Update app.py

app.py

@@ -2142,28 +2142,86 @@ from typing import List, Dict
 # Transformers pipeline for lightweight local LLM (text2text)
 from transformers import pipeline
 import math
+# =====================================================================
+#   OBJECTIVE 7 — INSIGHT & RECOMMENDATION (LLM FIRST, RULE-BASED IF FAIL)
+# =====================================================================
+
+st.markdown("<h3 class='section-title'>OBJECTIVE 7 — Insight and Recommendation</h3>", unsafe_allow_html=True)
+
+
+# ============================================================
+# 1. LLM LOADER
+# ============================================================
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM
 
-st.markdown("<h3 class='section-title'>Objective 7 — Insight & Recommendation (LLM)</h3>", unsafe_allow_html=True)
-# --------------------
-# Utility: load small local model (flan-t5-small)
-# --------------------
 @st.cache_resource
-def load_local_model(model_name="google/flan-t5-small"):
-    # text2text pipeline works well for instruction-style prompts
-    return pipeline("text2text-generation", model=model_name, device_map="auto" if hasattr(__import__('torch'),'cuda') else None)
+def load_llm_model():
+    try:
+        tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-2b-it")
+        model = AutoModelForCausalLM.from_pretrained(
+            "google/gemma-2-2b-it",
+            torch_dtype=torch.float16,
+            device_map="auto"
+        )
+        return tokenizer, model
+    except:
+        return None, None
 
-# Try to load model once
-try:
-    llm_pipe = load_local_model()
-    local_llm_available = True
-except Exception as e:
-    llm_pipe = None
-    local_llm_available = False
-    st.warning("Local LLM not available or failed to load (will fallback to rule-based recommendations).")
 
-# --------------------
-# Insert your extract_agentic_insights_v5 function (kept faithful to your original)
-# --------------------
+def llm_generate_recommendation(insight_summary_text):
+    """
+    Try to generate JSON recommendation using LLM.
+    If LLM fails → return None, and rule-based will be used.
+    """
+    tok, mdl = load_llm_model()
+    if tok is None or mdl is None:
+        return None
+
+    try:
+        prompt = f"""
+You are an expert Industrial Safety Analyst AI.
+Below is an INSIGHT SUMMARY from a Safety Reporting System:
+
+---
+{insight_summary_text}
+---
+
+Generate 5 Recommended Actions and Risk Mitigation in clean JSON:
+
+{{
+ "recommendations": [
+    {{"point":"1","rec":"...","mit":"..."}},
+    {{"point":"2","rec":"...","mit":"..."}},
+    {{"point":"3","rec":"...","mit":"..."}}
+ ]
+}}
+"""
+
+        inputs = tok(prompt, return_tensors="pt").to(mdl.device)
+        out = mdl.generate(
+            **inputs,
+            max_new_tokens=380,
+            temperature=0.25,
+            do_sample=True
+        )
+        text = tok.decode(out[0], skip_special_tokens=True)
+
+        import re, json
+        json_match = re.search(r"\{[\s\S]*\}", text)
+        if not json_match:
+            return None
+
+        return json.loads(json_match.group(0))
+
+    except:
+        return None
+
+
+# ============================================================
+# 2. RULE-BASED ENGINE (YOUR ORIGINAL SCRIPT)
+# ============================================================
+
 def extract_agentic_insights_v5(df: pd.DataFrame):
     dev = {
         "lowest_ratio_9_locs": [],
@@ -2179,7 +2237,7 @@ def extract_agentic_insights_v5(df: pd.DataFrame):
         "obj6_top2_categories": [],
     }
 
-    # === 1. 9 locations with lowest finding-to-reporter ratio ===
+    # === 1. 9 lowest locations ===
     if {'nama_lokasi_full', 'creator_nid', 'created_at', 'kode_temuan'}.issubset(df.columns):
         calc = df[['nama_lokasi_full', 'creator_nid', 'created_at', 'kode_temuan']].copy()
         calc['created_at'] = pd.to_datetime(calc['created_at'], errors='coerce')
@@ -2195,7 +2253,7 @@ def extract_agentic_insights_v5(df: pd.DataFrame):
         lowest_9 = loc_avg.nsmallest(9)
         dev["lowest_ratio_9_locs"] = [(loc, round(ratio, 3)) for loc, ratio in lowest_9.items()]
 
-    # === 2a: Division — lowest finding-to-reporter ratio (Obj 3a) ===
+    # === 2a: Lowest division ratio ===
     if {'nama', 'creator_nid', 'created_at', 'kode_temuan'}.issubset(df.columns):
         calc = df[['nama', 'creator_nid', 'created_at', 'kode_temuan']].copy()
         calc['bulan'] = pd.to_datetime(calc['created_at']).dt.to_period('M')
@@ -2207,22 +2265,18 @@ def extract_agentic_insights_v5(df: pd.DataFrame):
         agg['ratio'] = agg['findings'] / agg['reporters']
         div_ratio = agg.groupby('nama')['ratio'].mean()
         if not div_ratio.empty:
-            name = div_ratio.idxmin()
-            val = round(div_ratio.min(), 2)
-            dev["obj3a_lowest_div"] = (name, val)
+            dev["obj3a_lowest_div"] = (div_ratio.idxmin(), round(div_ratio.min(), 2))
 
-    # === 2b: Executor — longest average resolution time (Obj 3b) ===
+    # === 2b: Slowest executor ===
     if 'days_to_close' in df.columns:
         valid = df[df['days_to_close'].notna() & (df['days_to_close'] >= 0)]
         exec_col = 'nama_pic' if 'nama_pic' in valid.columns else 'creator_name'
         if exec_col in valid.columns:
             lead = valid.groupby(exec_col)['days_to_close'].mean()
             if not lead.empty:
-                name = lead.idxmax()
-                val = round(lead.max(), 1)
-                dev["obj3b_slowest_executor"] = (name, val)
+                dev["obj3b_slowest_executor"] = (lead.idxmax(), round(lead.max(), 1))
 
-    # === 2c: Reporter — lowest reporting frequency (Obj 3c) ===
+    # === 2c: Lowest reporter ===
     if {'creator_name', 'created_at'}.issubset(df.columns):
         calc = df[['creator_name', 'created_at']].copy()
         calc['bulan'] = pd.to_datetime(calc['created_at']).dt.to_period('M')
@@ -2230,28 +2284,24 @@ def extract_agentic_insights_v5(df: pd.DataFrame):
         avg = monthly.groupby('creator_name')['count'].mean()
         avg = avg[avg > 0]
         if not avg.empty:
-            name = avg.idxmin()
-            val = round(avg.min(), 2)
-            dev["obj3c_lowest_reporter"] = (name, val)
+            dev["obj3c_lowest_reporter"] = (avg.idxmin(), round(avg.min(), 2))
 
-    # === 2d: Division — longest average resolution time (Obj 3d) ===
+    # === 2d: Slowest division ===
     if 'days_to_close' in df.columns and 'nama' in df.columns:
         valid = df[df['days_to_close'].notna() & (df['days_to_close'] >= 0)]
         if not valid.empty:
             lead = valid.groupby('nama')['days_to_close'].mean()
             if not lead.empty:
-                name = lead.idxmax()
-                val = round(lead.max(), 1)
-                dev["obj3d_slowest_div"] = (name, val)
+                dev["obj3d_slowest_div"] = (lead.idxmax(), round(lead.max(), 1))
 
-    # === 3. Composition of non-Positive findings ===
+    # === 3. Non-positive ===
     if 'temuan_kategori' in df.columns:
         cnt = df['temuan_kategori'].value_counts(normalize=True) * 100
         dev["obj4_unsafe_condition_pct"] = round(cnt.get("Unsafe Condition", 0), 1)
         dev["obj4_unsafe_action_pct"] = round(cnt.get("Unsafe Action", 0), 1)
         dev["obj4_near_miss_pct"] = round(cnt.get("Near Miss", 0), 1)
 
-    # === 4. Risk Quadrants (X=20 findings/month, Y=3 days avg lead time) ===
+    # === 4. Risk Quadrant ===
     X_LIMIT, Y_LIMIT = 20, 3
     if {'nama', 'created_at', 'days_to_close', 'kode_temuan'}.issubset(df.columns):
         calc = df.copy()
@@ -2261,198 +2311,37 @@ def extract_agentic_insights_v5(df: pd.DataFrame):
         avg_count = monthly_counts.groupby('nama')['kode_temuan'].mean().reset_index(name='Finding Count')
         leadtime = calc.groupby('nama')['days_to_close'].mean().reset_index(name='Avg Lead Time')
         mat = avg_count.merge(leadtime, on='nama', how='left').fillna(0)
+
         for _, r in mat.iterrows():
             if r['Finding Count'] >= X_LIMIT and r['Avg Lead Time'] >= Y_LIMIT:
                 dev["obj5_q1_divs"].append(r['nama'])
             elif r['Finding Count'] < X_LIMIT and r['Avg Lead Time'] >= Y_LIMIT:
                 dev["obj5_q2_divs"].append(r['nama'])
 
-    # === 5. Top 2 non-Positive categories (avg per month) ===
+    # === 5. Top categories ===
     if {'kategori', 'temuan_kategori', 'created_at'}.issubset(df.columns):
         nonpos = df[df['temuan_kategori'] != 'Positive']
         if not nonpos.empty:
             start = nonpos['created_at'].min().to_period('M')
             end = nonpos['created_at'].max().to_period('M')
             n_months = len(pd.period_range(start=start, end=end, freq='M'))
-            if n_months == 0:
-                n_months = 1
             cat_avg = (nonpos.groupby('kategori').size() / n_months).sort_values(ascending=False).head(2)
-            dev["obj6_top2_categories"] = [(cat, round(val, 1)) for cat, val in cat_avg.items()]
+            dev["obj6_top2_categories"] = [(cat, round(v, 1)) for cat, v in cat_avg.items()]
 
     return dev
 
-# --------------------
-# Sample fallback data (if user didn't provide)
-# --------------------
-def sample_dataframe():
-    sample = {
-        "nama_lokasi_full": ["Loc A","Loc B","Loc C","Loc D","Loc E","Loc F","Loc G","Loc H","Loc I","Loc J"],
-        "creator_nid": [1,2,3,4,5,6,7,8,9,10],
-        "created_at": pd.date_range(end=pd.Timestamp("now"), periods=10).astype(str),
-        "kode_temuan": list(range(10)),
-        "nama": ["Div1","Div1","Div2","Div2","Div3","Div3","Div4","Div4","Div5","Div5"],
-        "days_to_close": [2,5,10,1,20,3,4,6,2,8],
-        "creator_name": ["A","B","C","D","E","F","G","H","I","J"],
-        "temuan_kategori": ["Unsafe Condition","Positive","Unsafe Action","Near Miss","Positive","Unsafe Condition","Unsafe Action","Positive","Near Miss","Unsafe Condition"],
-        "kategori": ["Electrical","Mechanical","Electrical","Civil","Mechanical","Electrical","Civil","Mechanical","Civil","Electrical"]
-    }
-    return pd.DataFrame(sample)
 
-# --------------------
-# LLM prompt & call (local using flan T5 small via pipeline)
-# --------------------
-def generate_recs_with_local_llm(insight_text: str, dev: dict, n_items: int = 5) -> List[Dict]:
-    """
-    Return list of objects: [{"point":"1","rec":"...","mit":"..."} ...]
-    Uses local text2text pipeline.
-    """
-    if not local_llm_available or llm_pipe is None:
-        return None
+# ============================================================
+# 3. RUN INSIGHT ENGINE
+# ============================================================
 
-    # Build very clear prompt asking for JSON list
-    prompt = (
-        "You are an expert Safety Analytics advisor for a power utility. "
-        "Based on the structured analytics below, produce EXACTLY "
-        f"{n_items} items. For each item return a JSON object with keys: point, rec, mit. "
-        "rec = Recommended Action (concise, 1-2 sentences). "
-        "mit = Risk Mitigation Strategy (concise, 1-2 sentences). "
-        "Return only a JSON array (no extra text).\n\n"
-        "Structured analytics (python dict):\n"
-        f"{json.dumps(dev, ensure_ascii=False, indent=2)}\n\n"
-        "Insight summary (short):\n"
-        f"{insight_text}\n\n"
-        "Output example:\n"
-        '[{"point":"1","rec":"...","mit":"..."}, ..., {"point":"5","rec":"...","mit":"..."}]\n'
-    )
+dev = extract_agentic_insights_v5(df_filtered)
 
-    # use pipeline to generate
-    try:
-        gen = llm_pipe(prompt, max_new_tokens=400, do_sample=False)[0]["generated_text"]
-    except Exception as e:
-        st.error(f"Local LLM generation failed: {e}")
-        return None
-
-    # try parse JSON array inside output text
-    try:
-        start = gen.index('[')
-        end = gen.rindex(']') + 1
-        arr_text = gen[start:end]
-        obj = json.loads(arr_text)
-        # Normalize: ensure point fields exist; if not, assign numbers
-        for idx, item in enumerate(obj):
-            if 'point' not in item:
-                item['point'] = str(idx+1)
-        return obj
-    except Exception:
-        # fallback: try to parse lines heuristically
-        lines = [ln.strip() for ln in gen.splitlines() if ln.strip()]
-        items = []
-        num = 1
-        for ln in lines:
-            if len(items) >= n_items:
-                break
-            # naive split by '|' or '-' if present
-            if '|' in ln:
-                parts = [p.strip() for p in ln.split('|') if p.strip()]
-                if len(parts) >= 2:
-                    rec = parts[0]
-                    mit = parts[1] if len(parts) > 1 else ""
-                    items.append({"point": str(num), "rec": rec, "mit": mit})
-                    num += 1
-            else:
-                # fallback short capture
-                items.append({"point": str(num), "rec": ln[:200], "mit": ""})
-                num += 1
-        if items:
-            return items
-    return None
-
-# --------------------
-# Rule-based fallback generator (if LLM unavailable)
-# --------------------
-def rule_based_recs(dev: dict) -> List[Dict]:
-    recs = []
-    if dev["lowest_ratio_9_locs"]:
-        recs.append({
-            "point": "1",
-            "rec": "Launch Agency Activation Sprint across the identified low-ratio locations: weekly micro-inspection by Area PIC.",
-            "mit": "Enable QR-based checklists with automatic reminders; monitor ratio weekly and target improvement within 45 days."
-        })
-    parts_exist = any([dev["obj3a_lowest_div"], dev["obj3c_lowest_reporter"], dev["obj3d_slowest_div"], dev["obj3b_slowest_executor"]])
-    if parts_exist:
-        recs.append({
-            "point": "2",
-            "rec": "Deploy Agentic Capacity Dashboard to monitor reporting & resolution KPIs per division/individual.",
-            "mit": "Auto-trigger coaching alerts when deviation >20% from baseline; weekly manager reports."
-        })
-    uc, ua, nm = dev["obj4_unsafe_condition_pct"], dev["obj4_unsafe_action_pct"], dev["obj4_near_miss_pct"]
-    if uc + ua + nm > 0:
-        recs.append({
-            "point": "3",
-            "rec": "Enforce photo-based validation for Unsafe categories to improve classification fidelity.",
-            "mit": "Block submission without evidence and require mandatory justification for manual overrides."
-        })
-    if dev["obj5_q1_divs"] or dev["obj5_q2_divs"]:
-        recs.append({
-            "point": "4",
-            "rec": "Assign Rapid Response Teams for Q1 divisions and apply One-Finding-One-Day policy for Q2.",
-            "mit": "Auto-escalate to senior ops if division remains Q1/Q2 for >=2 months."
-        })
-    if dev["obj6_top2_categories"]:
-        c1, c2 = dev["obj6_top2_categories"]
-        recs.append({
-            "point": "5",
-            "rec": f"Form RCA Task Force for {c1[0]} and {c2[0]} with cross-functional owners.",
-            "mit": "Update SOP and tender templates to include mitigations based on historical findings."
-        })
-    # Ensure at least 5 items
-    i = 1
-    while len(recs) < 5:
-        recs.append({"point": str(len(recs)+1), "rec": "Operational review and monitoring.", "mit": "Periodic review & KPIs."})
-        i += 1
-    return recs[:5]
-
-# --------------------
-# Streamlit UI
-# --------------------
-st.markdown("<h3 class='section-title'>OBJECTIVE 7 — Insight and Recommendation</h3>", unsafe_allow_html=True)
-
-col1, col2 = st.columns([2,1])
-
-with col1:
-    st.subheader("Upload data (or use sample)")
-    uploaded = st.file_uploader("Upload CSV (must contain relevant columns)", type=["csv"])
-    if uploaded is not None:
-        try:
-            df = pd.read_csv(uploaded)
-            st.success("CSV loaded.")
-        except Exception as e:
-            st.error(f"Failed to read CSV: {e}")
-            df = sample_dataframe()
-            st.info("Using sample dataframe due to read error.")
-    else:
-        st.info("No file uploaded — using sample dataset.")
-        df = sample_dataframe()
-
-    st.markdown("### Preview data (first 5 rows)")
-    st.dataframe(df.head(), use_container_width=True)
-
-with col2:
-    st.subheader("LLM / Mode")
-    st.write("Local LLM (flan-t5-small) will be used if available.")
-    st.write(f"Local LLM available: {local_llm_available}")
-    st.button("Reload model", on_click=lambda: st.experimental_rerun())
-
-# --------------------
-# Compute insights
-# --------------------
-dev = extract_agentic_insights_v5(df)
-
-# Build insight_text exactly similar to your format
+# Build Insight Summary Text
 insight_lines = []
 if dev["lowest_ratio_9_locs"]:
     loc_list = ", ".join([f"<strong>{loc}</strong> ({ratio})" for loc, ratio in dev["lowest_ratio_9_locs"]])
-    insight_lines.append(f"1. Nine locations with the <em>lowest</em> finding-to-reporter ratio: {loc_list}.")
+    insight_lines.append(f"1. Nine locations with the lowest finding-to-reporter ratio: {loc_list}.")
 
 parts = []
 if dev["obj3a_lowest_div"]:
@@ -2460,104 +2349,98 @@ if dev["obj3a_lowest_div"]:
 if dev["obj3c_lowest_reporter"]:
     parts.append(f"reporter <strong>{dev['obj3c_lowest_reporter'][0]}</strong> ({dev['obj3c_lowest_reporter'][1]} findings/month)")
 if dev["obj3d_slowest_div"]:
-    parts.append(f"division <strong>{dev['obj3d_slowest_div'][0]}</strong> (avg. resolution: {dev['obj3d_slowest_div'][1]} days)")
+    parts.append(f"division <strong>{dev['obj3d_slowest_div'][0]}</strong> ({dev['obj3d_slowest_div'][1]} days)")
 if dev["obj3b_slowest_executor"]:
-    parts.append(f"executor <strong>{dev['obj3b_slowest_executor'][0]}</strong> (avg. resolution: {dev['obj3b_slowest_executor'][1]} days)")
+    parts.append(f"executor <strong>{dev['obj3b_slowest_executor'][0]}</strong> ({dev['obj3b_slowest_executor'][1]} days)")
 
 if parts:
-    insight_lines.append(
-        f"2. Agentic AI highlights operational imbalance: {'; '.join(parts)}. These patterns indicate uneven engagement and resolution capability."
-    )
+    insight_lines.append("2. Agentic AI detection flags uneven operational capacity: " + "; ".join(parts))
 
 uc, ua, nm = dev["obj4_unsafe_condition_pct"], dev["obj4_unsafe_action_pct"], dev["obj4_near_miss_pct"]
 if uc + ua + nm > 0:
     insight_lines.append(f"3. Non-Positive composition: Unsafe Condition ({uc}%), Unsafe Action ({ua}%), Near Miss ({nm}%).")
 
 if dev["obj5_q1_divs"] or dev["obj5_q2_divs"]:
-    q1 = ", ".join([f"<strong>{d}</strong>" for d in dev["obj5_q1_divs"]]) or "—"
-    q2 = ", ".join([f"<strong>{d}</strong>" for d in dev["obj5_q2_divs"]]) or "—"
-    insight_lines.append(f"4. Quadrant I high-risk divisions: {q1}. Quadrant II hidden-risk divisions: {q2}.")
+    q1 = ", ".join([f"<strong>{x}</strong>" for x in dev["obj5_q1_divs"][:5]])
+    q2 = ", ".join([f"<strong>{x}</strong>" for x in dev["obj5_q2_divs"][:5]])
+    insight_lines.append(f"4. High-risk divisions (QI): {q1 or '—'}. Hidden-risk (QII): {q2 or '—'}.")
 
 if dev["obj6_top2_categories"]:
     c1, c2 = dev["obj6_top2_categories"]
-    insight_lines.append(f"5. Top recurring categories: <strong>{c1[0]}</strong> ({c1[1]}/month) and <strong>{c2[0]}</strong> ({c2[1]}/month).")
+    insight_lines.append(f"5. Top recurring non-Positive categories: <strong>{c1[0]}</strong> ({c1[1]}/mo) & <strong>{c2[0]}</strong> ({c2[1]}/mo).")
+
+insight_text = "<br>".join(insight_lines)
+
+
+# ============================================================
+# 4. TRY LLM FIRST → ELSE FALLBACK RULE-BASED
+# ============================================================
+
+llm_json = llm_generate_recommendation(insight_text)
+
+if llm_json is not None and "recommendations" in llm_json:
+    recs = llm_json["recommendations"]
+else:
+    # FALLBACK RULE-BASED
+    recs = []
+    if dev["lowest_ratio_9_locs"]:
+        recs.append({"point":"1","rec":"Launch spot-inspection sprint at low-ratio locations.","mit":"Enable 3-min QR checklist + auto-reminder."})
 
-insight_text_html = "<br>".join(insight_lines)
-insight_text_plain = "\n".join([s.replace("<strong>", "").replace("</strong>", "").replace("<em>", "").replace("</em>", "") for s in insight_lines])
+    if parts:
+        recs.append({"point":"2","rec":"Activate capacity dashboard.","mit":"Trigger coaching alerts if deviation >20%."})
 
-# --------------------
-# Render Insight card
-# --------------------
+    if uc + ua + nm > 0:
+        recs.append({"point":"3","rec":"Enforce photo-based validation.","mit":"Block submission without evidence."})
+
+    if dev["obj5_q1_divs"] or dev["obj5_q2_divs"]:
+        recs.append({"point":"4","rec":"Assign safety crews to QI divisions.","mit":"Auto escalate if stuck 2 months."})
+
+    if dev["obj6_top2_categories"]:
+        c1, c2 = dev["obj6_top2_categories"]
+        recs.append({"point":"5","rec":f"Create RCA task force for {c1[0]} & {c2[0]}.","mit":"Update tender specs with required mitigations."})
+
+
+# ============================================================
+# 5. RENDERING (NO CHANGES)
+# ============================================================
+
+# Insight Summary Card
 st.markdown(
     f"""
-    <div class="card" style="
-        background-color: #f8f9fa;
-        border-left: 4px solid #003DA5;
-        padding: 16px;
-        margin-bottom: 20px;
-        border-radius: 4px;
-        box-shadow: 0 2px 4px rgba(0,0,0,0.05);
-    ">
-        <h4 style="margin-top: 0; color: #FF6B6B;">Insight Summary</h4>
-        <p style="margin-bottom: 0; line-height: 1.6; font-size: 0.98em;">{insight_text_html if insight_text_html else 'No insights (missing columns).'}</p>
+    <div class="card" style="background-color:#f8f9fa;border-left:4px solid #003DA5;padding:16px;margin-bottom:20px;border-radius:4px;box-shadow:0 2px 4px rgba(0,0,0,0.05);">
+        <h4 style="margin-top:0;color:#FF6B6B;">Insight Summary</h4>
+        <p style="margin-bottom:0;line-height:1.6;font-size:0.98em;">{insight_text}</p>
     </div>
     """,
     unsafe_allow_html=True
 )
 
-# --------------------
-# Generate Recs via LLM (or fallback)
-# --------------------
-st.markdown("## Recommendations & Risk Mitigation (generated)")
-with st.spinner("Generating recommendations..."):
-    recs = generate_recs_with_local_llm(insight_text_plain, dev, n_items=5) if local_llm_available else None
-
-if not recs:
-    st.warning("LLM not available or failed to parse — using fallback rule-based recommendations.")
-    recs = rule_based_recs(dev)
-
-# --------------------
-# Render Recommendation table (same style)
-# --------------------
+# Recommendation Table
 if recs:
-    rows_html = ""
-    for r in recs[:5]:
-        rows_html += (
-            f"<tr>"
-            f"<td style='text-align:center; font-weight:bold; width:5%;'>{r.get('point','')}</td>"
-            f"<td style='padding:8px;'>{r.get('rec','')}</td>"
-            f"<td style='padding:8px;'>{r.get('mit','')}</td>"
-            f"</tr>"
-        )
+    rows = "".join([
+        f"<tr><td style='text-align:center;font-weight:bold;width:5%;'>{r['point']}</td>"
+        f"<td style='padding:8px;'>{r['rec']}</td>"
+        f"<td style='padding:8px;'>{r['mit']}</td></tr>"
+        for r in recs
+    ])
 
-    table_html = f"""
-    <div class="card" style="
-        background-color: #e8f5e9;
-        border-left: 4px solid #4CAF50;
-        padding: 16px;
-        margin-bottom: 20px;
-        border-radius: 4px;
-        box-shadow: 0 2px 4px rgba(0,0,0,0.05);
-    ">
-        <h4 style="margin-top: 0; color: #2E7D32;">Recommended Actions & Agentic Risk Mitigation</h4>
-        <table style="width:100%; border-collapse:collapse; font-size:0.95em; margin-top:12px;">
-            <thead>
-                <tr style="background-color:#e8f5ee;">
-                    <th style="padding:10px; text-align:center; border:1px solid #ccc;">#</th>
-                    <th style="padding:10px; text-align:left; border:1px solid #ccc;">Recommended Action</th>
-                    <th style="padding:10px; text-align:left; border:1px solid #ccc;">Risk Mitigation Strategy</th>
-                </tr>
-            </thead>
-            <tbody>
-                {rows_html}
-            </tbody>
-        </table>
-    </div>
-    """
-    st.markdown(table_html, unsafe_allow_html=True)
+    st.markdown(
+        f"""
+        <div class="card" style="background-color:#e8f5e9;border-left:4px solid #4CAF50;padding:16px;margin-bottom:20px;border-radius:4px;box-shadow:0 2px 4px rgba(0,0,0,0.05);">
+            <h4 style="margin-top:0;color:#2E7D32;">Recommended Actions & Agentic Risk Mitigation</h4>
+            <table style="width:100%;border-collapse:collapse;font-size:0.95em;margin-top:12px;">
+                <thead>
+                    <tr style="background-color:#e8f5ee;">
+                        <th style="padding:10px;text-align:center;border:1px solid #ccc;">#</th>
+                        <th style="padding:10px;text-align:left;border:1px solid #ccc;">Recommended Action</th>
+                        <th style="padding:10px;text-align:left;border:1px solid #ccc;">Risk Mitigation Strategy</th>
+                    </tr>
+                </thead>
+                <tbody>{rows}</tbody>
+            </table>
+        </div>
+        """,
+        unsafe_allow_html=True
+    )
 else:
-    st.info("No recommendations available.")
-
-# --------------------
-# End
-# --------------------
+    st.info("No actionable insights generated.")