blocks
text = re.sub(r'.*?', '', text, flags=re.DOTALL).strip()
# Extract + strip JSON tool-call objects {"field":..., "value":...}
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))
if field in valid_fields:
adjustments[field] = val
text = json_pattern.sub('', text).strip()
# Extract [ADJUST: field=value] tags
for m in re.finditer(r'\[ADJUST:\s*(\w+)\s*=\s*([\d.]+)\]', text):
field, val = m.group(1), float(m.group(2))
if field not in adjustments:
adjustments[field] = val
# Strip the tag itself (leave the surrounding sentence intact)
text = re.sub(r'\s*\[ADJUST:[^\]]+\]\s*', ' ', text).strip()
# โโ ORPHAN TRAILING SENTENCE STRIPPER (v6) โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# LLMs append unwanted recommendation sentences at the end of responses.
# Two generations of patterns observed in production:
#
# Gen 1 (Doc7 - after v4 fix): action-verb style
# "..., coba naikkan digital score ke 80 โ ini akan memberikan dampak..."
#
# Gen 2 (Doc8 - after Rule 9 ban): pivot-back-to-credit style
# "...aku bisa membantu kamu dengan menyarankan untuk meningkatkan X ke N..."
# "...Jadi, mari kita fokus pada membicarakan tentang cara meningkatkan X..."
# "...Atau, kalau kamu ingin kembali ke topik kredit, aku bisa menyarankan..."
#
# STRATEGY: run both passes, then strip_to_last_sentence as final safety net.
# โโ Pass 1: Gen 1 action-verb orphans โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
_orphan_id = re.compile(
r'[,\s]*coba\s+(?:naikkan|tingkatkan|optimalkan|kurangi|pertahankan|daftarkan|perbaiki)\s+'
r'[\w\s]+\s+ke\s+[\w\s./0-9]+(?:\s*[โโ-]\s*[^.]*)?\.?\s*$',
re.IGNORECASE | re.DOTALL
)
_orphan_id2 = re.compile(
r'[,\n]\s*(?:naikkan|tingkatkan|optimalkan|kurangi|pertahankan|daftarkan|perbaiki)\s+'
r'[\w\s]+\s+ke\s+[\w\s./0-9]+(?:\s*[โโ-]\s*[^.]*)?\.?\s*$',
re.IGNORECASE | re.DOTALL
)
_orphan_en = re.compile(
r'[,\s]*(?:try\s+)?(?:raise|increase|lower|reduce|optimize|improve|register)\s+'
r'[\w\s]+\s+to\s+[\w\s./0-9]+(?:\s*[โโ-]\s*[^.]*)?\.?\s*$',
re.IGNORECASE | re.DOTALL
)
for _pat in [_orphan_id, _orphan_id2, _orphan_en]:
text = _pat.sub('', text).strip()
# โโ Pass 2: Gen 2 pivot-back-to-credit orphans โโโโโโโโโโโโโโโโโโโโโโโโโโโ
_pivot_patterns = [
# "aku bisa membantu kamu dengan menyarankan untuk meningkatkan X ke N untuk mendapatkan Y"
re.compile(
r',?\s*aku\s+bisa\s+(?:membantu\s+kamu\s+)?dengan\s+menyarankan\s+untuk\s+meningkatkan\s+'
r'[\w\s]+\s+ke\s+\d+\s+untuk\s+mendapatkan[^.]*\.?\s*$',
re.IGNORECASE | re.DOTALL
),
# "meningkatkan digital score kamu ke N untuk mendapatkan pinjaman..."
re.compile(
r',?\s*(?:untuk\s+)?meningkatkan\s+(?:digital\s+)?(?:score|presence|skor)\s+kamu\s+ke\s+\d+'
r'\s+untuk\s+mendapatkan[^.]*\.?\s*$',
re.IGNORECASE | re.DOTALL
),
# "Jadi, mari kita fokus pada membicarakan tentang cara meningkatkan X ke N..."
re.compile(
r'[.\n]\s*(?:jadi|dan|tapi|,)?\s*(?:mari\s+kita\s+)?fokus\s+pada\s+'
r'(?:membicarakan|cara)\s+(?:tentang\s+)?(?:cara\s+)?meningkatkan\s+[\w\s]+\s+ke\s+\d+[^.]*\.?\s*$',
re.IGNORECASE | re.DOTALL
),
# "Atau/Tapi, kalau kamu ingin kembali ke topik kredit..."
re.compile(
r'[.\n]\s*(?:atau|tapi|jadi|dan|,)?\s*kalau\s+kamu\s+(?:ingin|mau|bisa)\s+kembali\s+ke\s+'
r'topik\s+(?:kredit|keuangan|bisnis)[^.]*\.?\s*$',
re.IGNORECASE | re.DOTALL
),
# "Tapi, kalau kamu adalah seorang pengusaha UMKM yang ingin mengajukan kredit, aku bisa..."
re.compile(
r',?\s*tapi,?\s+kalau\s+kamu\s+(?:adalah\s+seorang|ingin\s+mengajukan)\s+[^.]{0,120}\.?\s*$',
re.IGNORECASE | re.DOTALL
),
]
for _pat in _pivot_patterns:
text = _pat.sub('', text).strip()
# โโ Pass 3: strip_to_last_sentence โ final safety net โโโโโโโโโโโโโโโโโโโโ
# If text ends without sentence-ending punctuation (incomplete fragment),
# cut back to the last complete sentence. Handles any new orphan variant.
def _strip_to_last_sentence(t):
t = t.strip()
if not t or t[-1] in '.!?':
return t
last_end = max(t.rfind('.'), t.rfind('!'), t.rfind('?'))
if last_end > len(t) * 0.25: # Don't cut more than 75% of response
return t[:last_end + 1].strip()
return t
text = _strip_to_last_sentence(text)
# Cleanup
text = re.sub(r' +', ' ', text)
text = re.sub(r' ([,.])', r'\1', text)
text = re.sub(r'\n\*\*\s*\*\*\s*$', '', text, flags=re.MULTILINE).strip()
text = re.sub(r'\n{3,}', '\n\n', text).strip()
text = re.sub(r'[,\sโโ-]+$', '', 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 โ only fills in fields not already captured
by explicit [ADJUST:] tags or formal tool calls.
Examples handled:
"naikkan digital score ke 75" โ digital_presence_score=75
"digital score jadi 80" โ digital_presence_score=80
"Rp 25jt/bln" near "cash flow" โ monthly_cash_flow=25_000_000
"pinjaman ideal Rp 128jt" โ loan_rp=128_000_000
"tenor 36 bulan" โ duration=36
"""
if not text:
return {}
adjustments = {}
low = text.lower()
# โโ Digital presence score โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
for pat in [
r'digital\s+score\s+(?:ke|jadi|to|โ|->|=)\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
# โโ Monthly cash flow โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# Match "Rp NNjt" or "NNjt" near cash flow context
for pat in [
r'cash\s+flow\s+(?:ke|jadi|to|โ)\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
# โโ Loan amount โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
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) โ 128.6 โ 128_600_000
# Handle "128" (integer) โ 128 โ 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
# โโ Duration / tenor โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
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
# โโ Business age (only if explicitly recommended, not just mentioned) โโ
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 โ Groq tools โ OR no-tools โ Groq no-tools."""
_or = st.session_state.get("openrouter_key", "")
_grq = st.session_state.get("groq_key", "")
if _or:
text, tool_calls, err = call_openrouter_tools(messages, _or, MAYA_TOOLS)
if text or tool_calls:
return text, tool_calls, 'OpenRouter (tool-calling)', None
if _grq:
text, tool_calls, err = call_groq_tools(messages, _grq, MAYA_TOOLS)
if text or tool_calls:
return text, tool_calls, 'Groq (tool-calling)', None
if _or:
text, tool_calls, err = call_openrouter_tools(messages, _or, tools=None)
if text:
return text, [], 'OpenRouter', err
if _grq:
text, tool_calls, err = call_groq_tools(messages, _grq, tools=None)
if text:
return text, [], 'Groq', err
return None, [], None, "No LLM available"
def _parse_tool_calls(tool_calls):
"""Parse formal tool_calls from LLM response into adjustments dict."""
adjustments = {}
for tc in (tool_calls or []):
try:
fn = tc.get('function', {})
if fn.get('name') == 'adjust_whatif_slider':
args = json.loads(fn.get('arguments', '{}'))
field = args.get('field')
value = args.get('value')
if field and value is not None:
adjustments[field] = float(value)
except Exception:
pass
return adjustments
def _tool_call_to_text(tool_calls, lang):
"""Convert tool calls to human-readable confirmation text."""
if not tool_calls:
return ""
field_labels = {
'digital_presence_score': {'id':'Digital Score','en':'Digital Score','hi':'Digital Score'},
'business_age_years': {'id':'Umur Bisnis', 'en':'Business Age', 'hi':'Vyapaar Aayu'},
'num_employees': {'id':'Karyawan', 'en':'Employees', 'hi':'Karmachaaree'},
'monthly_cash_flow': {'id':'Cash Flow', 'en':'Cash Flow', 'hi':'Naqad Pravaah'},
'duration': {'id':'Tenor', 'en':'Duration', 'hi':'Avadhi'},
'loan_rp': {'id':'Pinjaman', 'en':'Loan', 'hi':'Rin'},
}
parts = []
for tc in tool_calls:
try:
fn = tc.get('function', {})
if fn.get('name') == 'adjust_whatif_slider':
args = json.loads(fn.get('arguments', '{}'))
field = args.get('field', '')
value = args.get('value', 0)
reason = args.get('reason', '')
fl = field_labels.get(field, {}).get(lang, field)
if field in ('monthly_cash_flow', 'loan_rp'):
val_str = "Rp " + str(int(float(value) / 1e6)) + "jt"
else:
val_str = str(int(float(value)))
reason_str = " โ " + reason if reason else ""
parts.append("Rekomendasi ubah **" + fl + "** ke **" + val_str + "**" + reason_str + ".")
except Exception:
pass
if parts:
suffix = {
'id': 'What-If sliders sudah diupdate!',
'en': 'What-If sliders updated!',
'hi': 'What-If sliders update ho gaye!',
}[lang]
return '\n'.join(parts) + '\n\n' + suffix
return ""
# ============================================================
# MEMORY PERSISTENCE
# ============================================================
def _get_session_id():
if 'session_id' not in st.session_state:
import random
st.session_state.session_id = hashlib.md5(
(str(_time.time()) + str(random.random())).encode()
).hexdigest()[:10]
return st.session_state.session_id
def _save_chat_memory(history, summary=''):
try:
sid = _get_session_id()
data = {"history": history[-30:], "summary": summary}
with open("/tmp/chat_mem_" + sid + ".json", 'w', encoding='utf-8') as f:
json.dump(data, f, ensure_ascii=False)
except Exception:
pass
def _load_chat_memory():
try:
sid = _get_session_id()
path = "/tmp/chat_mem_" + sid + ".json"
if os.path.exists(path):
with open(path, encoding='utf-8') as f:
data = json.load(f)
return data.get("history", []), data.get("summary", "")
except Exception:
pass
return [], ""
# ============================================================
# MEMORY SUMMARIZATION
# ============================================================
def _summarize_history(history, lang):
if len(history) <= 8:
return history, st.session_state.get('chat_summary', '')
old_turns = history[:-6]
recent_turns = history[-6:]
prev_summary = st.session_state.get('chat_summary', '')
lang_str = {'id':'Bahasa Indonesia','en':'English','hi':'Hindi'}
prompt_lines = [
"Summarize this conversation in 3-4 sentences in " + lang_str.get(lang,'Bahasa Indonesia') + ".",
"Preserve: key questions asked, recommendations given, any slider values suggested.",
]
if prev_summary:
prompt_lines.append("Previous summary to extend: " + prev_summary)
prompt_lines.append("")
for m in old_turns:
prompt_lines.append(m['role'].upper() + ": " + m['content'][:160])
r, _, _ = _call_llm([{"role": "user", "content": '\n'.join(prompt_lines)}])
fallback = {
'id': "[Ringkasan " + str(len(old_turns)//2) + " topik sebelumnya tidak tersedia]",
'en': "[Summary of " + str(len(old_turns)//2) + " earlier topics unavailable]",
'hi': "[Pehle ke " + str(len(old_turns)//2) + " vishay ka saaraansh upalabdh nahi]",
}
new_summary = r.strip() if r else (prev_summary or fallback[lang])
st.session_state.chat_summary = new_summary
return recent_turns, new_summary
# Early session state init
for _k, _v in [('llm_status', ''), ('api_log', [])]:
if _k not in st.session_state:
st.session_state[_k] = _v
# ============================================================
# SIDEBAR
# ============================================================
with st.sidebar:
st.markdown("## ๐ฆ SME Credit Risk AI")
st.markdown("---")
if 'lang_sel' not in st.session_state:
st.session_state.lang_sel = 'id'
lang = st.radio(
"๐ Language / Bahasa / Roman Hindi",
['id','en','hi'],
format_func=lambda x: LANG_LABELS[x],
key='lang_sel'
)
st.markdown("---")
openrouter_key = st.session_state.openrouter_key
groq_key = st.session_state.groq_key
st.caption(
"๐ **OpenRouter:** " + ('โ
' if openrouter_key else 'โ') +
" **Groq:** " + ('โ
' if groq_key else 'โ')
)
if not any([openrouter_key, groq_key]):
st.warning("โ ๏ธ Tidak ada API key. Tambahkan OPENROUTER_API_KEY / GROQ_API_KEY di HF Secrets.")
# Live LLM status placeholder โ updated by _set_status() during calls
_llm_ph = st.empty()
st.session_state['_llm_ph'] = _llm_ph
if st.session_state.get('llm_status'):
try:
_llm_ph.info(st.session_state.llm_status, icon="โณ")
except Exception:
pass
st.markdown("---")
st.markdown("### ๐ก API Tracker")
_active_status = st.session_state.get('llm_status', '')
if _active_status:
st.markdown(
""
"โณ " + _active_status + "
",
unsafe_allow_html=True
)
if st.session_state.api_log:
for entry in st.session_state.api_log[-5:]:
icon = "โ
" if entry['ok'] else "โ"
st.caption(icon + " `" + entry['step'] + "` โ " + entry['src'] + " (" + str(entry['ms']) + "ms)")
else:
st.caption("_Belum ada aktivitas API_")
if st.button("๐๏ธ Reset Log", key="reset_api_log", use_container_width=True):
st.session_state.api_log = []
st.rerun()
st.markdown("---")
st.markdown(T('sidebar_stack', lang))
stack_items = {
'id': "- ๐ค XGBoost + LightGBM + RF\n- ๐ SHAP Eksplainabilitas\n- ๐ฌ Narasi AI (OpenRouter + Groq)\n- ๐ค Maya AI Chat + Tool Calling\n- ๐ง Chain-of-Thought + Few-shot\n- ๐พ Memory Summarization\n- ๐ ID / EN / HI",
'en': "- ๐ค XGBoost + LightGBM + RF\n- ๐ SHAP Explainability\n- ๐ฌ AI Narrative (OpenRouter + Groq)\n- ๐ค Maya AI Chat + Tool Calling\n- ๐ง Chain-of-Thought + Few-shot\n- ๐พ Memory Summarization\n- ๐ ID / EN / HI",
'hi': "- ๐ค XGBoost + LightGBM + RF\n- ๐ SHAP Vyaakhyaa\n- ๐ฌ AI Vivarana (OpenRouter + Groq)\n- ๐ค Maya AI Chat + Tool Calling\n- ๐ง Chain-of-Thought + Few-shot\n- ๐พ Memory Saaraansh\n- ๐ ID / EN / HI",
}
st.markdown(stack_items[lang])
st.caption(T('sidebar_footer', lang))
# ============================================================
# LOAD MODEL & RAG
# ============================================================
@st.cache_resource(show_spinner="Loading AI model...")
def load_model():
d = 'model'
try:
ensemble = joblib.load(d + '/ensemble_model.pkl')
xgb_model = joblib.load(d + '/xgb_model.pkl')
scaler = joblib.load(d + '/scaler.pkl')
feature_names = joblib.load(d + '/feature_names.pkl')
explainer = shap.TreeExplainer(xgb_model)
meta = {}
if os.path.exists(d + '/metadata.json'):
with open(d + '/metadata.json') as f:
meta = json.load(f)
return ensemble, scaler, feature_names, explainer, meta
except FileNotFoundError:
return None, None, None, None, {}
@st.cache_resource(show_spinner="Loading knowledge base...")
def load_rag_index():
try:
from rag import load_index
return load_index(kb_path="knowledge_base", index_path="/tmp/kb_index.pkl")
except ImportError:
st.sidebar.warning("rag.py not found. RAG disabled.")
return None
except Exception as e:
st.sidebar.warning("RAG load failed: " + str(e))
return None
ensemble, scaler, feature_names, explainer, meta = load_model()
rag_index = load_rag_index()
if rag_index:
chunk_count = len(rag_index.get('chunks', []))
st.sidebar.success("Knowledge base: " + str(chunk_count) + " chunks")
else:
st.sidebar.warning("Knowledge base not loaded. RAG disabled.")
if ensemble is None:
st.markdown(
'',
unsafe_allow_html=True
)
st.error(T('no_model', lang))
st.stop()
# ============================================================
# SESSION STATE INIT
# ============================================================
_defaults = dict(
result=None, shap_vals=None, raw_input=None,
narrative='', llm_src='', narrative_lang='',
chat_history=[], shap_png=None,
wi_dig=None, wi_biz=None, wi_emp=None,
wi_cash=None, wi_dur=None, wi_loan=None,
chat_summary='', memory_loaded=False,
llm_status='', api_log=[],
)
for k, v in _defaults.items():
if k not in st.session_state:
st.session_state[k] = v
if not st.session_state.memory_loaded:
loaded_history, loaded_summary = _load_chat_memory()
if loaded_history and not st.session_state.chat_history:
st.session_state.chat_history = loaded_history
if loaded_summary and not st.session_state.chat_summary:
st.session_state.chat_summary = loaded_summary
st.session_state.memory_loaded = True
# ============================================================
# HELPER FUNCTIONS
# ============================================================
def preprocess(raw, scaler, feature_names):
df = pd.DataFrame([raw])
ohe = pd.get_dummies(df)
for col in feature_names:
if col not in ohe.columns:
ohe[col] = 0
ohe = ohe[feature_names].apply(pd.to_numeric, errors='coerce').fillna(0)
return pd.DataFrame(scaler.transform(ohe), columns=feature_names)
def risk_result(pd_score, loan_rp, lgd=0.40):
el = pd_score * lgd * loan_rp
if pd_score < 0.20:
css, color = 'risk-low', '#11998e'
cat = {l: T('risk_approved', l) for l in ['id','en','hi']}
elif pd_score < 0.50:
css, color = 'risk-med', '#f7971e'
cat = {l: T('risk_review', l) for l in ['id','en','hi']}
else:
css, color = 'risk-high', '#e74c3c'
cat = {l: T('risk_high', l) for l in ['id','en','hi']}
return dict(pd=pd_score, el=el, lgd=lgd, ead=loan_rp, css=css, color=color, cat=cat)
def shap_summary(vals, names, n=5):
idx = np.argsort(np.abs(vals))[-n:][::-1]
return '\n'.join([
"- " + names[i] + ": " + ('increases' if vals[i] > 0 else 'decreases') +
" risk (SHAP=" + str(round(vals[i], 3)) + ")"
for i in idx
])
def make_shap_png(shap_vals, feature_names):
"""Returns base64 data URI โ avoids Streamlit MediaFileStorage expiry."""
sv = np.array(shap_vals)
names = list(feature_names)
n = min(12, len(sv))
idx = np.argsort(np.abs(sv))[-n:]
fig, ax = plt.subplots(figsize=(9, 5))
colors = ['#ef4444' if sv[i] > 0 else '#3b82f6' for i in idx]
ax.barh(range(n), sv[idx], color=colors, height=0.6)
ax.set_yticks(range(n))
ax.set_yticklabels([names[i] for i in idx], fontsize=8)
ax.axvline(0, color='black', linewidth=0.8)
ax.set_xlabel('SHAP value')
ax.set_title('Feature Impact โ SHAP', fontweight='bold', fontsize=10)
plt.tight_layout()
buf = io.BytesIO()
fig.savefig(buf, format='png', dpi=130, bbox_inches='tight')
plt.close(fig)
buf.seek(0)
b64 = base64.b64encode(buf.read()).decode()
return 'data:image/png;base64,' + b64
# ============================================================
# LLM HELPERS โ Narrative cascade
# ============================================================
_OR_FREE_MODELS = [
"google/gemini-2.0-flash-exp:free",
"qwen/qwen3-32b:free",
"meta-llama/llama-3.3-70b-instruct:free",
"qwen/qwen2.5-72b-instruct:free",
"deepseek/deepseek-chat-v3-0324:free",
"microsoft/phi-4:free",
"mistralai/mistral-7b-instruct:free",
]
_GROQ_FREE_MODELS = [
"llama-3.3-70b-versatile",
"llama-3.1-8b-instant",
"gemma2-9b-it",
"mixtral-8x7b-32768",
]
def call_openrouter(messages, api_key, model=None):
if not api_key:
return None, "No OpenRouter API key"
models_to_try = [model] if model else _OR_FREE_MODELS
last_error = None
for m in models_to_try:
t0 = _time.time()
try:
resp = requests.post(
"https://openrouter.ai/api/v1/chat/completions",
headers={
"Authorization": "Bearer " + api_key,
"Content-Type": "application/json",
"HTTP-Referer": "https://huggingface.co",
"X-Title": "SME Credit Risk AI"
},
json={"model": m, "messages": messages, "max_tokens": 600, "temperature": 0.75},
timeout=25,
)
ms = int((_time.time() - t0) * 1000)
if resp.status_code == 200:
short = m.split('/')[-1].replace(':free','')
_log_api("Narasi", "OR/" + short, True, ms)
return resp.json()['choices'][0]['message']['content'], None
short = m.split('/')[-1].replace(':free','')
last_error = "HTTP " + str(resp.status_code) + " (" + m + "): " + resp.text[:150]
_log_api("Narasi", "OR/" + short, False, ms)
except requests.exceptions.Timeout:
ms = int((_time.time() - t0) * 1000)
last_error = "Timeout (" + m + ")"
short = m.split('/')[-1].replace(':free','')
_log_api("Narasi", "OR/" + short, False, ms)
except Exception as e:
ms = int((_time.time() - t0) * 1000)
last_error = m + ": " + str(e)[:100]
short = m.split('/')[-1].replace(':free','')
_log_api("Narasi", "OR/" + short, False, ms)
return None, last_error
def call_groq(messages, api_key):
if not api_key:
return None, "No Groq API key"
last_error = None
for m in _GROQ_FREE_MODELS:
t0 = _time.time()
try:
resp = requests.post(
"https://api.groq.com/openai/v1/chat/completions",
headers={
"Authorization": "Bearer " + api_key,
"Content-Type": "application/json"
},
json={"model": m, "messages": messages, "max_tokens": 600, "temperature": 0.75},
timeout=20,
)
ms = int((_time.time() - t0) * 1000)
if resp.status_code == 200:
_log_api("Narasi", "Groq/" + m, True, ms)
return resp.json()['choices'][0]['message']['content'], None
last_error = "HTTP " + str(resp.status_code) + " (" + m + "): " + resp.text[:150]
_log_api("Narasi", "Groq/" + m, False, ms)
except requests.exceptions.Timeout:
ms = int((_time.time() - t0) * 1000)
last_error = "Timeout (" + m + ")"
_log_api("Narasi", "Groq/" + m, False, ms)
except Exception as e:
ms = int((_time.time() - t0) * 1000)
last_error = m + ": " + str(e)[:100]
_log_api("Narasi", "Groq/" + m, False, ms)
return None, last_error
def _call_llm(messages):
"""Narrative LLM: OpenRouter free โ Groq free โ None."""
_or = st.session_state.get("openrouter_key", "")
_grq = st.session_state.get("groq_key", "")
last_err = None
if _or:
r, last_err = call_openrouter(messages, _or)
if r:
model_used = _OR_FREE_MODELS[0].split("/")[-1].replace(":free","")
return r, 'OpenRouter (' + model_used + ')', None
if _grq:
r, last_err = call_groq(messages, _grq)
if r:
model_used = _GROQ_FREE_MODELS[0]
return r, 'Groq (' + model_used + ')', None
return None, None, last_err or "No API key configured"
# ============================================================
# NARRATIVE GENERATION
# ============================================================
def get_narrative(shap_vals, feature_names, result, lang, raw_input):
pd_pct = result['pd'] * 100
factors = shap_summary(shap_vals, feature_names, 5)
lang_str = {'id':'Bahasa Indonesia profesional','en':'Professional English','hi':'Hindi'}
prompt = (
"You are a senior credit analyst for Indonesian SME lending.\n"
"Reply ONLY in " + lang_str.get(lang,'English') + ".\n"
"Applicant: PD=" + str(round(pd_pct,1)) + "%, EL=Rp " + str(int(result['el'])) + ", "
"BizAge=" + str(raw_input.get('business_age_years')) + "yr, "
"Employees=" + str(raw_input.get('num_employees')) + ", "
"DigitalScore=" + str(raw_input.get('digital_presence_score')) + "/100, "
"CashFlow=Rp " + str(int(raw_input.get('monthly_cash_flow',0))) + "/mo, "
"NPWP=" + ('yes' if raw_input.get('has_npwp') else 'no') + ", "
"SIUP=" + ('yes' if raw_input.get('has_siup') else 'no') + "\n"
"Top SHAP factors:\n" + factors + "\n"
"Write: 1) " + T('n_risk_summary',lang) + " (1-2 sentences) "
"2) " + T('n_strengths',lang) + " (2-3 bullets) "
"3) " + T('n_risks',lang) + " (2-3 bullets) "
"4) " + T('n_recommendations',lang) + " (2-3 specific steps). "
"Be data-driven and reference actual values."
)
msgs = [{"role": "user", "content": prompt}]
r, src, _ = _call_llm(msgs)
if r:
return r, src
# Static fallback
biz = raw_input.get('business_age_years', 0)
emp = raw_input.get('num_employees', 0)
dig = raw_input.get('digital_presence_score', 0)
cf = raw_input.get('monthly_cash_flow', 0)
npwp = raw_input.get('has_npwp', 0)
siup = raw_input.get('has_siup', 0)
dur = raw_input.get('duration', 24)
loan = raw_input.get('loan_rp', 50e6)
ch = raw_input.get('credit_history', '')
def _t(i, e, h):
return {'id': i, 'en': e, 'hi': h}[lang]
if pd_pct < 20:
status = _t('risiko rendah โ **direkomendasikan disetujui**',
'low risk โ **recommended for approval**',
'kam jokhim โ **swikriti anushansit**')
elif pd_pct < 50:
status = _t('risiko sedang โ **perlu review**',
'moderate risk โ **review required**',
'madhyam jokhim โ **sameeksha zaroori**')
else:
status = _t('risiko tinggi โ **butuh jaminan/penolakan**',
'high risk โ **collateral or rejection recommended**',
'uchch jokhim โ **zamanat ya aswikriti anushansit**')
s = []
if biz >= 5:
s.append(_t("Bisnis " + str(int(biz)) + " tahun โ track record terbukti",
str(int(biz)) + "-year business โ proven track record",
str(int(biz)) + " saal ka vyapaar โ siddh track record"))
elif biz >= 3:
s.append(_t("Bisnis " + str(int(biz)) + " tahun โ melewati fase startup",
str(int(biz)) + "-year business โ past startup phase",
str(int(biz)) + " saal ka vyapaar โ startup charan paar"))
if emp >= 10:
s.append(_t(str(emp) + " karyawan โ skala usaha signifikan",
str(emp) + " employees โ significant scale",
str(emp) + " karmachaaree โ mahatvapurn paimaana"))
elif emp >= 5:
s.append(_t(str(emp) + " karyawan โ tim solid",
str(emp) + " employees โ solid team",
str(emp) + " karmachaaree โ mazboot team"))
if dig >= 70:
s.append(_t("Digital score " + str(dig) + "/100 โ kehadiran online sangat aktif",
"Digital score " + str(dig) + "/100 โ very active online presence",
"Digital score " + str(dig) + "/100 โ bahut active online upasthiti"))
elif dig >= 50:
s.append(_t("Digital score " + str(dig) + "/100 โ kehadiran digital cukup",
"Digital score " + str(dig) + "/100 โ decent digital presence",
"Digital score " + str(dig) + "/100 โ theek-thaak online"))
if cf >= 15e6:
s.append(_t("Cash flow Rp " + str(int(cf/1e6)) + "jt/bln โ mendukung kemampuan bayar",
"Cash flow Rp " + str(int(cf/1e6)) + "M/month โ supports repayment",
"Naqad pravaah Rp " + str(int(cf/1e6)) + "M/maah โ bhugtaan samarthan"))
if npwp:
s.append(_t("NPWP terverifikasi โ kepatuhan pajak terbukti",
"NPWP verified โ tax compliance proven",
"NPWP satthapit โ kar anupalan siddh"))
if siup:
s.append(_t("SIUP/NIB aktif โ legalitas usaha terpenuhi",
"SIUP/NIB active โ business license verified",
"SIUP/NIB sakriy โ vyapaar laaisens satthapit"))
if 'existing paid' in ch or 'all paid' in ch:
s.append(_t("Riwayat kredit baik โ tidak ada tunggakan",
"Good credit history โ no defaults",
"Achha credit itihaas โ koi chook nahi"))
r_list = []
if biz < 2:
r_list.append(_t("Bisnis sangat baru (" + str(int(biz)) + " thn) โ belum ada track record",
"Very young business (" + str(int(biz)) + " yr) โ no track record",
"Bahut naya vyapaar (" + str(int(biz)) + " saal) โ koi track record nahi"))
elif biz < 3:
r_list.append(_t("Bisnis " + str(int(biz)) + " thn โ masih di fase awal",
str(int(biz)) + "-year business โ early growth stage",
str(int(biz)) + " saal ka vyapaar โ prarambhik charan"))
if dig < 30:
r_list.append(_t("Digital score rendah (" + str(dig) + "/100) โ minim online",
"Low digital score (" + str(dig) + "/100) โ minimal presence",
"Kam digital score (" + str(dig) + "/100) โ nyoonatam upasthiti"))
elif dig < 50:
r_list.append(_t("Digital score " + str(dig) + "/100 โ di bawah rata-rata",
"Digital score " + str(dig) + "/100 โ below average",
"Digital score " + str(dig) + "/100 โ ausat se neeche"))
if cf < 5e6:
r_list.append(_t("Cash flow Rp " + str(int(cf/1e6)) + "jt/bln โ kemampuan bayar diragukan",
"Cash flow Rp " + str(int(cf/1e6)) + "M/month โ repayment questionable",
"Naqad pravaah Rp " + str(int(cf/1e6)) + "M/maah โ bhugtaan sandehaapad"))
elif cf < 10e6:
r_list.append(_t("Cash flow Rp " + str(int(cf/1e6)) + "jt/bln โ rasio cicilan mungkin ketat",
"Cash flow Rp " + str(int(cf/1e6)) + "M/month โ installment ratio tight",
"Naqad pravaah Rp " + str(int(cf/1e6)) + "M/maah โ kist anupaat tang"))
if not npwp:
r_list.append(_t("NPWP belum ada โ wajib untuk pinjaman formal",
"No NPWP โ required for formal loans",
"NPWP nahi โ aupchaarik rin ke liye zaroori"))
if not siup:
r_list.append(_t("SIUP/NIB belum ada โ legalitas perlu diperkuat",
"No SIUP/NIB โ legality needs strengthening",
"SIUP/NIB nahi โ vaaneeya ko mazboot karna hoga"))
if 'delayed' in ch or 'critical' in ch:
r_list.append(_t("Riwayat kredit bermasalah โ sinyal negatif kuat",
"Problematic credit history โ strong negative signal",
"Samasyaagrast credit itihaas โ nkaraatmak sanket"))
est_inst = loan / dur if dur > 0 else loan
if cf > 0 and (est_inst / cf) > 0.40:
r_list.append(_t(
"Rasio cicilan " + str(int(est_inst/cf*100)) + "% โ di atas batas aman 40%",
"Installment ratio " + str(int(est_inst/cf*100)) + "% โ exceeds 40% safe limit",
"Kist anupaat " + str(int(est_inst/cf*100)) + "% โ 40% seema se adhik"
))
rec = []
if not npwp:
rec.append(_t("Urus NPWP โ daftar di pajak.go.id",
"Register NPWP โ via pajak.go.id",
"NPWP register karein โ pajak.go.id par"))
if dig < 60:
rec.append(_t("Naikkan digital score ke 60+ โ Google Business + marketplace aktif",
"Raise digital score to 60+ โ Google Business + active marketplace",
"Digital score 60+ karein โ Google Business + marketplace"))
if cf < 15e6:
rec.append(_t("Optimalkan cash flow โ dokumentasikan semua pemasukan",
"Optimize cash flow โ document all income streams",
"Naqad pravaah optimize karein โ sabhi aay document karein"))
if loan > 0 and dur > 0 and cf > 0 and (loan / dur / cf) > 0.40:
safe = cf * dur * 0.35
rec.append(_t(
"Pertimbangkan pinjaman Rp " + str(int(safe/1e6)) + "jt โ lebih aman",
"Consider loan of Rp " + str(int(safe/1e6)) + "M โ safer option",
"Rp " + str(int(safe/1e6)) + "M ka rin sochein โ surakshit vikalp"
))
rec.append(_t("Gunakan tab What-If โ simulasikan sebelum mengajukan ulang",
"Use the What-If tab โ simulate before reapplying",
"What-If tab use karein โ dobara apply se pehle simulate karein"))
if not s:
s = [_t("Profil sedang dievaluasi", "Profile under evaluation", "Parichay mulyankan mein")]
if not r_list:
r_list = [_t("Tidak ada risiko signifikan",
"No significant risk factors detected",
"Koi mahatvapurn jokhim nahi")]
el_fmt = result['el'] / 1e6
intro = _t(
"Skor PD **" + str(round(pd_pct,1)) + "%** menunjukkan " + status +
". Expected Loss diestimasi **Rp " + str(round(el_fmt,2)) + "jt** (PD ร LGD " +
str(int(result['lgd']*100)) + "% ร EAD Rp " + str(int(result['ead']/1e6)) + "jt).",
"PD score of **" + str(round(pd_pct,1)) + "%** indicates " + status +
". Expected Loss estimated at **Rp " + str(round(el_fmt,2)) + "M** (PD ร LGD " +
str(int(result['lgd']*100)) + "% ร EAD Rp " + str(int(result['ead']/1e6)) + "M).",
"PD score **" + str(round(pd_pct,1)) + "%** darshata hai " + status +
". Anumaanit haani **Rp " + str(round(el_fmt,2)) + "M** (PD ร LGD " +
str(int(result['lgd']*100)) + "% ร EAD Rp " + str(int(result['ead']/1e6)) + "M)."
)
txt = (
"**" + T('n_risk_summary', lang) + "**\n" + intro + "\n\n"
"**" + T('n_strengths', lang) + "**\n" + '\n'.join(s[:3]) +
"\n\n**" + T('n_risks', lang) + "**\n" + '\n'.join(r_list[:3]) +
"\n\n**" + T('n_recommendations', lang) + "**\n" + '\n'.join(rec[:3])
)
return txt, 'Smart Template'
# ============================================================
# HELPER: top issue for fallback chat
# ============================================================
def _get_top_issue(raw_input, pd_pct, lang):
dig = raw_input.get('digital_presence_score', 0)
cf = raw_input.get('monthly_cash_flow', 0)
npwp = raw_input.get('has_npwp', 0)
biz = raw_input.get('business_age_years', 0)
def _t(i, e, h):
return {'id': i, 'en': e, 'hi': h}[lang]
if not npwp:
return _t("prioritas utama: **urus NPWP** dulu di pajak.go.id",
"top priority: **register NPWP** at pajak.go.id",
"mukhya prathamikta: **NPWP register karein** pajak.go.id par")
elif dig < 40:
return _t("**digital score " + str(dig) + "/100** adalah area paling kritis untuk ditingkatkan",
"**digital score " + str(dig) + "/100** is the most critical area to improve",
"**digital score " + str(dig) + "/100** sabse mahatvapurn sudhaar kshetra hai")
elif cf < 10e6:
return _t("**cash flow Rp " + str(int(cf/1e6)) + "jt/bln** perlu dioptimalkan",
"**cash flow Rp " + str(int(cf/1e6)) + "M/month** needs optimization",
"**naqad pravaah Rp " + str(int(cf/1e6)) + "M/maah** optimize zaroori hai")
elif biz < 2:
return _t("bisnis yang baru **" + str(int(biz)) + " tahun** jadi faktor risiko utama",
"**" + str(int(biz)) + "-year** business age is the main risk factor",
"**" + str(int(biz)) + " saal** ka vyapaar mukhya jokhim kaarak hai")
elif pd_pct < 20:
return _t("profil kamu sudah sangat bagus!",
"your profile is already excellent!",
"aapka parichay pehle se behtareen hai!")
else:
return _t("beberapa area bisa diperbaiki โ tanya aku lebih spesifik!",
"several areas can be improved โ ask me specifically!",
"kai kshetra sudhaare jaa sakte hain โ vishesh roop se puchein!")
# ============================================================
# AI CHAT โ Maya persona (FULL โ CoT + Few-shot + Tool Calling)
# ============================================================
def get_chat_response(user_msg, history, result, raw_input, shap_vals, feature_names, lang, rag_index=None):
pd_pct = result['pd'] * 100
factors = shap_summary(shap_vals, feature_names, 5)
lang_full = {'id': 'Bahasa Indonesia', 'en': 'English', 'hi': 'Hindi'}
# โโ Memory summarization โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
recent_history, chat_summary = _summarize_history(history, lang)
# โโ RAG context retrieval โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
context = ""
if rag_index:
try:
from rag import retrieve
retrieved = retrieve(user_msg, rag_index, k=2)
if retrieved:
context = "\nINFORMASI TAMBAHAN DARI KNOWLEDGE BASE:\n" + retrieved + "\n"
except Exception:
context = ""
# โโ Few-shot examples (RESTORED from Doc3) โโโโโโโโโโโโโโโโ
fewshot = """
CONTOH DIALOG MAYA (few-shot โ 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 [ADJUST: digital_presence_score=75] โ ini faktor SHAP terbesar. 2) **Optimalkan cash flow ke Rp 25jt/bln** [ADJUST: monthly_cash_flow=25000000] โ 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) โ bisnis muda = risiko lebih tinggi di mata model, dan **digital_presence_score** (SHAP=+0.189) โ 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 โ masih aman! Kalau mau tambah, pastikan cash flow naik dulu."
---
"""
# โโ Summary context block โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
summary_block = ""
if chat_summary:
lbl = {
'id': 'Ringkasan percakapan sebelumnya',
'en': 'Earlier conversation summary',
'hi': 'Pehle ki baatcheet ka saaraansh',
}
summary_block = "\n" + lbl.get(lang, lbl['en']) + ":\n" + chat_summary + "\n"
# โโ Chain-of-Thought instructions (RESTORED from Doc3) โโโโ
cot_block = """
CARA BERPIKIR โ CHAIN OF THOUGHT (lakukan ini secara SILENT sebelum menjawab):
โ Apa yang sebenarnya ditanyakan user? (bukan hanya kata-katanya โ cari intent sebenarnya)
โก Data pemohon mana yang paling relevan? (PD score, cash flow, digital score, NPWP, dll)
โข Apa saran paling KONKRET & berdampak tinggi untuk UMKM spesifik ini?
โฃ Apakah perlu adjust slider What-If? Jika ya, embed tag [ADJUST: field=value] LANGSUNG di dalam kalimat rekomendasi.
โ Tulis HANYA jawaban akhir โ jangan tampilkan proses berpikir di output.
"""
# โโ Full persona system prompt โโโโโโโโโโโโโโโโโโโโโโโโโโโโ
system = (
"Kamu adalah Maya, AI Credit Advisor yang cerdas, hangat, dan sedikit kasual untuk UMKM Indonesia.\n"
"Kepribadian: ramah tapi profesional, suka pakai analogi sederhana, TIDAK kaku, bisa ngobrol bebas, humoris kalau situasinya santai.\n"
"Bahasa: gunakan " + lang_full.get(lang,'Bahasa Indonesia') + " yang natural โ boleh semi-formal, boleh pakai slang ringan (dong/nih/ya/banget/sih/kan/kok).\n\n"
"PEMAHAMAN BAHASA INDONESIA:\n"
"- Pahami kata ambigu/homofon dengan tepat sesuai konteks percakapan:\n"
" * 'beruang' bisa = hewan beruang (bear) ATAU ber-uang (punya uang) โ baca konteks!\n"
" * 'bisa' bisa = mampu ATAU racun ular\n"
" * 'bunga' bisa = bunga tanaman ATAU bunga kredit\n"
"- Kalau pesannya lucu/absurd, ikuti humor-nya dengan ringan sambil tetap hubungkan ke konteks kredit UMKM.\n\n"
"DATA PEMOHON SAAT INI:\n"
"- PD Score: " + str(round(pd_pct,1)) + "% โ " + ('LAYAK' if pd_pct < 20 else 'PERLU REVIEW' if pd_pct < 50 else 'RISIKO TINGGI') + "\n"
"- Expected Loss: Rp " + str(round(result['el']/1e6, 2)) + "M\n"
"- Umur Bisnis: " + str(raw_input.get('business_age_years')) + " tahun | Karyawan: " + str(raw_input.get('num_employees')) + " orang\n"
"- Digital Score: " + str(raw_input.get('digital_presence_score')) + "/100\n"
"- Cash Flow: Rp " + str(round(raw_input.get('monthly_cash_flow',0)/1e6, 1)) + "M/bulan\n"
"- NPWP: " + ('ada' if raw_input.get('has_npwp') else 'belum ada') + " | SIUP: " + ('ada' if raw_input.get('has_siup') else 'belum ada') + "\n"
"- Pinjaman: Rp " + str(int(raw_input.get('loan_rp',50e6)/1e6)) + "M | Tenor: " + str(raw_input.get('duration',24)) + " bulan\n"
"- Riwayat Kredit: " + str(raw_input.get('credit_history')) + "\n\n"
"FAKTOR RISIKO UTAMA (SHAP):\n" + factors + "\n"
+ context
+ summary_block
+ cot_block
+ fewshot
+ "ATURAN MENJAWAB โ BACA SEMUA DENGAN SEKSAMA:\n"
"1. Jawab BEBAS โ tidak harus soal kredit. Kalau ditanya soal diri sendiri, perkenalkan sebagai Maya.\n"
"2. Kalau relevan, selalu hubungkan ke data pemohon dengan menyebut angka aktualnya.\n"
"3. Berikan saran KONKRET dan SPESIFIK, bukan generik.\n"
"4. Maksimal 150 kata kecuali diminta lebih panjang.\n"
"5. FORMAT TAG [ADJUST:] โ WAJIB IKUTI ATURAN INI:\n"
" BENAR: Embed tag langsung di akhir kalimat saran:\n"
" 'Coba naikkan digital score ke 75 [ADJUST: digital_presence_score=75] โ dampaknya paling besar.'\n"
" SALAH: Menulis kalimat saran TANPA tag, lalu menambah baris terpisah\n"
" SALAH: Menulis tag di baris sendiri di akhir response\n"
" SALAH BESAR: Menulis ulang rekomendasi di baris terpisah setelah selesai\n"
" RULE: Setiap rekomendasi bernilai angka = SATU kalimat + SATU tag [ADJUST:] inline. TIDAK LEBIH.\n"
" RULE: JANGAN pernah tulis JSON object {...} di response.\n"
" RULE: Field valid: digital_presence_score(1-100), business_age_years(1-20),\n"
" num_employees(1-50), monthly_cash_flow(angka Rp), duration(4-72), loan_rp(angka Rp)\n"
"6. JANGAN pernah balik ke template. Jawab seperti manusia cerdas.\n"
"7. Baca context percakapan sebelumnya sebelum menjawab โ jaga konsistensi topik.\n"
"8. AKHIRI response dengan kalimat lengkap. Jangan tambahkan apapun setelah kalimat terakhir.\n"
"9. LARANGAN KERAS โ JANGAN PERNAH LAKUKAN INI:\n"
" โ Menambah kalimat 'coba naikkan X ke N โ ...' di AKHIR response sebagai penutup\n"
" โ Menambah kalimat 'raise/increase X to N โ ...' di akhir response\n"
" โ Mengulang rekomendasi yang sudah ada di badan response sebagai trailing sentence\n"
" PENJELASAN: Kalau kamu sudah embed [ADJUST:] di dalam kalimat yang relevan, STOP.\n"
" Jangan append kalimat rekomendasi baru di akhir. Response harus berakhir dengan\n"
" kalimat natural yang sesuai dengan topik yang ditanyakan, bukan dengan action item.\n"
)
messages = [{"role": "system", "content": system}]
for m in recent_history[-8:]:
messages.append({"role": m['role'], "content": m['content']})
messages.append({"role": "user", "content": user_msg})
# โโ Call LLM with tool calling support โโโโโโโโโโโโโโโโโโโ
response, tool_calls, _llm_src, _last_error = _call_chat_llm(messages)
# Parse formal tool calls โ adjustments dict
adjustments = _parse_tool_calls(tool_calls)
# If LLM only returned tool calls (no text), convert to text
if not response and tool_calls:
response = _tool_call_to_text(tool_calls, lang)
# โโ Smart Keyword Fallback (only if ALL LLMs failed) โโโโโ
if response is None:
low = user_msg.lower()
dig = raw_input.get('digital_presence_score', 0)
cf = raw_input.get('monthly_cash_flow', 0)
npwp = raw_input.get('has_npwp', 0)
biz = raw_input.get('business_age_years', 0)
def _t(i, e, h):
return {'id': i, 'en': e, 'hi': h}[lang]
# Helper: word-boundary check for short keywords (avoids 'hi' matching 'mempengaruhi')
def _kw(msg_low, word_list):
for w in word_list:
if len(w) <= 3:
if re.search(r'\b' + re.escape(w) + r'\b', msg_low):
return True
else:
if w in msg_low:
return True
return False
# โ Who is Maya / greetings
# FIXED: removed 'kamu','km' (too broad โ appears in almost every sentence)
# FIXED: 'hi','hei','hai' use word-boundary via _kw helper
if _kw(low, ['siapa','who are you','perkenalan','halo','hello','hei','hai','hi',
'maya','nama kamu','you are']):
response = _t(
"Hei! Aku **Maya**, AI Credit Advisor kamu Aku dirancang untuk bantu kamu pahami skor kredit dan strategi bisnis UMKM. "
"Skor PD kamu sekarang **" + str(round(pd_pct,1)) + "%** โ " +
('sudah bagus banget!' if pd_pct < 20 else 'masih ada ruang untuk diperbaiki.') +
" Mau aku jelasin lebih detail atau ada yang mau ditanyain?",
"Hey! I'm **Maya**, your AI Credit Advisor. I help you understand your credit score and SME business strategy. "
"Your PD score is **" + str(round(pd_pct,1)) + "%** โ " +
('looking great!' if pd_pct < 20 else 'there is room to improve.') +
" Anything you want to ask?",
"Namaste! Main **Maya** hoon, aapki AI Credit Advisor. "
"Aapka PD score **" + str(round(pd_pct,1)) + "%** hai โ " +
('bahut badhiya!' if pd_pct < 20 else 'sudhaar ki gunjaish hai.') +
" Kuch poochna hai?"
)
# โก Quick profit / cuan intent โ CoT detects bisnis intent
elif any(w in low for w in ['cuan','profit','untung','laba','cepet cuan',
'penghasilan cepat','cara cepet','duit cepet']):
response = _t(
"Cuan cepet? Sah-sah aja! Tapi di kredit, yang bikin bank percaya bukan seberapa cepat, tapi **konsistensi**. "
"Dari profil kamu (cash flow Rp " + str(int(cf/1e6)) + "jt/bln, digital " + str(dig) + "/100):\n\n"
"1. **Aktifin marketplace** โ cash flow naik [ADJUST: monthly_cash_flow=25000000]\n"
"2. **Google Business aktif** โ digital score naik [ADJUST: digital_presence_score=75]\n"
"3. **Dokumentasikan semua pemasukan** โ pisah rekening bisnis & pribadi",
"Quick profit? Totally valid! But in credit, **consistency** matters more. "
"From your profile (CF Rp " + str(int(cf/1e6)) + "M/mo, digital " + str(dig) + "/100):\n\n"
"1. Active marketplace โ CF up [ADJUST: monthly_cash_flow=25000000]\n"
"2. Google Business โ digital up [ADJUST: digital_presence_score=75]\n"
"3. Document all income streams",
"Jaldi munafa? **Niyamitata** zaroori hai.\n"
"1) Marketplace active karein [ADJUST: monthly_cash_flow=25000000] "
"2) Google Business [ADJUST: digital_presence_score=75] 3) Aay document karein"
)
adjustments['digital_presence_score'] = 75
adjustments['monthly_cash_flow'] = 25000000
# โฅ Digital score targeted โ MOVED UP before โข and โค
# FIXED: was after โขimprove which had 'naik' catching 'naikin digital score'
elif any(w in low for w in ['digital score','skor digital','digital presence',
'naikin digital','naik digital','digital jadi','digital ke']):
target = 80
try:
nums = re.findall(r'\d+', user_msg)
if nums:
target = int(nums[-1])
except Exception:
pass
target = max(1, min(100, target))
adjustments['digital_presence_score'] = target
cur_dig = raw_input.get('digital_presence_score', 0)
response = _t(
"Digital score **" + str(cur_dig) + " โ " + str(target) + "** sudah aku set di What-If!\n\n"
"Cara naik digital score:\n"
"1. **Google Business Profile** โ verifikasi & lengkapi info\n"
"2. **Aktif di marketplace** โ Tokopedia/Shopee/TikTok Shop\n"
"3. **Media sosial konsisten** โ posting minimal 3x/minggu\n\n"
"Cek tab What-If untuk lihat dampak ke PD!",
"Digital score **" + str(cur_dig) + " โ " + str(target) + "** set in What-If!\n\n"
"Ways to improve:\n"
"1. **Google Business Profile** โ verify & complete\n"
"2. **Active marketplace** โ Tokopedia/Shopee/TikTok Shop\n"
"3. **Consistent social media** โ post 3x/week\n\n"
"Check the What-If tab for PD impact!",
"Digital score **" + str(cur_dig) + " โ " + str(target) + "** What-If mein set!\n\n"
"Sudhaar ke tarike:\n"
"1. Google Business Profile verify karein\n"
"2. Marketplace active rahein\n"
"3. Social media niyamit rahein\n\n"
"What-If tab mein PD prabhav dekhein!"
)
# โฉ Humor/absurd โ MOVED UP before โคloan
# FIXED: 'kredit' in โคloan was catching "beruang... minta kredit bisa gak?"
elif any(w in low for w in ['beruang','maling','jerapah','babi','kucing','lucu',
'absurd','random','hewan','binatang','gajah','dinosaurus']):
response = _t(
"Haha, oke oke! ๐ Tapi balik ke topik serius โ bahkan si beruang pun butuh skor kredit bagus buat pinjam madu dari bank!\n\n"
"Profil kamu: PD **" + str(round(pd_pct,1)) + "%** โ " +
('udah bagus banget nih!' if pd_pct < 20 else 'masih ada yang bisa diperbaiki.') +
" Ada yang mau ditanyain soal kredit atau bisnis?",
"Haha, fair enough! ๐ But back to business โ even a bear needs good credit to borrow honey from the bank!\n\n"
"Your profile: PD **" + str(round(pd_pct,1)) + "%** โ " +
('already looking great!' if pd_pct < 20 else 'some room to improve.') +
" Anything credit or business related?",
"Haha, theek hai! ๐ Lekin credit ki baat karein โ bhaaloo ko bhi bank se madhu udhaarne ke liye achha score chahiye!\n\n"
"Aapka PD **" + str(round(pd_pct,1)) + "%** โ " +
('pehle se badhiya!' if pd_pct < 20 else 'kuch sudhaar ho sakta hai.') +
" Kuch aur poochna hai?"
)
# โช 6-month action plan โ MOVED UP before โฃshap
# FIXED: 'apa yang perlu' was triggering โฃshap via 'apa yang' substring
elif any(w in low for w in ['6 bulan','rencana','plan','action plan','ke depan',
'persiapan','apply lagi','sebelum apply']):
rec_items = []
if not npwp:
rec_items.append(_t(
"**Bulan 1**: Urus NPWP di pajak.go.id (gratis, 1-3 hari)",
"**Month 1**: Register NPWP at pajak.go.id (free, 1-3 days)",
"**Maah 1**: NPWP register karein pajak.go.id par"
))
if dig < 60:
rec_items.append(_t(
"**Bulan 1-2**: Aktifkan Google Business + daftar marketplace",
"**Month 1-2**: Activate Google Business + register on marketplace",
"**Maah 1-2**: Google Business activate + marketplace register karein"
))
if cf < 15e6:
rec_items.append(_t(
"**Bulan 2-4**: Pisah rekening bisnis, dokumentasi semua pemasukan",
"**Month 2-4**: Separate business account, document all income",
"**Maah 2-4**: Vyapaar khaata alag karein, sabhi aay document karein"
))
rec_items.append(_t(
"**Bulan 5-6**: Simulasikan ulang di What-If, ajukan kredit jika PD < 20%",
"**Month 5-6**: Re-simulate in What-If, apply when PD < 20%",
"**Maah 5-6**: What-If mein re-simulate karein, PD < 20% ho tab apply karein"
))
response = _t(
"Action plan 6 bulan buat kamu (PD sekarang **" + str(round(pd_pct,1)) + "%**):\n\n",
"6-month action plan (current PD **" + str(round(pd_pct,1)) + "%**):\n\n",
"6 maah ka plan (abhi PD **" + str(round(pd_pct,1)) + "%**):\n\n"
) + '\n'.join(rec_items[:4])
# โข How to improve / lower PD โ triggers sliders
elif any(w in low for w in ['improve','better','lower','reduce','tingkatkan','kurangi',
'turunkan','cara','gimana','bagaimana','naik','turun',
'meningkat','naikkan','optimalkan']):
tips = []
if not npwp:
tips.append(_t(
"1. **Urus NPWP** โ wajib untuk pinjaman formal, daftar di pajak.go.id",
"1. **Register NPWP** โ required for formal loans, apply at pajak.go.id",
"1. **NPWP register karein** โ aupchaarik rin ke liye zaroori"
))
adjustments['has_npwp'] = 1
if dig < 75:
n = '2' if not npwp else '1'
tips.append(_t(
n + ". **Naikkan Digital Score dari " + str(dig) + " ke 75+** โ aktif di Google Business, Tokopedia/Shopee",
n + ". **Raise Digital Score from " + str(dig) + " to 75+** โ Google Business, marketplace",
n + ". **Digital Score " + str(dig) + " se 75+ karein** โ Google Business, marketplace"
))
adjustments['digital_presence_score'] = 75
if cf < 25e6:
n = str(len(tips) + 1)
tips.append(_t(
n + ". **Optimalkan cash flow dari Rp " + str(int(cf/1e6)) + "jt ke 25jt+/bln** โ diversifikasi produk",
n + ". **Grow cash flow from Rp " + str(int(cf/1e6)) + "M to 25M+/month**",
n + ". **Naqad pravaah Rp " + str(int(cf/1e6)) + "M se 25M+ karein**"
))
adjustments['monthly_cash_flow'] = 25000000
if not tips:
tips.append(_t(
"Profil kamu sudah solid dengan PD " + str(round(pd_pct,1)) + "%! Coba simulasikan di tab What-If.",
"Your profile is solid at PD " + str(round(pd_pct,1)) + "%! Try the What-If tab.",
"Aapka parichay PD " + str(round(pd_pct,1)) + "% ke saath mazboot hai!"
))
response = (
_t(
"Untuk turunkan PD dari **" + str(round(pd_pct,1)) + "%**, fokus ke:\n\n",
"To lower PD from **" + str(round(pd_pct,1)) + "%**, focus on:\n\n",
"PD **" + str(round(pd_pct,1)) + "%** kam karne ke liye:\n\n"
)
+ '\n'.join(tips[:3])
)
# โฃ Why high score / SHAP factors โ CoT picks actual SHAP values
# FIXED: removed 'apa yang' (matched 'berapa yang'), 'factor' (too broad)
# Use specific phrases instead
elif any(w in low for w in ['why','kenapa','mengapa','shap','pengaruh','faktor risiko',
'driver','jelasin','jelaskan','explain','definisi',
'kenapa skor','mengapa skor','yang mempengaruhi','apa penyebab']):
ti = int(np.argsort(np.abs(shap_vals))[-1])
ti2 = int(np.argsort(np.abs(shap_vals))[-2])
fn = feature_names[ti]
if 'business_age' in fn:
analogi = _t(
" โ ibarat pengalaman kerja, makin lama makin dipercaya",
" โ like work experience, longer = more trustworthy",
" โ kaam ke anubhav ki tarah, zyada = zyada bharosemand"
)
elif 'cash_flow' in fn:
analogi = _t(
" โ ibarat gaji bulanan kamu, makin besar makin mudah bayar cicilan",
" โ like your monthly income, higher = easier to repay",
" โ maasik aay ki tarah, zyada = kist bhrna aasaan"
)
elif 'digital' in fn:
analogi = _t(
" โ ibarat 'nilai reputasi online' bisnis kamu di mata bank",
" โ your business's 'online reputation score' in the bank's eyes",
" โ bank ki nazar mein vyapaar ka 'online pratishtha score'"
)
else:
analogi = ""
response = _t(
"Dua faktor terbesar yang drive skor PD **" + str(round(pd_pct,1)) + "%** kamu:\n\n"
"1. **" + fn + "**" + analogi + "\n โ " +
('meningkatkan' if shap_vals[ti] > 0 else 'menurunkan') +
" risiko (SHAP=" + str(round(shap_vals[ti],3)) + ")\n\n"
"2. **" + feature_names[ti2] + "**\n โ " +
('meningkatkan' if shap_vals[ti2] > 0 else 'menurunkan') +
" risiko (SHAP=" + str(round(shap_vals[ti2],3)) + ")\n\n"
+ ('Skor bagus! Kedua faktor ini justru mendukung kelayakan kamu.' if pd_pct < 20
else 'Fokus perbaiki faktor pertama dulu โ dampaknya paling besar.'),
"Two biggest drivers of your PD **" + str(round(pd_pct,1)) + "%**:\n\n"
"1. **" + fn + "**" + analogi + "\n โ " +
('increases' if shap_vals[ti] > 0 else 'decreases') +
" risk (SHAP=" + str(round(shap_vals[ti],3)) + ")\n\n"
"2. **" + feature_names[ti2] + "**\n โ " +
('increases' if shap_vals[ti2] > 0 else 'decreases') +
" risk (SHAP=" + str(round(shap_vals[ti2],3)) + ")\n\n"
+ ('Great score! Both factors support your eligibility.' if pd_pct < 20
else 'Focus on the first factor โ it has the biggest impact.'),
"Aapke PD **" + str(round(pd_pct,1)) + "%** ke do mukhya kaarak:\n\n"
"1. **" + fn + "**" + analogi + "\n โ " +
('badhata' if shap_vals[ti] > 0 else 'ghataata') +
" jokhim (SHAP=" + str(round(shap_vals[ti],3)) + ")\n\n"
"2. **" + feature_names[ti2] + "**\n โ " +
('badhata' if shap_vals[ti2] > 0 else 'ghataata') +
" jokhim (SHAP=" + str(round(shap_vals[ti2],3)) + ")\n\n"
+ ('Badhiya score! Dono kaarak yogyata ka samarthan karte hain.' if pd_pct < 20
else 'Pehle pehle kaarak sudhaarein โ sabse bada prabhav.')
)
# โค Ideal loan amount โ trigger loan_rp slider
# FIXED: removed 'besar' (too broad); โฅdigit and โฉhumor now come before this
elif any(w in low for w in ['pinjaman','loan','kredit','berapa','amount',
'limit','ideal','rekomendasi pinjaman']):
if cf > 0:
dur_val = raw_input.get('duration', 24)
safe = cf * dur_val * 0.35
cur = raw_input.get('loan_rp', 50e6)
if cur > safe:
adjustments['loan_rp'] = safe
response = _t(
"Berdasarkan cash flow kamu **Rp " + str(int(cf/1e6)) + "jt/bulan** dan tenor " +
str(int(dur_val)) + " bulan, pinjaman ideal maksimal **Rp " + str(int(safe/1e6)) +
"jt** (rasio cicilan 35%).\n\nPinjaman kamu sekarang Rp " + str(int(cur/1e6)) + "jt โ " +
('masih aman, good job!' if cur <= safe
else 'melebihi batas aman. Pertimbangkan kurangi ke Rp ' + str(int(safe/1e6)) + 'jt atau perpanjang tenor.'),
"Based on your cash flow **Rp " + str(int(cf/1e6)) + "M/month** and " +
str(int(dur_val)) + "-month tenure, ideal loan is max **Rp " + str(int(safe/1e6)) +
"M** (35% installment ratio).\n\nCurrent loan Rp " + str(int(cur/1e6)) + "M โ " +
('within safe limits, good job!' if cur <= safe
else 'exceeds safe limit. Consider reducing to Rp ' + str(int(safe/1e6)) + 'M.'),
"Aapke naqad pravaah **Rp " + str(int(cf/1e6)) + "M/maah** aur " +
str(int(dur_val)) + " maah ke aadhaar par adhiktam rin **Rp " + str(int(safe/1e6)) +
"M** (35% kist anupaat).\n\nVartamaan rin Rp " + str(int(cur/1e6)) + "M โ " +
('surakshit!' if cur <= safe else 'seema se adhik.')
)
else:
response = _t(
"Isi cash flow bulanan di form dulu ya, biar aku bisa kasih rekomendasi akurat!",
"Fill in your monthly cash flow first for an accurate recommendation!",
"Sahi sujhaav ke liye pehle maasik naqad pravaah bharein!"
)
# โฆ Savings / cash flow topics
elif any(w in low for w in ['tabungan','saving','nabung','menabung','savings',
'uang','keuangan','finance','financial','simpan']):
response = _t(
"Soal tabungan dan cash flow kamu (saat ini **Rp " + str(int(cf/1e6)) + "jt/bln**):\n\n"
"3 cara praktis tingkatkan tabungan UMKM:\n"
"1. Pisahkan rekening bisnis & pribadi โ biar tidak tercampur\n"
"2. Sisihkan minimal 10-15% dari omzet setiap bulan secara otomatis\n"
"3. Dokumentasikan semua pemasukan โ ini juga naikkan digital score!\n\n"
"Tabungan naik โ cash flow lebih sehat โ PD turun",
"About savings and your cash flow (currently **Rp " + str(int(cf/1e6)) + "M/month**):\n\n"
"3 practical SME savings tips:\n"
"1. Separate business & personal bank accounts\n"
"2. Auto-transfer 10-15% of monthly revenue to savings\n"
"3. Document all income streams โ this also boosts digital score!\n\n"
"More savings โ healthier cash flow โ lower PD",
"Bachat aur naqad pravaah (**Rp " + str(int(cf/1e6)) + "M/maah**) ke baare mein:\n\n"
"3 vyaavhaarik tarike:\n"
"1. Vyapaar aur vyaktigat khaate alag rakhein\n"
"2. Maasik aay ka 10-15% bachaen\n"
"3. Sabhi aay document karein โ digital score bhi badhega!\n\n"
"Zyada bachat โ swasth naqad pravaah โ PD kam"
)
adjustments['monthly_cash_flow'] = int(cf * 1.3)
# โง Credit history
elif any(w in low for w in ['riwayat kredit','credit history','kredit history',
'credit record','riwayat','history kredit']):
ch_val = raw_input.get('credit_history', '-')
response = _t(
"**Riwayat kredit** = catatan sejarah bayar utangmu โ ibarat rapor keuangan.\n\n"
"Status kamu: **" + ch_val + "**\n\n"
"Dampak ke skor kredit:\n"
"- All paid / Existing paid โ sinyal positif\n"
"- Delayed previously โ bank waspada\n"
"- Critical โ risiko naik signifikan\n\n"
"Cara bangun riwayat bagus: bayar cicilan tepat waktu, jangan ambil pinjaman melebihi kemampuan bayar.",
"**Credit history** = your track record of paying debts โ your financial report card.\n\n"
"Your status: **" + ch_val + "**\n\n"
"Score impact:\n"
"- All paid / Existing paid โ positive signal\n"
"- Delayed previously โ bank is cautious\n"
"- Critical โ significant risk increase\n\n"
"Build good history: pay on time, don't borrow beyond your repayment capacity.",
"**Credit itihaas** = karz bhugtaan ka record โ vitteey report card.\n\n"
"Aapki sthiti: **" + ch_val + "**\n\n"
"Score prabhav:\n"
"- All paid / Existing paid โ sakaraatmak sanket\n"
"- Delayed previously โ bank saavdhan\n"
"- Critical โ jokhim zyada\n\n"
"Samay par bhugtaan karein, kshamta se adhik rin na lein."
)
# โจ Business/SME/digital tips
elif any(w in low for w in ['tips','bisnis','usaha','umkm','sme','digital','marketplace',
'online','ecommerce','strategi','business tips']):
response = _t(
"Tips bisnis UMKM buat kamu (PD " + str(round(pd_pct,1)) + "%):\n\n"
"1. **Digital hadir** โ daftar Google Business Profile, aktif di Tokopedia/Shopee/TikTok Shop\n"
" (Digital score kamu " + str(dig) + "/100 โ masih bisa naik!)\n"
"2. **Pisah keuangan** โ rekening bisnis terpisah dari pribadi\n"
"3. **Dokumentasi rutin** โ catat semua transaksi, ini bukti ke bank\n"
"4. **Legalitas** โ NPWP & SIUP/NIB buka akses ke KUR & kredit formal\n\n"
"Mau aku simulasikan dampaknya ke skor di tab What-If?",
"SME business tips for you (PD " + str(round(pd_pct,1)) + "%):\n\n"
"1. **Go digital** โ Google Business Profile, Tokopedia/Shopee/TikTok Shop\n"
" (Your digital score " + str(dig) + "/100 โ room to improve!)\n"
"2. **Separate finances** โ dedicated business bank account\n"
"3. **Document everything** โ track all transactions as proof for banks\n"
"4. **Get legal** โ NPWP & SIUP/NIB unlock KUR & formal credit\n\n"
"Want me to simulate the impact in the What-If tab?",
"SME vyapaar tips (PD " + str(round(pd_pct,1)) + "%):\n\n"
"1. **Digital upasthiti** โ Google Business Profile, Tokopedia/Shopee\n"
" (Aapka digital score " + str(dig) + "/100 โ sudhaar ki gunjaish!)\n"
"2. **Alag vitteey khaata** โ vyapaar ka alag bank khaata\n"
"3. **Sab document karein** โ sabhi laanden-denden ka record\n"
"4. **Kanooni rahein** โ NPWP & SIUP/NIB KUR unlock karta hai\n\n"
"Kya main What-If tab mein prabhav simulate karun?"
)
if dig < 75:
adjustments['digital_presence_score'] = 75
# โซ Generic fallback
else:
top_issue = _get_top_issue(raw_input, pd_pct, lang)
has_key = any([st.session_state.get("openrouter_key"), st.session_state.get("groq_key")])
err_note = "\n\nLLM error: " + str(_last_error)[:80] if (has_key and _last_error) else ""
response = _t(
"Pertanyaan menarik! Untuk ini aku butuh koneksi AI.\n\n"
"Yang bisa aku kasih tahu: PD kamu **" + str(round(pd_pct,1)) + "%** dan " + top_issue + "\n\n"
"Coba tanya yang lebih spesifik: cara turunkan skor, faktor risiko, pinjaman ideal, tips tabungan, atau riwayat kredit." + err_note,
"Great question! For this I need an AI connection.\n\n"
"What I can share: your PD is **" + str(round(pd_pct,1)) + "%** and " + top_issue + "\n\n"
"Try asking specifically: how to lower score, risk factors, ideal loan, savings tips, or credit history." + err_note,
"Accha sawaal! Iske liye AI connection chahiye.\n\n"
"Aapka PD **" + str(round(pd_pct,1)) + "%** hai aur " + top_issue + "\n\n"
"Vishesh roop se puchein: score kaise kam karein, jokhim kaarak, ideal rin, bachat tips." + err_note
)
# โโ Final cleanup: clean response text + extract explicit [ADJUST:] tags โโ
if response:
response, extra_adj = _clean_response(response)
for field, val in extra_adj.items():
if field not in adjustments:
adjustments[field] = val
# โโ Semantic fallback: parse numeric values from natural LLM text โโโโโโโโโ
# 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
# ============================================================
# HEADER
# ============================================================
roc_b = ("ROC-AUC: " + str(meta['roc_auc']) + "") if meta.get('roc_auc') else ""
ks_b = ("KS: " + str(meta['ks_stat']) + "") if meta.get('ks_stat') else ""
st.markdown(
"",
unsafe_allow_html=True
)
# ============================================================
# INPUT FORM
# ============================================================
st.markdown('' + T("form_title",lang) + '
', unsafe_allow_html=True)
with st.form("credit_form"):
c1, c2, c3 = st.columns(3)
with c1:
st.markdown(T('form_loan', lang))
loan_rp = st.number_input(T('f_loan_amt',lang), 5_000_000, 500_000_000, 50_000_000, 5_000_000, format="%d")
lgd_val = st.slider(T('f_lgd',lang), 0.20, 0.80, 0.40, 0.05)
duration = st.slider(T('f_duration',lang), 4, 72, 24)
credit_amt = st.number_input(T('f_credit_amt',lang), 250, 20000, 2500, 250)
purpose = st.selectbox(T('f_purpose',lang), PURPOSE_OPTS, index=9,
format_func=lambda x: PURPOSE_LABELS[x][lang])
with c2:
st.markdown(T('form_financial', lang))
age = st.slider(T('f_age',lang), 18, 75, 35)
checking = st.selectbox(T('f_checking',lang), CHECKING_OPTS,
format_func=lambda x: CHECKING_LABELS[x][lang])
savings = st.selectbox(T('f_savings',lang), SAVINGS_OPTS, index=4,
format_func=lambda x: SAVINGS_LABELS[x][lang])
credit_hist = st.selectbox(T('f_credit_hist',lang), CREDIT_H_OPTS, index=2,
format_func=lambda x: CREDIT_H_LABELS[x][lang])
employment = st.selectbox(T('f_employment',lang), EMPLOY_OPTS, index=2,
format_func=lambda x: EMPLOY_LABELS[x][lang])
housing = st.selectbox(T('f_housing',lang), HOUSING_OPTS, index=2,
format_func=lambda x: HOUSING_LABELS[x][lang])
installment_c = st.slider(T('f_installment',lang), 1, 4, 2)
with c3:
st.markdown(T('form_sme', lang))
digital_score = st.slider(T('f_digital',lang), 1, 100, 50)
has_social = st.toggle(T('f_social',lang), True)
ecomm_vol = st.number_input(T('f_ecomm',lang), 0, 100_000_000, 5_000_000, 1_000_000)
has_npwp = st.toggle(T('f_npwp',lang), True)
has_siup = st.toggle(T('f_siup',lang), True)
biz_age = st.slider(T('f_biz_age',lang), 1, 20, 5)
cash_flow = st.number_input(T('f_cashflow',lang), 0, 100_000_000, 15_000_000, 1_000_000)
num_emp = st.slider(T('f_employees',lang), 1, 50, 5)
submitted = st.form_submit_button(T('f_submit',lang), use_container_width=True, type="primary")
# ============================================================
# EMPTY STATE (before first submission)
# ============================================================
if not submitted and st.session_state.result is None:
st.markdown("
", unsafe_allow_html=True)
for col, (icon, lbl_k, val) in zip(st.columns(4), [
('๐ค','empty_ensemble','XGB+LGBM+RF'),
('๐','empty_xai','SHAP'),
('๐ฌ','empty_narrative','Multi-LLM'),
('๐ค','empty_chat','Maya AI'),
]):
with col:
st.markdown(
'' + icon + '
'
'
' + T(lbl_k,lang) + '
'
'
' + val + '
", unsafe_allow_html=True)
for col, (lbl, key) in zip(st.columns(4), [
('ROC-AUC','roc_auc'),('KS Stat','ks_stat'),('CV Mean','cv_mean'),('Features','n_features')
]):
with col:
st.markdown(
'' + lbl + '
'
'
' + str(meta.get(key,"โ")) + '
"
+ result['cat'][lang] +
"
"
+ T('kpi_pd',lang) + ": " + str(round(result['pd']*100,1)) + "%"
"
' + T(lbl_k,lang) + '
'
'
' + val + '
'
'
' + T(sub_k,lang) + '
", unsafe_allow_html=True)
# ============================================================
# TAB LOADING PROGRESS โ FIXED: always defined before use
# ============================================================
_tab_load_steps = {
'id': [
'๐ Menyiapkan tab SHAP...',
'๐ฌ Menyiapkan Narasi AI...',
'๐ค Menyiapkan AI Chat...',
'๐ฎ Menyiapkan What-If...',
'๐ Menyiapkan Formula...',
'โ
Semua tab siap!',
],
'en': [
'๐ Preparing SHAP tab...',
'๐ฌ Preparing AI Narrative...',
'๐ค Preparing AI Chat...',
'๐ฎ Preparing What-If...',
'๐ Preparing Formula...',
'โ
All tabs ready!',
],
'hi': [
'๐ SHAP tab taiyaar ho raha hai...',
'๐ฌ AI Vivarana taiyaar ho raha hai...',
'๐ค AI Chat taiyaar ho raha hai...',
'๐ฎ What-If taiyaar ho raha hai...',
'๐ Sutra taiyaar ho raha hai...',
'โ
Sabhi tab taiyaar!',
],
}
_tls = _tab_load_steps.get(lang, _tab_load_steps['id'])
_tab_prog_ph = st.empty() # Always defined here โ no conditional
def _tab_step(n):
"""Update tab loading progress. Safe to call at any time."""
try:
pct = int(n / 5 * 100)
_tab_prog_ph.progress(pct, text=_tls[n])
except Exception:
pass
_tab_step(0) # Show immediately before rendering first tab content
# ============================================================
# TABS
# ============================================================
t1, t2, t3, t4, t5 = st.tabs([
T('tab_shap',lang), T('tab_narrative',lang),
T('tab_chat',lang), T('tab_whatif',lang), T('tab_formula',lang)
])
# โโ TAB 1: SHAP โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
with t1:
_tab_step(0)
st.markdown('' + T("shap_title",lang) + '
', unsafe_allow_html=True)
_shap_all_zero = (shap_vals is not None and np.all(shap_vals == 0))
if _shap_all_zero:
st.warning(
"โ ๏ธ **SHAP tidak tersedia** โ model XGBoost (`xgb_model.pkl`) mungkin belum di-upload atau "
"tidak kompatibel dengan SHAP TreeExplainer. "
"Pastikan `model/xgb_model.pkl` ada di HF Space dan di-train ulang jika perlu.\n\n"
"Skor PD dan narasi tetap valid (menggunakan ensemble model)."
)
else:
if st.session_state.get("shap_png"):
_spng = st.session_state.shap_png
if isinstance(_spng, str) and _spng.startswith('data:'):
# Render base64 directly โ avoids MediaFileStorage expiry
st.markdown(
'
',
unsafe_allow_html=True
)
elif _spng:
st.image(_spng, use_container_width=True)
st.caption(T('shap_caption', lang))
top_idx = np.argsort(np.abs(shap_vals))[-10:][::-1]
st.dataframe(pd.DataFrame({
T('shap_col_feature',lang): [feature_names[i] for i in top_idx],
'SHAP': [round(shap_vals[i],4) for i in top_idx],
T('shap_col_impact',lang): [
T('shap_risk_up',lang) if shap_vals[i] > 0 else T('shap_risk_down',lang)
for i in top_idx
],
}), use_container_width=True, hide_index=True)
# โโ TAB 2: NARRATIVE โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
with t2:
_tab_step(1)
flag = {'id':'๐ฎ๐ฉ','en':'๐ฌ๐ง','hi':'๐ฎ๐ณ'}[lang]
st.markdown(
'' + T("narr_title",lang) + ' ' + flag + '
',
unsafe_allow_html=True
)
st.caption(T('narr_source',lang) + ": " + (llm_src if llm_src else "โ"))
if narrative and narrative.strip():
fmt = re.sub(r'\*\*(.+?)\*\*', r'\1', narrative.replace('\n','
'))
st.markdown('' + fmt + '
', unsafe_allow_html=True)
else:
narrative_fb, src_fb = get_narrative(shap_vals, feature_names, result, lang, raw_input)
st.session_state.narrative = narrative_fb
st.session_state.llm_src = src_fb
fmt = re.sub(r'\*\*(.+?)\*\*', r'\1', narrative_fb.replace('\n','
'))
st.markdown('' + fmt + '
', unsafe_allow_html=True)
_or_key = st.session_state.get("openrouter_key","")
_grq_key = st.session_state.get("groq_key","")
with st.expander("LLM Debug Info"):
st.code(
"OR key : " + ('OK ' + _or_key[:12] + '...' if _or_key else 'missing') + "\n"
"Groq key : " + ('OK ' + _grq_key[:12] + '...' if _grq_key else 'missing') + "\n"
"Source : " + str(src_fb) + "\n"
"OR model : " + _OR_FREE_MODELS[0] + "\n"
"Groq mdl : " + _GROQ_FREE_MODELS[0]
)
# โโ TAB 3: AI CHAT โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
with t3:
_tab_step(2)
st.markdown(
'' + T("chat_title",lang) + '
',
unsafe_allow_html=True
)
# Memory badge โ shows when summary exists (after 8+ messages)
if st.session_state.get('chat_summary'):
n_turns = len(st.session_state.chat_history)
mem_lbl = {
'id': '๐ง Memory aktif ยท ' + str(n_turns) + ' pesan tersimpan',
'en': '๐ง Memory active ยท ' + str(n_turns) + ' messages saved',
'hi': '๐ง Memory active ยท ' + str(n_turns) + ' sandesh save hue',
}
st.markdown(
'' + mem_lbl.get(lang, mem_lbl["en"]) + '
',
unsafe_allow_html=True
)
# Quick chip buttons
chips = [T('chat_chip1',lang), T('chat_chip2',lang), T('chat_chip3',lang), T('chat_chip4',lang)]
chip_cols = st.columns(len(chips))
chip_clicked = None
for i, (col, chip) in enumerate(zip(chip_cols, chips)):
with col:
if st.button(chip, key="chip_" + str(i), use_container_width=True):
chip_clicked = chip
st.markdown("---")
# Chat history display
if st.session_state.chat_history:
html_chat = ""
for msg in st.session_state.chat_history:
content = re.sub(
r'\*\*(.+?)\*\*', r'\1',
msg['content'].replace('\n','
')
)
cls = "chat-bubble-user" if msg['role'] == 'user' else "chat-bubble-ai"
icon = "๐ค" if msg['role'] == 'user' else "๐ค"
html_chat += '' + icon + ' ' + content + '
'
st.markdown(html_chat, unsafe_allow_html=True)
else:
st.info("๐ " + T('chat_welcome', lang))
# Chat input + processing
user_input = st.chat_input(T('chat_input', lang))
to_process = chip_clicked or user_input
if to_process:
st.session_state.chat_history.append({'role': 'user', 'content': to_process})
# Live status indicator while LLM runs
_chat_status_ph = st.empty()
_spinner_label = {
'id': '๐ค Maya sedang berpikir...',
'en': '๐ค Maya is thinking...',
'hi': '๐ค Maya soch rahi hai...',
}.get(lang, '๐ค Maya sedang berpikir...')
with st.spinner(_spinner_label):
_chat_status_ph.info(
{
'id': 'โณ Menghubungi LLM โ bisa 5โ15 detik tergantung model yang dipilih...',
'en': 'โณ Connecting to LLM โ may take 5โ15 seconds depending on the model...',
'hi': 'โณ LLM se connect ho raha hai โ model ke hisaab se 5โ15 second lag sakte hain...',
}[lang],
icon="โณ"
)
ai_resp, adjustments, _debug_err = get_chat_response(
to_process,
st.session_state.chat_history,
result, raw_input, shap_vals, feature_names, lang,
rag_index=rag_index
)
_chat_status_ph.empty()
# Show error hint only if keys exist but all LLMs failed
if _debug_err and any([
st.session_state.get("openrouter_key"),
st.session_state.get("groq_key")
]):
st.warning(
"Semua LLM gagal, menggunakan Smart Fallback. Error: " + str(_debug_err)[:100]
)
# Append AI response
st.session_state.chat_history.append({'role': 'assistant', 'content': ai_resp})
_save_chat_memory(st.session_state.chat_history, st.session_state.get('chat_summary',''))
# Apply tool call / [ADJUST:] results to What-If sliders
adj_map = {
'digital_presence_score': 'wi_dig',
'business_age_years': 'wi_biz',
'num_employees': 'wi_emp',
'monthly_cash_flow': 'wi_cash',
'duration': 'wi_dur',
'loan_rp': 'wi_loan',
}
if adjustments:
for field, val in adjustments.items():
if field in adj_map:
st.session_state[adj_map[field]] = val
st.success("๐ก " + T('chat_updated', lang))
st.rerun()
# Clear chat button
if st.session_state.chat_history:
if st.button(T('chat_clear', lang)):
st.session_state.chat_history = []
st.session_state.chat_summary = ''
_save_chat_memory([], '')
st.rerun()
# โโ TAB 4: WHAT-IF โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
with t4:
_tab_step(3)
@st.fragment
def render_whatif():
_lang = st.session_state.get("lang_sel","id")
result_f = st.session_state.result
raw_input_f = st.session_state.raw_input
if result_f is None or raw_input_f is None:
st.info(T('wi_form_first', _lang))
return
st.markdown(
'' + T("wi_title",_lang) + '
',
unsafe_allow_html=True
)
st.caption("๐ก " + T('wi_caption', _lang))
d_dig = int(st.session_state.wi_dig or raw_input_f.get('digital_presence_score',50))
d_biz = int(st.session_state.wi_biz or raw_input_f.get('business_age_years',5))
d_emp = int(st.session_state.wi_emp or raw_input_f.get('num_employees',5))
d_cash = int((st.session_state.wi_cash or raw_input_f.get('monthly_cash_flow',15e6))//1e6)
d_dur = int(st.session_state.wi_dur or raw_input_f.get('duration',24))
d_loan = int((st.session_state.wi_loan or raw_input_f.get('loan_rp',50e6))//1e6)
wc1, wc2 = st.columns(2)
with wc1:
wi_loan = st.slider(T('wi_loan',_lang), 5, 500, d_loan, 5)
wi_dur = st.slider(T('wi_duration',_lang), 4, 72, d_dur)
wi_cash = st.slider(T('wi_cashflow',_lang), 1, 100, d_cash)
with wc2:
wi_dig = st.slider(T('wi_digital',_lang), 1, 100, d_dig)
wi_biz = st.slider(T('wi_bizage',_lang), 1, 20, d_biz)
wi_emp = st.slider(T('wi_employees',_lang),1, 50, d_emp)
try:
wi_raw = {
**raw_input_f,
'duration': wi_dur,
'digital_presence_score': wi_dig,
'business_age_years': float(wi_biz),
'monthly_cash_flow': float(wi_cash * 1e6),
'num_employees': wi_emp,
}
X_wi = preprocess(wi_raw, scaler, feature_names)
wi_pd = float(ensemble.predict_proba(X_wi)[0][1])
wi_res = risk_result(wi_pd, wi_loan * 1e6, result_f['lgd'])
d_pd = wi_pd - result_f['pd']
d_el = wi_res['el'] - result_f['el']
_log_api("What-If", "Local Model", True, 0)
r1, r2, r3 = st.columns(3)
with r1:
st.metric(T('kpi_pd',_lang), str(round(wi_pd*100,1)) + "%",
str(round(d_pd*100,1)) + "pp", delta_color="inverse")
with r2:
st.metric(T('kpi_el',_lang), "Rp " + str(round(wi_res['el']/1e6,2)) + "jt",
"Rp " + str(round(d_el/1e6,2)) + "jt", delta_color="inverse")
with r3:
icon_m = {'risk-low':'๐ข','risk-med':'๐ก','risk-high':'๐ด'}
label_m = {
'risk-low': T('wi_approved',_lang),
'risk-med': T('wi_review',_lang),
'risk-high': T('wi_highrisk',_lang),
}
st.metric(T('wi_status',_lang),
icon_m[wi_res['css']] + " " + label_m[wi_res['css']])
fig2, ax = plt.subplots(figsize=(7, 2.5))
bars = ax.barh(
['Original','What-If'],
[result_f['pd']*100, wi_pd*100],
color=[result_f['color'], wi_res['color']],
height=0.4
)
ax.axvline(20, color='#f7971e', ls='--', lw=1.5, label='20%')
ax.axvline(50, color='#e74c3c', ls='--', lw=1.5, label='50%')
for bar, v in zip(bars, [result_f['pd']*100, wi_pd*100]):
ax.text(bar.get_width()+0.5, bar.get_y()+bar.get_height()/2,
str(round(v,1)) + "%", va='center', fontweight='bold', fontsize=11)
ax.set_xlim(0, 110)
ax.legend(fontsize=8)
ax.grid(axis='x', alpha=0.3)
ax.set_title(T('wi_chart_title',_lang), fontweight='bold')
plt.tight_layout()
buf2 = io.BytesIO()
fig2.savefig(buf2, format='png', dpi=130, bbox_inches='tight')
plt.close(fig2)
buf2.seek(0)
_wi_b64 = base64.b64encode(buf2.read()).decode()
st.markdown(
'
',
unsafe_allow_html=True
)
if wi_pd < result_f['pd'] - 0.005:
st.success(
"PD " + T('wi_pd_down',_lang) + " **" +
str(round((result_f['pd']-wi_pd)*100,1)) + "pp** (" +
T('wi_pd_from',_lang) + " " + str(round(result_f['pd']*100,1)) +
"% โ " + str(round(wi_pd*100,1)) + "%)"
)
elif wi_pd > result_f['pd'] + 0.005:
st.warning(
"PD " + T('wi_pd_up',_lang) + " **" +
str(round((wi_pd-result_f['pd'])*100,1)) + "pp** (" +
T('wi_pd_from',_lang) + " " + str(round(result_f['pd']*100,1)) +
"% โ " + str(round(wi_pd*100,1)) + "%)"
)
else:
st.info("โ๏ธ " + T('wi_no_change',_lang))
def _t(i, e, h):
return {'id': i, 'en': e, 'hi': h}[_lang]
with st.expander(T('wi_tips_title', _lang)):
tips = []
if wi_dig < 70:
tips.append(_t(
"๐ฑ Naikkan Digital Score ke 70+ โ marketplace & Google Business",
"๐ฑ Raise Digital Score to 70+ โ marketplace & Google Business",
"๐ฑ Digital Score 70+ karein โ marketplace & Google Business"
))
if wi_cash < 20:
tips.append(_t(
"๐ต Target cash flow Rp 20jt+/bulan",
"๐ต Target Rp 20M+/month cash flow",
"๐ต Naqad pravaah Rp 20M+/maah target"
))
if wi_biz < 3:
tips.append(_t(
"๐ข Bisnis < 3 thn lebih berisiko โ bangun track record",
"๐ข Business < 3 yr is riskier โ build track record",
"๐ข 3 saal se kam vyapaar โ track record banaen"
))
if wi_emp < 5:
tips.append(_t(
"๐ฅ Tambah karyawan = skala bisnis lebih sehat",
"๐ฅ More employees signals healthy scale",
"๐ฅ Zyada karmachaaree = swasth paimaana"
))
if not tips:
tips.append(_t(
"๐ Profil sudah optimal!",
"๐ Profile already well-optimized!",
"๐ Parichay pehle se anukoolit!"
))
for tip in tips:
st.markdown(tip)
except Exception as e:
st.error("Error: " + str(e))
render_whatif()
# โโ TAB 5: FORMULA โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
with t5:
_tab_step(4)
st.markdown(
'' + T("form_formula_title",lang) + '
',
unsafe_allow_html=True
)
st.latex(r"EL = PD \times LGD \times EAD")
st.latex(
"EL = " + str(round(result['pd'],4)) + r" \times " + str(round(result['lgd'],2))
+ r" \times Rp\," + "{:,}".format(int(result['ead']))
+ r" = Rp\," + "{:,}".format(int(result['el']))
)
cf1, cf2 = st.columns(2)
with cf1:
st.markdown(T('formula_def', lang))
with cf2:
cat_lbl = (
T('formula_low',lang) if result['pd'] < .2 else
T('formula_medium',lang) if result['pd'] < .5 else
T('formula_high',lang)
)
st.markdown(
"| " + T('formula_komponen',lang) + " | " + T('formula_nilai',lang) + " |\n"
"|--|--|\n"
"| PD | " + str(round(result['pd']*100,1)) + "% (" + cat_lbl + ") |\n"
"| LGD | " + str(int(result['lgd']*100)) + "% |\n"
"| EAD | Rp " + str(int(result['ead']/1e6)) + "jt |\n"
"| **EL** | **Rp " + str(round(result['el']/1e6,2)) + "jt** |"
)
_log_api("Formula", "Local", True, 0)
# ============================================================
# All tabs rendered โ clear progress bar
# ============================================================
_tab_step(5)
_tab_prog_ph.empty()
# ============================================================
# DOWNLOAD REPORT
# ============================================================
st.markdown("---")
report_txt = (
"SME CREDIT RISK REPORT โ 1na37 AI ยท Batch 10\n" + "="*50 + "\n"
+ T('kpi_pd',lang) + ": " + str(round(result['pd']*100,1)) + "% โ " + result['cat'][lang] + "\n"
+ T('kpi_el',lang) + ": Rp " + "{:,}".format(int(result['el'])) + "\n"
"LGD: " + str(int(result['lgd']*100)) + "% (Basel II)\n"
"EAD: Rp " + "{:,}".format(int(result['ead'])) + "\n"
+ "="*50 + "\n"
+ T('narr_title',lang) + "\n" + narrative + "\n"
+ "="*50 + "\n"
"TOP SHAP\n" + shap_summary(shap_vals, feature_names, 5) + "\n"
+ "="*50 + "\n"
"DISCLAIMER: Educational use only. Not financial advice."
)
st.download_button(
T('download_btn', lang),
data=report_txt,
file_name=T('download_file', lang),
mime="text/plain"
)