")
+                .reset_index(drop=True)
+            )
+            yr_country_pillar = df_pillar_by_country[df_pillar_by_country["year"] == yr]
+
+            # Strongest / weakest pillar this year (for context sentence)
+            strongest_pillar = yr_pillars.iloc[0]  if len(yr_pillars) > 0 else None
+            weakest_pillar   = yr_pillars.iloc[-1] if len(yr_pillars) > 0 else None
+
+            # Biggest improvement / decline across all pillars this year
+            yr_pillars_yoy = yr_pillars.dropna(subset=["year_over_year_change"])
+            if not yr_pillars_yoy.empty:
+                best_p_idx  = yr_pillars_yoy["year_over_year_change"].idxmax()
+                worst_p_idx = yr_pillars_yoy["year_over_year_change"].idxmin()
+                most_improved_pillar = str(yr_pillars_yoy.loc[best_p_idx,  "pillar_name"])
+                most_improved_delta  = round(float(yr_pillars_yoy.loc[best_p_idx,  "year_over_year_change"]), 2)
+                most_declined_pillar = str(yr_pillars_yoy.loc[worst_p_idx, "pillar_name"])
+                most_declined_delta  = round(float(yr_pillars_yoy.loc[worst_p_idx, "year_over_year_change"]), 2)
+            else:
+                most_improved_pillar = most_declined_pillar = None
+                most_improved_delta  = most_declined_delta  = None
+
+            for _, prow in yr_pillars.iterrows():
+                p_id      = int(prow["pillar_id"])
+                p_name    = str(prow["pillar_name"])
+                p_score   = float(prow["pillar_score_1_100"])
+                p_rank    = int(prow["rank_in_year"])
+                p_yoy     = prow["year_over_year_change"]
+                p_yoy_val = float(p_yoy) if pd.notna(p_yoy) else None
+
+                # Top / bottom country within this pillar & year
+                p_country = (
+                    yr_country_pillar[yr_country_pillar["pillar_id"] == p_id]
+                    .sort_values("rank_in_pillar_year")
+                    .reset_index(drop=True)
+                )
+                if not p_country.empty:
+                    top_country       = str(p_country.iloc[0]["country_name"])
+                    top_country_score = round(float(p_country.iloc[0]["pillar_country_score_1_100"]), 2)
+                    bot_country       = str(p_country.iloc[-1]["country_name"])
+                    bot_country_score = round(float(p_country.iloc[-1]["pillar_country_score_1_100"]), 2)
+                else:
+                    top_country = bot_country = None
+                    top_country_score = bot_country_score = None
+
+                # -- Build narrative ---------------------------------------
+                narrative = _build_pillar_narrative(
+                    year                 = yr,
+                    pillar_name          = p_name,
+                    pillar_score         = p_score,
+                    rank_in_year         = p_rank,
+                    n_pillars            = len(yr_pillars),
+                    yoy_val              = p_yoy_val,
+                    top_country          = top_country,
+                    top_country_score    = top_country_score,
+                    bot_country          = bot_country,
+                    bot_country_score    = bot_country_score,
+                    strongest_pillar     = str(strongest_pillar["pillar_name"])          if strongest_pillar is not None else None,
+                    strongest_score      = round(float(strongest_pillar["pillar_score_1_100"]), 2) if strongest_pillar is not None else None,
+                    weakest_pillar       = str(weakest_pillar["pillar_name"])            if weakest_pillar   is not None else None,
+                    weakest_score        = round(float(weakest_pillar["pillar_score_1_100"]), 2)   if weakest_pillar   is not None else None,
+                    most_improved_pillar = most_improved_pillar,
+                    most_improved_delta  = most_improved_delta,
+                    most_declined_pillar = most_declined_pillar,
+                    most_declined_delta  = most_declined_delta,
+                )
+
+                records.append({
+                    "year":                 yr,
+                    "pillar_id":            p_id,
+                    "pillar_name":          p_name,
+                    "pillar_score":         round(p_score, 2),
+                    "rank_in_year":         p_rank,
+                    "yoy_change":           p_yoy_val,
+                    "top_country":          top_country,
+                    "top_country_score":    top_country_score,
+                    "bottom_country":       bot_country,
+                    "bottom_country_score": bot_country_score,
+                    "narrative_pillar":     narrative,
+                })
+
+        df = pd.DataFrame(records)
+        df["year"]         = df["year"].astype(int)
+        df["pillar_id"]    = df["pillar_id"].astype(int)
+        df["rank_in_year"] = df["rank_in_year"].astype(int)
+        for col in ["pillar_score", "yoy_change", "top_country_score", "bottom_country_score"]:
+            df[col] = pd.to_numeric(df[col], errors="coerce").astype(float)
+
+        schema = [
+            bigquery.SchemaField("year",                 "INTEGER", mode="REQUIRED"),
+            bigquery.SchemaField("pillar_id",            "INTEGER", mode="REQUIRED"),
+            bigquery.SchemaField("pillar_name",          "STRING",  mode="REQUIRED"),
+            bigquery.SchemaField("pillar_score",         "FLOAT",   mode="REQUIRED"),
+            bigquery.SchemaField("rank_in_year",         "INTEGER", mode="REQUIRED"),
+            bigquery.SchemaField("yoy_change",           "FLOAT",   mode="NULLABLE"),
+            bigquery.SchemaField("top_country",          "STRING",  mode="NULLABLE"),
+            bigquery.SchemaField("top_country_score",    "FLOAT",   mode="NULLABLE"),
+            bigquery.SchemaField("bottom_country",       "STRING",  mode="NULLABLE"),
+            bigquery.SchemaField("bottom_country_score", "FLOAT",   mode="NULLABLE"),
+            bigquery.SchemaField("narrative_pillar",     "STRING",  mode="REQUIRED"),
+        ]
+        rows = load_to_bigquery(
+            self.client, df, table_name, layer='gold',
+            write_disposition="WRITE_TRUNCATE", schema=schema,
+        )
+        self._finalize(table_name, rows)
+        return df
+
     # =========================================================================
     # HELPERS
     # =========================================================================
@@ -698,21 +1257,36 @@ class FoodSecurityAggregator:
         log_update(self.client, "DW", table_name, "full_load", 0, "failed", str(error))
 
     # =========================================================================
-    # RUN
+    # RUN — 6 tabel (4 lama + 2 narrative baru)
     # =========================================================================
 
     def run(self):
         start = datetime.now()
         self.logger.info("\n" + "=" * 70)
-        self.logger.info("FOOD SECURITY AGGREGATION v8.0 — 4 TABLES -> fs_asean_gold")
+        self.logger.info("FOOD SECURITY AGGREGATION v9.0 — 6 TABLES -> fs_asean_gold")
+        self.logger.info("  agg_pillar_composite    | agg_pillar_by_country")
+        self.logger.info("  agg_framework_by_country| agg_framework_asean")
+        self.logger.info("  agg_narrative_overview  | agg_narrative_pillar")
         self.logger.info("=" * 70)
 
         self.load_data()
         self._classify_indicators()
-        self.calc_pillar_composite()
-        self.calc_pillar_by_country()
-        self.calc_framework_by_country()
-        self.calc_framework_asean()
+
+        # -- 4 tabel lama (tidak ada perubahan) ----------------------------
+        df_pillar_composite     = self.calc_pillar_composite()
+        df_pillar_by_country    = self.calc_pillar_by_country()
+        df_framework_by_country = self.calc_framework_by_country()
+        df_framework_asean      = self.calc_framework_asean()
+
+        # -- 2 tabel narrative baru ----------------------------------------
+        self.calc_narrative_overview(
+            df_framework_asean      = df_framework_asean,
+            df_framework_by_country = df_framework_by_country,
+        )
+        self.calc_narrative_pillar(
+            df_pillar_composite  = df_pillar_composite,
+            df_pillar_by_country = df_pillar_by_country,
+        )
 
         duration   = (datetime.now() - start).total_seconds()
         total_rows = sum(m["rows_loaded"] for m in self.load_metadata.values())
@@ -735,6 +1309,7 @@ def run_aggregation():
     """
     Airflow task: Hitung semua agregasi dari analytical_food_security.
     Dipanggil setelah analytical_layer_to_gold selesai.
+    Menjalankan 6 tabel sekaligus: 4 agregasi + 2 narrative.
     """
     from scripts.bigquery_config import get_bigquery_client
     client = get_bigquery_client()
@@ -757,7 +1332,7 @@ if __name__ == "__main__":
         _sys.stderr = io.TextIOWrapper(_sys.stderr.buffer, encoding="utf-8", errors="replace")
 
     print("=" * 70)
-    print("FOOD SECURITY AGGREGATION v8.0 — 4 TABLES -> fs_asean_gold")
+    print("FOOD SECURITY AGGREGATION-> fs_asean_gold")
     print(f"  NORMALIZE_FRAMEWORKS_JOINTLY = {NORMALIZE_FRAMEWORKS_JOINTLY}")
     print("=" * 70)