Update app.py

app.py

@@ -1987,158 +1987,328 @@ else:
     st.info("No data available for non-positive issue categories with 100% coverage and positive trend.")
 # =================== OBJECTIVE 7 — Insight and Recommendation (Agentic AI LLM Style — Final) ===================
 # =================== OBJECTIVE 7 — Insight and Recommendation (Final — Agentic AI, No markdown bold) ===================
-# === RENDER INSIGHT CARD (unchanged logic, improved styling & center header) ===
-st.markdown(
-    f"""
-    <div class="card" style="
-        background-color: #f8f9fa;
-        border-left: 4px solid #003DA5;
-        padding: 18px;
-        margin-bottom: 24px;
-        border-radius: 6px;
-        box-shadow: 0 3px 6px rgba(0,0,0,0.06);
-    ">
-        <h4 style="margin-top: 0; color: #003DA5; text-align: center;">🔍 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 RECOMMENDATION & MITIGATION VIA LLM (DeepSeek-7B) ===
-@st.cache_resource
-def get_pipe():
-    return pipeline(
-        "text-generation",
-        model="deepseek-ai/deepseek-llm-7b-chat",
-        device_map="auto",
-        torch_dtype="auto",
-        trust_remote_code=True
-    )
+import streamlit as st
+import pandas as pd
+import re
+import os
 
-pipe = get_pipe()
+# ==============================
+# 1. IMPORT & INSTALL CHECK
+# ==============================
+try:
+    from transformers import pipeline
+except ImportError:
+    st.error("❌ `transformers` not installed. Run: `pip install transformers torch accelerate sentencepiece einops`")
+    st.stop()
 
-def safe_llm_call(insight: str, mode: str = "rec"):  # mode: 'rec' or 'mit'
-    sys_msg = (
-        "You are PLN's Lead Safety AI. Output ONLY a short, professional sentence. "
-        "Be directive, system-aware, and action-oriented. No markdown, no emoticons."
-    )
-    if mode == "rec":
-        sys_msg += " Focus on high-leverage intervention: 'Deploy...', 'Establish...', 'Launch...'"
-    else:
-        sys_msg += " Focus on automated/systemic control: 'Enforce...', 'Trigger...', 'Block...'"
-    
-    prompt = f"""<｜begin▁of▁sentence｜><｜User｜>Insight: {insight}
+# ==============================
+# 2. LOAD LLM (Phi-3-mini — ringan & stabil)
+# ==============================
+@st.cache_resource
+def load_llm():
+    try:
+        st.info("🧠 Loading Phi-3-mini-4k-instruct (optimized for safety recommendations)...")
+        pipe = pipeline(
+            "text-generation",
+            model="microsoft/Phi-3-mini-4k-instruct",
+            device_map="auto",
+            torch_dtype="auto",
+            trust_remote_code=True,
+            max_new_tokens=256
+        )
+        st.success("✅ Phi-3-mini loaded!")
+        return pipe
+    except Exception as e:
+        st.error(f"❌ Failed to load model: {e}")
+        st.stop()
+
+pipe = load_llm()
+
+# ==============================
+# 3. INSIGHT EXTRACTION (sama seperti kode Anda)
+# ==============================
+def extract_agentic_insights_v5(df: pd.DataFrame):
+    dev = {
+        "lowest_ratio_9_locs": [],
+        "obj3a_lowest_div": None,
+        "obj3b_slowest_executor": None,
+        "obj3c_lowest_reporter": None,
+        "obj3d_slowest_div": None,
+        "obj4_unsafe_condition_pct": 0.0,
+        "obj4_unsafe_action_pct": 0.0,
+        "obj4_near_miss_pct": 0.0,
+        "obj5_q1_divs": [],
+        "obj5_q2_divs": [],
+        "obj6_top2_categories": [],
+    }
 
-{sys_msg}
-<｜Assistant｜>"""
+    # === 1. 9 locations with lowest finding-to-reporter ratio ===
+    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')
+        calc = calc.dropna(subset=['created_at', 'nama_lokasi_full', 'creator_nid'])
+        calc['bulan'] = calc['created_at'].dt.to_period('M')
+        monthly = calc.groupby(['nama_lokasi_full', 'bulan']).agg(
+            findings=('kode_temuan', 'size'),
+            reporters=('creator_nid', 'nunique')
+        ).reset_index()
+        monthly = monthly[monthly['reporters'] > 0]
+        monthly['ratio'] = monthly['findings'] / monthly['reporters']
+        loc_avg = monthly.groupby('nama_lokasi_full')['ratio'].mean()
+        lowest_9 = loc_avg.nsmallest(9)
+        dev["lowest_ratio_9_locs"] = [(loc, round(ratio, 2)) for loc, ratio in lowest_9.items()]
+
+    # === 2a: Division — lowest 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')
+        agg = calc.groupby(['nama', 'bulan']).agg(
+            findings=('kode_temuan', 'size'),
+            reporters=('creator_nid', 'nunique')
+        )
+        agg = agg[agg['reporters'] > 0].reset_index()
+        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)
+
+    # === 2b: Executor — slowest resolution ===
+    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(), 2)
+                dev["obj3b_slowest_executor"] = (name, val)
+
+    # === 2c: Reporter — lowest frequency ===
+    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')
+        monthly = calc.groupby(['creator_name', 'bulan']).size().reset_index(name='count')
+        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)
+
+    # === 2d: Division — slowest resolution ===
+    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(), 2)
+                dev["obj3d_slowest_div"] = (name, val)
+
+    # === 3. Non-Positive composition ===
+    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), 2)
+        dev["obj4_unsafe_action_pct"] = round(cnt.get("Unsafe Action", 0), 2)
+        dev["obj4_near_miss_pct"] = round(cnt.get("Near Miss", 0), 2)
+
+    # === 4. Risk Quadrants ===
+    X_LIMIT, Y_LIMIT = 20, 3
+    if {'nama', 'created_at', 'days_to_close', 'kode_temuan'}.issubset(df.columns):
+        calc = df.copy()
+        calc['created_at'] = pd.to_datetime(calc['created_at'], errors='coerce')
+        calc = calc.assign(month=calc['created_at'].dt.to_period('M').astype(str))
+        monthly_counts = calc.groupby(['nama', 'month'])['kode_temuan'].nunique().reset_index()
+        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 ===
+    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'))
+            cat_avg = (nonpos.groupby('kategori').size() / n_months).sort_values(ascending=False).head(2)
+            dev["obj6_top2_categories"] = [(cat, round(val, 2)) for cat, val in cat_avg.items()]
+
+    return dev
+
+# ==============================
+# 4. LLM UTILS (aman & cepat)
+# ==============================
+def generate_llm_text(insight: str, mode: str = "rec") -> str:
+    """Generate rec or mit text using Phi-3-mini."""
+    suffix = "Recommend a single high-leverage action." if mode == "rec" else "Propose one automated/systemic risk control."
+    messages = [
+        {"role": "system", "content": "You are PLN's Lead Safety AI. Output ONLY a short, professional sentence. Be directive. No markdown, no emoticons."},
+        {"role": "user", "content": f"Insight: {insight}\n\n{suffix}"}
+    ]
     
     try:
         out = pipe(
-            prompt,
-            max_new_tokens=128,
+            messages,
             do_sample=False,
             temperature=0.1,
             return_full_text=False
         )
         text = out[0]["generated_text"].strip()
-        # Clean common artifacts
-        text = re.sub(r"^.*?:\s*", "", text)  # remove "Recommendation: "
-        text = re.sub(r"[\n\"`]", " ", text).strip()
-        if len(text) < 10 or len(text) > 200:
-            raise ValueError("Invalid length")
-        return text
+        # Clean
+        text = re.sub(r"^(Recommendation|Mitigation|Action|Control):\s*", "", text, flags=re.IGNORECASE)
+        text = re.sub(r"[\n\"`*]", " ", text).strip(". ")
+        return text[:250]  # Batas panjang
     except Exception as e:
-        # Fallback — still your preferred phrasing
+        st.warning(f"LLM fallback for {mode}: {e}")
+        # Fallback — tetap profesional & sesuai gaya Anda
         fallbacks = {
             ("1", "rec"): "Launch Agency Activation Sprint: ≥1 spot inspection/week per low-ratio location.",
-            ("1", "mit"): "Deploy QR-code checklists + WhatsApp reminders; target ratio ≥0.5 in 45 days.",
-            ("2", "rec"): "Activate Agentic Capacity Dashboard for real-time ratio & lead-time monitoring.",
-            ("2", "mit"): "Auto-trigger coaching alerts if deviation >20% from divisional baseline.",
+            ("1", "mit"): "Deploy QR-code checklists + automated reminders; target ratio ≥0.5 in 45 days.",
+            ("2", "rec"): "Activate Agentic Capacity Dashboard for real-time monitoring of reporter engagement and resolution efficiency.",
+            ("2", "mit"): "Auto-trigger coaching alerts if performance deviates >20% from divisional baseline.",
             ("3", "rec"): "Enforce photo-based validation for all Unsafe Condition/Action/Near Miss submissions.",
             ("3", "mit"): "System blocks submission if photo evidence or justification is missing.",
-            ("4", "rec"): "Assign dedicated safety crews to Quadrant I; enforce ‘One Finding, One Day’ for Quadrant II.",
-            ("4", "mit"): "Auto-generate VP escalation reports if division remains in QI/QII ≥2 months.",
-            ("5", "rec"): "Form cross-functional RCA Task Force (Civil, Electrical, HSE, Contractors) for top categories.",
+            ("4", "rec"): "Assign dedicated safety crews to Quadrant I; enforce ‘One Finding, One Day’ closure for Quadrant II.",
+            ("4", "mit"): "Auto-generate VP escalation reports if division remains in risk quadrant ≥2 months.",
+            ("5", "rec"): "Form cross-functional RCA Task Force (Civil, Electrical, HSE, Contractors) for top recurring categories.",
             ("5", "mit"): "Update tender templates: all bids must include mitigations for these historical findings.",
         }
-        return fallbacks.get((insight_lines.index(insight)+1 if insight in insight_lines else "1", mode), 
-                            "Review insight and implement targeted action.")
+        return fallbacks.get((str(len(insight_list) + 1), mode), "Review insight and implement targeted action.")
+
+# ==============================
+# 5. MAIN EXECUTION
+# ==============================
+st.markdown("<h3 class='section-title'>OBJECTIVE 7 — Insight and Recommendation</h3>", unsafe_allow_html=True)
+
+# ✅ Pastikan df_filtered ada
+if 'df_filtered' not in st.session_state:
+    st.error("⚠️ `df_filtered` not found in session state. Please load data first.")
+    st.stop()
+
+df_filtered = st.session_state.df_filtered
+dev = extract_agentic_insights_v5(df_filtered)
+
+# === BUILD INSIGHT LINES (2 desimal, clean) ===
+insight_lines = []
 
-# Extract clean insight strings (for LLM input)
-clean_insights = []
 if dev["lowest_ratio_9_locs"]:
-    locs = ", ".join([f"{loc} ({ratio})" for loc, ratio in dev["lowest_ratio_9_locs"]])
-    clean_insights.append(f"Nine locations with the lowest finding-to-reporter ratio: {locs}.")
+    loc_list = ", ".join([f"<strong>{loc}</strong> ({ratio:.2f})" 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}.")
 
-parts_desc = []
+parts = []
 if dev["obj3a_lowest_div"]:
-    parts_desc.append(f"division {dev['obj3a_lowest_div'][0]} (ratio: {dev['obj3a_lowest_div'][1]})")
+    name, val = dev["obj3a_lowest_div"]
+    parts.append(f"division <strong>{name}</strong> (ratio: {val:.2f})")
 if dev["obj3c_lowest_reporter"]:
-    parts_desc.append(f"reporter {dev['obj3c_lowest_reporter'][0]} ({dev['obj3c_lowest_reporter'][1]} findings/month)")
+    name, val = dev["obj3c_lowest_reporter"]
+    parts.append(f"reporter <strong>{name}</strong> ({val:.2f} findings/month)")
 if dev["obj3d_slowest_div"]:
-    parts_desc.append(f"division {dev['obj3d_slowest_div'][0]} (avg. resolution: {dev['obj3d_slowest_div'][1]} days)")
+    name, val = dev["obj3d_slowest_div"]
+    parts.append(f"division <strong>{name}</strong> (avg. resolution: {val:.2f} days)")
 if dev["obj3b_slowest_executor"]:
-    parts_desc.append(f"executor {dev['obj3b_slowest_executor'][0]} (avg. resolution: {dev['obj3b_slowest_executor'][1]} days)")
-if parts_desc:
-    clean_insights.append(f"Uneven operational capacity: {'; '.join(parts_desc)}.")
+    name, val = dev["obj3b_slowest_executor"]
+    parts.append(f"executor <strong>{name}</strong> (avg. resolution: {val:.2f} days)")
+
+if parts:
+    insight_lines.append(
+        f"2. Uneven operational capacity detected: {'; '.join(parts)}. "
+        "This indicates systemic imbalance in reporting engagement and resolution efficiency."
+    )
 
+uc, ua, nm = dev["obj4_unsafe_condition_pct"], dev["obj4_unsafe_action_pct"], dev["obj4_near_miss_pct"]
 if uc + ua + nm > 0:
-    clean_insights.append(f"Non-Positive finding composition: Unsafe Condition ({uc}%), Unsafe Action ({ua}%), Near Miss ({nm}%).")
+    insight_lines.append(
+        f"3. Non-Positive finding composition: Unsafe Condition ({uc:.2f}%), Unsafe Action ({ua:.2f}%), Near Miss ({nm:.2f}%)."
+    )
 
 if dev["obj5_q1_divs"] or dev["obj5_q2_divs"]:
-    q1 = ", ".join([d for d in dev["obj5_q1_divs"][:3]]) or "—"
-    q2 = ", ".join([d for d in dev["obj5_q2_divs"][:3]]) or "—"
-    clean_insights.append(f"High-risk divisions (Q1): {q1}; Hidden-risk divisions (Q2): {q2}.")
+    q1 = ", ".join([f"<strong>{d}</strong>" for d in dev["obj5_q1_divs"][:3]]) or "—"
+    q2 = ", ".join([f"<strong>{d}</strong>" for d in dev["obj5_q2_divs"][:3]]) or "—"
+    insight_lines.append(f"4. High-risk divisions (Q1): {q1}; Hidden-risk divisions (Q2): {q2}.")
 
 if dev["obj6_top2_categories"]:
     c1, c2 = dev["obj6_top2_categories"]
-    clean_insights.append(f"Top non-Positive categories: {c1[0]} ({c1[1]}/month) and {c2[0]} ({c2[1]}/month).")
-
-# Generate paired rec & mit
-rec_lines, mit_lines = [], []
-with st.spinner("🧠 Generating Recommendation & Risk Mitigation with DeepSeek-7B..."):
-    for i, ins in enumerate(clean_insights, 1):
-        rec = safe_llm_call(ins, "rec")
-        mit = safe_llm_call(ins, "mit")
-        rec_lines.append(f"{i}. {rec}")
-        mit_lines.append(f"{i}. {mit}")
+    insight_lines.append(
+        f"5. Top two recurring non-Positive categories: <strong>{c1[0]}</strong> ({c1[1]:.2f}/month) and <strong>{c2[0]}</strong> ({c2[1]:.2f}/month)."
+    )
 
-rec_text = "<br>".join(rec_lines)
-mit_text = "<br>".join(mit_lines)
+insight_text = "<br>".join(insight_lines)
 
-# === RENDER RECOMMENDATION CARD ===
+# === RENDER 3 CARDS ===
+# Card 1: Insight
 st.markdown(
     f"""
     <div class="card" style="
-        background-color: #e8f5e9;
-        border-left: 4px solid #4CAF50;
-        padding: 18px;
+        background-color: #f8f9fa;
+        border-left: 5px solid #003DA5;
+        padding: 20px;
         margin-bottom: 24px;
         border-radius: 6px;
         box-shadow: 0 3px 6px rgba(0,0,0,0.06);
     ">
-        <h4 style="margin-top: 0; color: #2E7D32; text-align: center;">✅ Recommendation</h4>
-        <p style="margin-bottom: 0; line-height: 1.6; font-size: 0.98em;">{rec_text}</p>
+        <h4 style="margin-top: 0; color: #003DA5; text-align: center;">🔍 Insight Summary</h4>
+        <p style="margin-bottom: 0; line-height: 1.6; font-size: 0.98em;">{insight_text}</p>
     </div>
     """,
     unsafe_allow_html=True
 )
 
-# === RENDER RISK MITIGATION CARD ===
-st.markdown(
-    f"""
-    <div class="card" style="
-        background-color: #e3f2fd;
-        border-left: 4px solid #1976D2;
-        padding: 18px;
-        margin-bottom: 24px;
-        border-radius: 6px;
-        box-shadow: 0 3px 6px rgba(0,0,0,0.06);
-    ">
-        <h4 style="margin-top: 0; color: #0D47A1; text-align: center;">🛡️ Risk Mitigation Strategy</h4>
-        <p style="margin-bottom: 0; line-height: 1.6; font-size: 0.98em;">{mit_text}</p>
-    </div>
-    """,
-    unsafe_allow_html=True
-)
+if insight_lines:
+    # Generate rec & mit
+    rec_list, mit_list = [], []
+    with st.spinner("🧠 Generating Recommendation & Risk Mitigation with Phi-3-mini..."):
+        for i, ins in enumerate(insight_lines, 1):
+            # Ekstrak teks bersih untuk LLM
+            clean_ins = re.sub(r"<[^>]+>", "", ins).replace("1. ", "").replace("2. ", "").replace("3. ", "").replace("4. ", "").replace("5. ", "").strip()
+            rec = generate_llm_text(clean_ins, "rec")
+            mit = generate_llm_text(clean_ins, "mit")
+            rec_list.append(f"{i}. {rec}")
+            mit_list.append(f"{i}. {mit}")
+
+    rec_text = "<br>".join(rec_list)
+    mit_text = "<br>".join(mit_list)
+
+    # Card 2: Recommendation
+    st.markdown(
+        f"""
+        <div class="card" style="
+            background-color: #e8f5e9;
+            border-left: 5px solid #4CAF50;
+            padding: 20px;
+            margin-bottom: 24px;
+            border-radius: 6px;
+            box-shadow: 0 3px 6px rgba(0,0,0,0.06);
+        ">
+            <h4 style="margin-top: 0; color: #2E7D32; text-align: center;">✅ Recommendation</h4>
+            <p style="margin-bottom: 0; line-height: 1.6; font-size: 0.98em;">{rec_text}</p>
+        </div>
+        """,
+        unsafe_allow_html=True
+    )
+
+    # Card 3: Risk Mitigation
+    st.markdown(
+        f"""
+        <div class="card" style="
+            background-color: #e3f2fd;
+            border-left: 5px solid #1976D2;
+            padding: 20px;
+            margin-bottom: 24px;
+            border-radius: 6px;
+            box-shadow: 0 3px 6px rgba(0,0,0,0.06);
+        ">
+            <h4 style="margin-top: 0; color: #0D47A1; text-align: center;">🛡️ Risk Mitigation Strategy</h4>
+            <p style="margin-bottom: 0; line-height: 1.6; font-size: 0.98em;">{mit_text}</p>
+        </div>
+        """,
+        unsafe_allow_html=True
+    )
+else:
+    st.info("ℹ️ No insights generated. Ensure required columns are present in the dataset.")