Update app.py

app.py

@@ -10,6 +10,7 @@ import numpy as np
 import matplotlib
 matplotlib.use('Agg')
 import matplotlib.pyplot as plt
+import streamlit.components.v1 as components
 import shap, joblib, os, json, warnings, re, requests, io, hashlib
 warnings.filterwarnings('ignore')
 
@@ -401,7 +402,8 @@ def call_groq_tools(messages, api_key, tools=None):
 def _clean_response(text: str) -> tuple[str, dict]:
     """
     Strip all LLM artifacts from response text and extract any embedded adjustments.
-    Handles: <think>…</think> blocks, raw JSON objects {"field":…}, [ADJUST:…] tags.
+    Handles: <think>…</think>, raw JSON objects, [ADJUST:…] tags, and trailing
+    standalone recommendation lines that duplicate inline suggestions.
     Returns (clean_text, adjustments_dict).
     """
     if not text:
@@ -412,17 +414,16 @@ def _clean_response(text: str) -> tuple[str, dict]:
     text = re.sub(r'<think>.*?</think>', '', text, flags=re.DOTALL).strip()
 
     # ── Parse & strip inline JSON objects (tool-call leak) ────────
-    # Matches {"field": "...", "value": ..., "reason": "..."} anywhere in text
     json_pattern = re.compile(
         r'\{[^{}]*"field"\s*:\s*"([^"]+)"[^{}]*"value"\s*:\s*([\d.]+)[^{}]*\}',
         re.DOTALL
     )
+    valid_fields = {
+        'digital_presence_score', 'business_age_years', 'num_employees',
+        'monthly_cash_flow', 'duration', 'loan_rp'
+    }
     for m in json_pattern.finditer(text):
         field, val = m.group(1), float(m.group(2))
-        valid_fields = {
-            'digital_presence_score', 'business_age_years', 'num_employees',
-            'monthly_cash_flow', 'duration', 'loan_rp'
-        }
         if field in valid_fields:
             adjustments[field] = val
     text = json_pattern.sub('', text).strip()
@@ -434,18 +435,37 @@ def _clean_response(text: str) -> tuple[str, dict]:
             adjustments[field] = val
     text = re.sub(r'\[ADJUST:[^\]]+\]', '', text).strip()
 
-    # ── Strip trailing standalone slider-hint lines (LLM sometimes writes
-    #    "Naikkan X ke Y" or "Tingkatkan cash flow ke ..." as a bare last line
-    #    after the ADJUST tag has already been extracted and removed)
-    # Pattern: last 1-2 lines that are short (<= 80 chars) and match
-    # "Naikkan/Tingkatkan/Pertahankan/Kurangi ... ke ..." with no sentence context.
-    _standalone = re.compile(
-        r'\n+((?:Naikkan|Tingkatkan|Pertahankan|Kurangi|Raise|Increase|Maintain|Reduce|Lower)'
-        r'[^.\n]{5,70}(?:ke|to|juta|M|score)[^.\n]{0,40})\s*$',
-        re.IGNORECASE
+    # ── FIX: Strip ALL trailing standalone action-lines ──────────
+    # Only strip when there's meaningful content before the trailing line,
+    # i.e. the response has more than 1 non-empty line. This prevents
+    # incorrectly stripping single-line responses that start with action words.
+    _action_line = re.compile(
+        r'\n[ \t]*(?:'
+        # Indonesian
+        r'(?:Naikkan|Tingkatkan|Pertahankan|Kurangi|Optimalkan|Perbaiki|Tambahkan|Pastikan|Coba|Daftarkan)'
+        r'|'
+        # English
+        r'(?:Raise|Increase|Maintain|Reduce|Lower|Improve|Try|Optimize|Register|Ensure)'
+        r'|'
+        # Hindi
+        r'(?:बढ़ाएं|घटाएं|सुधारें|पंजीकृत|सुनिश्चित)'
+        r')[^\n]{3,90}$',
+        re.IGNORECASE | re.MULTILINE
     )
-    text = _standalone.sub('', text).strip()
-
+    # Run up to 5 passes; guard: only strip if text has >1 non-empty line
+    for _ in range(5):
+        non_empty_lines = [l for l in text.split('\n') if l.strip()]
+        if len(non_empty_lines) <= 1:
+            break  # Don't strip the only content line
+        new_text = _action_line.sub('', text).rstrip()
+        if new_text == text:
+            break
+        text = new_text
+
+    # ── Strip markdown bold artifacts left on their own line ─────
+    text = re.sub(r'\n\*\*\s*\*\*\s*$', '', text, flags=re.MULTILINE).strip()
+
+    # ── Collapse excessive blank lines ────────────────────────────
     text = re.sub(r'\n{3,}', '\n\n', text).strip()
 
     return text, adjustments
@@ -706,7 +726,6 @@ _defaults = dict(
     chat_history=[], shap_png=None,
     wi_dig=None, wi_biz=None, wi_emp=None,
     wi_cash=None, wi_dur=None, wi_loan=None,
-    # ② Memory persistence
     chat_summary='',
     memory_loaded=False,
 )
@@ -982,7 +1001,6 @@ def _get_top_issue(raw_input, pd_pct, lang):
 
 # ============================================================
 # AI CHAT — Maya persona
-# ③ Few-shot + ④ Chain-of-Thought + ① Tool Calling + ③ Memory
 # ============================================================
 def get_chat_response(user_msg, history, result, raw_input, shap_vals, feature_names, lang, rag_index=None):
     pd_pct    = result['pd'] * 100
@@ -1003,18 +1021,18 @@ def get_chat_response(user_msg, history, result, raw_input, shap_vals, feature_n
         except Exception:
             context = ""
 
-    # ── ② Few-shot examples (explicit Q&A demonstrations) ─────
+    # ── Few-shot examples ─────────────────────────────────────
     fewshot = """
 CONTOH DIALOG MAYA (few-shot — ikuti gaya & kualitas ini):
 ---
 User: "Gimana cara turunin skor PD saya?"
-Maya: "Berdasarkan data kamu, 3 prioritas utama: 1) **Naikkan digital score** ke 75+ via Google Business & marketplace aktif — ini faktor SHAP terbesar. 2) **Optimalkan cash flow** — dokumentasikan semua pemasukan, pisah rekening bisnis & pribadi. 3) **Urus NPWP** kalau belum ada. Mau saya simulasikan dampaknya ke skor sekarang?"
+Maya: "Berdasarkan data kamu, 3 prioritas utama: 1) **Naikkan digital score ke 75+** via Google Business & marketplace aktif [ADJUST: digital_presence_score=75] — ini faktor SHAP terbesar. 2) **Optimalkan cash flow ke Rp 25jt/bln** [ADJUST: monthly_cash_flow=25000000] — dokumentasikan semua pemasukan, pisah rekening. 3) **Urus NPWP** kalau belum ada. Mau lihat simulasinya di tab What-If?"
 ---
 User: "Kenapa skor saya tinggi banget?"
 Maya: "Dua driver terbesar menurut SHAP: **business_age_years** (SHAP=+0.234) — bisnis muda = risiko lebih tinggi di mata model, dan **digital_presence_score** (SHAP=+0.189) — bank pakai ini sebagai proxy kredibilitas online. Fokus perbaiki keduanya dulu, dampaknya paling signifikan!"
 ---
 User: "Pinjaman berapa yang paling aman buat saya?"
-Maya: "Dengan cash flow Rp 15jt/bln dan tenor 24 bln, batas aman kamu adalah **Rp 126jt** (rasio cicilan 35%). Pinjaman kamu sekarang Rp 50jt — ✅ masih aman! Kalau mau tambah, pastikan cash flow naik dulu biar rasio tetap sehat."
+Maya: "Dengan cash flow Rp 15jt/bln dan tenor 24 bln, batas aman kamu adalah **Rp 126jt** (rasio cicilan 35%). Pinjaman kamu sekarang Rp 50jt — ✅ masih aman! Kalau mau tambah, pastikan cash flow naik dulu."
 ---
 """
 
@@ -1024,13 +1042,13 @@ Maya: "Dengan cash flow Rp 15jt/bln dan tenor 24 bln, batas aman kamu adalah **R
         lbl = {'id': '📝 Ringkasan percakapan sebelumnya', 'en': '📝 Earlier conversation summary', 'hi': '📝 पूर्व वार्तालाप सारांश'}
         summary_block = f"\n{lbl.get(lang, lbl['en'])}:\n{chat_summary}\n"
 
-    # ── ④ Chain-of-Thought instructions ─────────────────────────
+    # ── Chain-of-Thought instructions ─────────────────────────
     cot_block = """
 CARA BERPIKIR — CHAIN OF THOUGHT (lakukan ini secara SILENT sebelum menjawab):
 ① Apa yang sebenarnya ditanyakan user? (bukan hanya kata-katanya — cari intent sebenarnya)
 ② Data pemohon mana yang paling relevan? (PD score, cash flow, digital score, NPWP, dll)
 ③ Apa saran paling KONKRET & berdampak tinggi untuk UMKM spesifik ini?
-④ Apakah perlu adjust slider What-If? Jika ya, gunakan tool adjust_whatif_slider dengan field & value spesifik.
+④ Apakah perlu adjust slider What-If? Jika ya, embed tag [ADJUST: field=value] LANGSUNG di dalam kalimat rekomendasi.
 → Tulis HANYA jawaban akhir — jangan tampilkan proses berpikir di output.
 """
 
@@ -1058,21 +1076,27 @@ CARA BERPIKIR — CHAIN OF THOUGHT (lakukan ini secara SILENT sebelum menjawab):
         f"{summary_block}"
         + cot_block
         + fewshot +
-        f"ATURAN MENJAWAB:\n"
+        f"ATURAN MENJAWAB — BACA SEMUA DENGAN SEKSAMA:\n"
         f"1. Jawab BEBAS — tidak harus soal kredit. Kalau ditanya soal diri sendiri, perkenalkan sebagai Maya.\n"
         f"2. Kalau relevan, selalu hubungkan ke data pemohon dengan menyebut angka aktualnya.\n"
         f"3. Berikan saran KONKRET dan SPESIFIK, bukan generik.\n"
         f"4. Maksimal 150 kata kecuali diminta lebih panjang.\n"
-        f"5. Jika merekomendasikan nilai spesifik, EMBED tag `[ADJUST: field=value]` LANGSUNG di dalam kalimat saran.\n"
-        f"   BENAR:  'Coba naikkan digital score-mu ke 75 [ADJUST: digital_presence_score=75] — dampaknya paling besar.'\n"
-        f"   SALAH:  'Saran: naikkan digital score.\\nNaikkan digital score ke 75 [ADJUST: digital_presence_score=75]'\n"
-        f"   SALAH:  Menulis baris terpisah seperti 'Naikkan digital score ke 75' di akhir response.\n"
-        f"   PENTING: Tag [ADJUST:] hanya boleh muncul SEKALI per field, INLINE di dalam kalimat, bukan di baris tersendiri.\n"
-        f"   PENTING: Jangan pernah menulis JSON object {{...}} di response.\n"
-        f"   Field valid: digital_presence_score(1-100), business_age_years(1-20),\n"
-        f"   num_employees(1-50), monthly_cash_flow(angka Rp), duration(4-72), loan_rp(angka Rp)\n"
+        f"5. FORMAT TAG [ADJUST:] — WAJIB IKUTI ATURAN INI:\n"
+        f"   ✅ BENAR: Embed tag langsung di akhir kalimat saran:\n"
+        f"      'Coba naikkan digital score ke 75 [ADJUST: digital_presence_score=75] — dampaknya paling besar.'\n"
+        f"   ❌ SALAH: Menulis kalimat saran TANPA tag, lalu menambah baris terpisah:\n"
+        f"      'Coba naikkan digital score ke 75.\\nNaikkan digital score ke 75 [ADJUST: ...]'\n"
+        f"   ❌ SALAH: Menulis tag di baris sendiri di akhir response:\n"
+        f"      '...skor kamu akan membaik.\\n[ADJUST: digital_presence_score=75]'\n"
+        f"   ❌ SALAH BESAR: Menulis ulang rekomendasi di baris terpisah setelah selesai:\n"
+        f"      '...semua sudah dijelaskan di atas.\\nNaikkan digital score ke 75'\n"
+        f"   RULE: Setiap rekomendasi bernilai angka = SATU kalimat + SATU tag [ADJUST:] inline. TIDAK LEBIH.\n"
+        f"   RULE: JANGAN pernah tulis JSON object {{...}} di response.\n"
+        f"   RULE: Field valid: digital_presence_score(1-100), business_age_years(1-20),\n"
+        f"         num_employees(1-50), monthly_cash_flow(angka Rp), duration(4-72), loan_rp(angka Rp)\n"
         f"6. JANGAN pernah balik ke template. Jawab seperti manusia cerdas.\n"
         f"7. Baca context percakapan sebelumnya sebelum menjawab — jaga konsistensi topik.\n"
+        f"8. AKHIRI response dengan kalimat lengkap. Jangan tambahkan apapun setelah kalimat terakhir.\n"
     )
 
     messages = [{"role": "system", "content": system}]
@@ -1080,7 +1104,7 @@ CARA BERPIKIR — CHAIN OF THOUGHT (lakukan ini secara SILENT sebelum menjawab):
         messages.append({"role": m['role'], "content": m['content']})
     messages.append({"role": "user", "content": user_msg})
 
-    # ── ① Formal Tool Calling via _call_chat_llm ──────────────
+    # ── Formal Tool Calling via _call_chat_llm ────────────────
     response, tool_calls, _llm_src, _last_error = _call_chat_llm(messages)
 
     # Parse tool_calls for adjustments (formal schema)
@@ -1120,14 +1144,14 @@ CARA BERPIKIR — CHAIN OF THOUGHT (lakukan ini secara SILENT sebelum menjawab):
                 adjustments['has_npwp'] = 1
             if dig < 75:
                 n = '2' if not npwp else '1'
-                tips.append(_t(f"{n}. **Naikkan Digital Score** dari {dig} ke 75+ — aktif di Google Business, Tokopedia/Shopee",
-                               f"{n}. **Raise Digital Score** from {dig} to 75+ — Google Business, marketplace",
+                tips.append(_t(f"{n}. **Naikkan Digital Score dari {dig} ke 75+** — aktif di Google Business, Tokopedia/Shopee",
+                               f"{n}. **Raise Digital Score from {dig} to 75+** — Google Business, marketplace",
                                f"{n}. **डिजिटल स्कोर** {dig} से 75+ करें"))
                 adjustments['digital_presence_score'] = 75
             if cf < 25e6:
                 n = str(len(tips)+1)
-                tips.append(_t(f"{n}. **Optimalkan cash flow** dari Rp {cf/1e6:.0f}jt ke 25jt+/bln — diversifikasi produk",
-                               f"{n}. **Grow cash flow** from Rp {cf/1e6:.0f}M to 25M+/month",
+                tips.append(_t(f"{n}. **Optimalkan cash flow dari Rp {cf/1e6:.0f}jt ke 25jt+/bln** — diversifikasi produk",
+                               f"{n}. **Grow cash flow from Rp {cf/1e6:.0f}M to 25M+/month**",
                                f"{n}. **नकद प्रवाह** Rp {cf/1e6:.0f}M से 25M+ करें"))
                 adjustments['monthly_cash_flow'] = 25000000
             if not tips:
@@ -1281,7 +1305,7 @@ CARA BERPIKIR — CHAIN OF THOUGHT (lakukan ini secara SILENT sebelum menjawab):
                 f"आपका PD **{pd_pct:.1f}%** है और {top_issue}"
             )
 
-    # Final cleanup: strip think-tags, leaked JSON, ADJUST tags
+    # Final cleanup: strip think-tags, leaked JSON, ADJUST tags, trailing lines
     if response:
         response, extra_adj = _clean_response(response)
         for field, val in extra_adj.items():
@@ -1291,6 +1315,155 @@ CARA BERPIKIR — CHAIN OF THOUGHT (lakukan ini secara SILENT sebelum menjawab):
     return response or "...", adjustments, _last_error
 
 
+# ============================================================
+# FLOATING DRAGGABLE GIF + SOUND PER TAB
+# ============================================================
+def render_media_manager():
+    """
+    Inject floating draggable GIFs + audio per tab.
+    Tab index mapping:
+      0 = SHAP   → 1 GIF (crying cat) + idk.mp3
+      3 = What-If → 3 GIFs (dancing chibi) + shake.mp3
+      others      → semua tersembunyi, audio pause
+    """
+    html = """
+<div id="media-container">
+    <!-- ── AUDIO ─────────────────────────────────────── -->
+    <audio id="audio-shap"   loop preload="auto" src="/static/idk.mp3"></audio>
+    <audio id="audio-whatif" loop preload="auto" src="/static/shake.mp3"></audio>
+
+    <!-- ── GIFs: SHAP tab ────────────────────────────── -->
+    <div id="gif-shap" class="drag-gif" style="bottom:20px;left:20px;display:none;">
+        <img src="https://media1.tenor.com/m/8606581414273585609/crying-cat-no-msgs-kitty-meme-screaming.gif" width="140" draggable="false">
+        <div class="drag-close" onclick="document.getElementById('gif-shap').style.display='none'">✕</div>
+    </div>
+
+    <!-- ── GIFs: What-If tab ─────────────────────────── -->
+    <div id="gif-wi-1" class="drag-gif" style="bottom:20px;left:20px;display:none;">
+        <img src="https://media1.tenor.com/m/20429653/chibi-haachama-hololive-dancing-squareofthelightones-akai-haato.gif" width="110" draggable="false">
+        <div class="drag-close" onclick="document.getElementById('gif-wi-1').style.display='none'">✕</div>
+    </div>
+    <div id="gif-wi-2" class="drag-gif" style="bottom:20px;left:170px;display:none;">
+        <img src="https://media1.tenor.com/m/13376873318217365775/evernight-hsr-lonely-lonely-i-guess-im-lonely-dance.gif" width="110" draggable="false">
+        <div class="drag-close" onclick="document.getElementById('gif-wi-2').style.display='none'">✕</div>
+    </div>
+    <div id="gif-wi-3" class="drag-gif" style="bottom:20px;left:320px;display:none;">
+        <img src="https://media1.tenor.com/m/129497420709683859/anime-chibi-azazal-silly-jumping.gif" width="110" draggable="false">
+        <div class="drag-close" onclick="document.getElementById('gif-wi-3').style.display='none'">✕</div>
+    </div>
+</div>
+
+<style>
+.drag-gif {
+    position: fixed;
+    z-index: 9999;
+    cursor: move;
+    user-select: none;
+    filter: drop-shadow(0 4px 12px rgba(0,0,0,0.4));
+    transition: opacity 0.2s;
+}
+.drag-gif:hover { opacity: 0.92; }
+.drag-gif img { display: block; border-radius: 8px; pointer-events: none; }
+.drag-close {
+    position: absolute;
+    top: -8px; right: -8px;
+    background: #e74c3c;
+    color: white;
+    border-radius: 50%;
+    width: 20px; height: 20px;
+    font-size: 11px;
+    line-height: 20px;
+    text-align: center;
+    cursor: pointer;
+    font-weight: 700;
+    box-shadow: 0 2px 6px rgba(0,0,0,0.3);
+    transition: transform 0.15s;
+}
+.drag-close:hover { transform: scale(1.2); }
+</style>
+
+<script>
+(function () {
+    'use strict';
+
+    // ── Draggable logic ───────────────────────────────────────
+    function makeDraggable(el) {
+        let ox = 0, oy = 0, mx = 0, my = 0;
+        el.addEventListener('mousedown', function (e) {
+            if (e.target.classList.contains('drag-close')) return;
+            e.preventDefault();
+            mx = e.clientX; my = e.clientY;
+            document.addEventListener('mousemove', drag);
+            document.addEventListener('mouseup', stopDrag);
+        });
+        function drag(e) {
+            ox = mx - e.clientX; oy = my - e.clientY;
+            mx = e.clientX; my = e.clientY;
+            el.style.top  = (el.offsetTop  - oy) + 'px';
+            el.style.left = (el.offsetLeft - ox) + 'px';
+            el.style.bottom = 'auto';
+        }
+        function stopDrag() {
+            document.removeEventListener('mousemove', drag);
+            document.removeEventListener('mouseup', stopDrag);
+        }
+    }
+    document.querySelectorAll('.drag-gif').forEach(makeDraggable);
+
+    // ── Audio + GIF per-tab controller ───────────────────────
+    const audShap   = document.getElementById('audio-shap');
+    const audWhatif = document.getElementById('audio-whatif');
+    const allGifs   = document.querySelectorAll('.drag-gif');
+
+    function stopAll() {
+        allGifs.forEach(g => g.style.display = 'none');
+        [audShap, audWhatif].forEach(a => { a.pause(); a.currentTime = 0; });
+    }
+
+    window.setActiveTab = function (idx) {
+        stopAll();
+        if (idx === 0) {                // SHAP tab
+            document.getElementById('gif-shap').style.display = 'block';
+            audShap.play().catch(() => {});
+        } else if (idx === 3) {         // What-If tab
+            ['gif-wi-1','gif-wi-2','gif-wi-3'].forEach(id => {
+                document.getElementById(id).style.display = 'block';
+            });
+            audWhatif.play().catch(() => {});
+        }
+    };
+
+    // ── Auto-detect active Streamlit tab ─────────────────────
+    // Streamlit renders tabs as <li role="tab" aria-selected="true">
+    function detectTab() {
+        const tabs = document.querySelectorAll('[role="tab"]');
+        for (let i = 0; i < tabs.length; i++) {
+            if (tabs[i].getAttribute('aria-selected') === 'true') {
+                window.setActiveTab(i);
+                return;
+            }
+        }
+    }
+
+    // Watch for Streamlit rerenders via MutationObserver
+    const observer = new MutationObserver(function (mutations) {
+        for (const m of mutations) {
+            for (const n of m.addedNodes) {
+                if (n.nodeType === 1) { detectTab(); return; }
+            }
+        }
+    });
+    observer.observe(document.body, { childList: true, subtree: true });
+
+    // Initial detection after short delay (Streamlit needs time to render tabs)
+    setTimeout(detectTab, 800);
+    setTimeout(detectTab, 2000);
+})();
+</script>
+"""
+    components.html(html, height=0, scrolling=False)
+
+
 # ============================================================
 # HEADER
 # ============================================================
@@ -1406,7 +1579,6 @@ if submitted:
             st.session_state.chat_history=[]
             st.session_state.chat_summary=''
             for k in ['wi_dig','wi_biz','wi_emp','wi_cash','wi_dur','wi_loan']: st.session_state[k]=None
-            # Clear persisted memory on new submission
             _save_chat_memory([], '')
         except Exception as e:
             st.error(f"Error: {e}"); st.stop()
@@ -1451,6 +1623,11 @@ for col,(lbl_k,val,sub_k,color) in zip(st.columns(4),[
                     f'<div class="kpi-sub">{T(sub_k,lang)}</div></div>', unsafe_allow_html=True)
 st.markdown("<br>", unsafe_allow_html=True)
 
+# ============================================================
+# FLOATING MEDIA (inject once, JS auto-detects active tab)
+# ============================================================
+render_media_manager()
+
 # ============================================================
 # TABS
 # ============================================================
@@ -1482,7 +1659,6 @@ with t2:
 with t3:
     st.markdown(f'<div class="sec-title">{T("chat_title",lang)}</div>', unsafe_allow_html=True)
 
-    # Show memory summary badge if exists
     if st.session_state.get('chat_summary'):
         n_turns = len(st.session_state.chat_history)
         lbl = {'id': f'🧠 Memory aktif · {n_turns} pesan tersimpan',
@@ -1520,7 +1696,6 @@ with t3:
                 st.warning(f"⚠️ Semua LLM gagal, menggunakan Smart Fallback. Error: `{_debug_err[:100]}`")
         st.session_state.chat_history.append({'role':'assistant','content':ai_resp})
 
-        # ② Persist memory after each turn
         _save_chat_memory(
             st.session_state.chat_history,
             st.session_state.get('chat_summary', '')