Update app.py

app.py

@@ -1,21 +1,26 @@
 # ============================================================
-# 🏦 SME Credit Risk Assessment — FINAL (FIXED v3)
+# 🏦 SME Credit Risk Assessment — FINAL (FIXED v4)
 # Final Project | AI Engineering Bootcamp Batch 10
 # Author: 1na37
 # ============================================================
-# BASE: Doc4 (no GIF/music) as foundation
-# RESTORED from Doc3:
+# BASE: Doc6 (live version — with GIF/music sidebar)
+# RESTORED:
 #   - Full CoT block with ① ② ③ ④ numbered steps
 #   - Full few-shot examples (3 dialogs)
-#   - _tab_step(0) called at correct place before tabs
-#   - inject_media_manager() REMOVED intentionally (no GIF/music)
-# FIXED:
-#   - _tab_label_ph always defined before use
-#   - _tab_step works even before result exists
-#   - Tool calling → slider trigger fully wired
-#   - _clean_response correctly strips [ADJUST:] without orphan sentences
-#   - Memory badge shows correctly
-#   - SHAP fallback message clear
+#   - _tab_step always defined before use
+# ROOT CAUSE FIX (screenshot bug — sliders never triggered):
+#   - Free-tier LLMs (gemini-flash, llama, etc.) frequently IGNORE
+#     [ADJUST:] tag instructions and respond in natural language only.
+#   - _clean_response only extracts explicit [ADJUST:] tags → empty dict
+#   - SOLUTION: _extract_adjustments_semantic() parses numeric values
+#     directly from natural LLM text as a guaranteed fallback.
+#     e.g. "naikkan digital score ke 75" → digital_presence_score=75
+#          "Rp 25jt/bln" near "cash flow" → monthly_cash_flow=25000000
+#          "digital score jadi 80"        → digital_presence_score=80
+# OTHER FIXES:
+#   - _tab_prog_ph always defined before use (no NameError)
+#   - Memory badge shows 🧠 icon
+#   - 3 new fallback handlers: digital-score-target, humor, 6-month plan
 # ============================================================
 
 import streamlit as st
@@ -504,6 +509,117 @@ def _clean_response(text: str) -> tuple:
     text = re.sub(r'\n{3,}', '\n\n', text).strip()
     return text, adjustments
 
+def _extract_adjustments_semantic(text: str, raw_input: dict) -> dict:
+    """
+    FALLBACK: parse numeric slider values from natural LLM text.
+
+    Free-tier LLMs (gemini-flash, llama-3, etc.) frequently ignore [ADJUST:] tag
+    instructions and respond in plain natural language. This function detects
+    numeric recommendations in the response text and maps them to What-If fields.
+
+    Called AFTER _clean_response — only fills in fields not already captured
+    by explicit [ADJUST:] tags or formal tool calls.
+
+    Examples handled:
+      "naikkan digital score ke 75"       → digital_presence_score=75
+      "digital score jadi 80"             → digital_presence_score=80
+      "Rp 25jt/bln" near "cash flow"      → monthly_cash_flow=25_000_000
+      "pinjaman ideal Rp 128jt"           → loan_rp=128_000_000
+      "tenor 36 bulan"                    → duration=36
+    """
+    if not text:
+        return {}
+    adjustments = {}
+    low = text.lower()
+
+    # ── Digital presence score ────────────────────────────────
+    for pat in [
+        r'digital\s+score\s+(?:ke|jadi|to|→|->|=)\s*(\d+)',
+        r'naikkan\s+digital\s+(?:score\s+)?(?:ke|jadi|to)\s*(\d+)',
+        r'raise\s+digital\s+(?:score\s+)?to\s*(\d+)',
+        r'digital\s+(?:ke|jadi)\s*(\d+)',
+        r'digital\s+presence\s+(?:score\s+)?(?:ke|jadi|to)\s*(\d+)',
+    ]:
+        m = re.search(pat, low)
+        if m:
+            val = int(m.group(1))
+            if 1 <= val <= 100:
+                adjustments['digital_presence_score'] = float(val)
+                break
+
+    # ── Monthly cash flow ─────────────────────────────────────
+    # Match "Rp NNjt" or "NNjt" near cash flow context
+    for pat in [
+        r'cash\s+flow\s+(?:ke|jadi|to|→)\s*rp\s*(\d+)\s*(?:jt|juta|m\b)',
+        r'(?:optimal|target|naikkan)\s+cash\s+flow.*?rp\s*(\d+)\s*(?:jt|juta|m\b)',
+        r'rp\s*(\d+)\s*(?:jt|juta)\s*/\s*(?:bln|bulan|month)',
+        r'cash\s+flow\s+.*?(\d+)\s*(?:jt|juta)\s*/\s*(?:bln|bulan|month)',
+        r'cash\s+flow.*?rp\s*(\d+)\s*(?:jt|juta)',
+    ]:
+        m = re.search(pat, low)
+        if m:
+            val_m = int(m.group(1)) * 1_000_000
+            cur_cf = raw_input.get('monthly_cash_flow', 0)
+            # Only update if it's a recommendation (different from current by >10%)
+            if val_m != cur_cf and 1_000_000 <= val_m <= 500_000_000:
+                adjustments['monthly_cash_flow'] = float(val_m)
+                break
+
+    # ── Loan amount ───────────────────────────────────────────
+    for pat in [
+        r'pinjaman\s+(?:ideal|aman|safe|maksimal|max)\s+.*?rp\s*(\d+(?:[,.]\d+)?)\s*(?:jt|juta|m\b)',
+        r'rp\s*(\d+(?:[,.]\d+)?)\s*(?:jt|juta)\s+(?:lebih aman|masih aman|safe|ideal)',
+        r'batas\s+aman\s+.*?rp\s*(\d+(?:[,.]\d+)?)\s*(?:jt|juta)',
+        r'ideal\s+loan\s+.*?rp\s*(\d+(?:[,.]\d+)?)\s*(?:jt|juta|m\b)',
+        r'max(?:imal)?\s+.*?rp\s*(\d+(?:[,.]\d+)?)\s*(?:jt|juta|m\b)',
+    ]:
+        m = re.search(pat, low)
+        if m:
+            raw_val = m.group(1)
+            try:
+                # Handle "128,6" (ID decimal) → 128.6 → 128_600_000
+                # Handle "128"   (integer)    → 128   → 128_000_000
+                if ',' in raw_val or '.' in raw_val:
+                    val_f = float(raw_val.replace(',', '.'))
+                else:
+                    val_f = float(raw_val)
+                val_m = int(val_f * 1_000_000)
+                cur_loan = raw_input.get('loan_rp', 50e6)
+                if val_m != cur_loan and 5_000_000 <= val_m <= 500_000_000:
+                    adjustments['loan_rp'] = float(val_m)
+                    break
+            except ValueError:
+                pass
+
+    # ── Duration / tenor ─────────────────────────────────────
+    for pat in [
+        r'tenor\s+(\d+)\s*bulan',
+        r'duration\s+(\d+)\s*months?',
+        r'perpanjang\s+tenor\s+(?:ke|jadi|to)\s*(\d+)',
+    ]:
+        m = re.search(pat, low)
+        if m:
+            val = int(m.group(1))
+            cur_dur = raw_input.get('duration', 24)
+            if val != cur_dur and 4 <= val <= 72:
+                adjustments['duration'] = float(val)
+                break
+
+    # ── Business age (only if explicitly recommended, not just mentioned) ──
+    for pat in [
+        r'bangun\s+bisnis\s+(?:selama\s+)?(\d+)\s*tahun',
+        r'business\s+age\s+(?:to|ke|jadi)\s*(\d+)',
+    ]:
+        m = re.search(pat, low)
+        if m:
+            val = int(m.group(1))
+            if 1 <= val <= 20:
+                adjustments['business_age_years'] = float(val)
+                break
+
+    return adjustments
+
+
 def _call_chat_llm(messages):
     """Cascade: OR tools → Groq tools → OR no-tools → Groq no-tools."""
     _or  = st.session_state.get("openrouter_key", "")
@@ -1660,13 +1776,24 @@ CARA BERPIKIR — CHAIN OF THOUGHT (lakukan ini secara SILENT sebelum menjawab):
                 "Vishesh roop se puchein: score kaise kam karein, jokhim kaarak, ideal rin, bachat tips." + err_note
             )
 
-    # Clean response and extract any remaining [ADJUST:] tags
+    # ── Final cleanup: clean response text + extract explicit [ADJUST:] tags ──
     if response:
         response, extra_adj = _clean_response(response)
         for field, val in extra_adj.items():
             if field not in adjustments:
                 adjustments[field] = val
 
+    # ── Semantic fallback: parse numeric values from natural LLM text ─────────
+    # Free-tier LLMs frequently ignore [ADJUST:] tag instructions even when
+    # the system prompt says to use them. This guarantees slider updates work
+    # even when the LLM gives a perfect natural-language recommendation but
+    # forgets to embed the tags. Only fills fields not already captured above.
+    if response and raw_input:
+        sem_adj = _extract_adjustments_semantic(response, raw_input)
+        for field, val in sem_adj.items():
+            if field not in adjustments:
+                adjustments[field] = val
+
     return response or "...", adjustments, _last_error
 
 # ============================================================
@@ -2351,4 +2478,4 @@ st.download_button(
     data=report_txt,
     file_name=T('download_file', lang),
     mime="text/plain"
-)
+)