Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,18 +1,28 @@
|
|
| 1 |
import os
|
| 2 |
import re
|
| 3 |
import textwrap
|
|
|
|
| 4 |
import pandas as pd
|
| 5 |
import gradio as gr
|
| 6 |
|
| 7 |
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
|
|
|
|
| 8 |
from langchain_core.documents import Document
|
| 9 |
from langchain_core.prompts import PromptTemplate
|
| 10 |
-
from langchain_community.vectorstores import FAISS
|
| 11 |
|
| 12 |
|
| 13 |
-
# =========================
|
| 14 |
-
#
|
| 15 |
-
# =========================
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
def one_sentence_per_line(text: str, width: int = 110) -> str:
|
| 17 |
if text is None:
|
| 18 |
return ""
|
|
@@ -23,8 +33,7 @@ def one_sentence_per_line(text: str, width: int = 110) -> str:
|
|
| 23 |
prefix, body = prefix_match.group(1), prefix_match.group(2)
|
| 24 |
wrapped = textwrap.wrap(body, width=max(20, width - len(prefix))) or [""]
|
| 25 |
return [prefix + wrapped[0]] + [(" " * len(prefix)) + w for w in wrapped[1:]]
|
| 26 |
-
|
| 27 |
-
return wrapped
|
| 28 |
|
| 29 |
out_lines = []
|
| 30 |
for raw_line in str(text).splitlines():
|
|
@@ -35,10 +44,8 @@ def one_sentence_per_line(text: str, width: int = 110) -> str:
|
|
| 35 |
parts = re.split(r"(?<=[.!?])\s+", line)
|
| 36 |
for part in parts:
|
| 37 |
part = part.strip()
|
| 38 |
-
if
|
| 39 |
-
|
| 40 |
-
out_lines.extend(_wrap_line(part))
|
| 41 |
-
|
| 42 |
return "\n".join(out_lines)
|
| 43 |
|
| 44 |
|
|
@@ -55,35 +62,30 @@ def enforce_third_person(text: str, customer_name: str) -> str:
|
|
| 55 |
return text
|
| 56 |
|
| 57 |
|
| 58 |
-
def format_customer_profile(profile) -> str:
|
| 59 |
-
if
|
| 60 |
return ""
|
| 61 |
-
|
| 62 |
-
if isinstance(profile, str):
|
| 63 |
-
return profile
|
| 64 |
-
|
| 65 |
d = dict(profile)
|
| 66 |
-
|
|
|
|
| 67 |
|
| 68 |
nat = str(d.get("Nationality", "")).strip().lower()
|
| 69 |
if nat == "singaporean":
|
| 70 |
d.pop("PR_Status", None)
|
| 71 |
|
| 72 |
lines = []
|
| 73 |
-
for k in
|
| 74 |
if k in d:
|
| 75 |
lines.append(f"{k}: {d.get(k)}")
|
| 76 |
-
|
| 77 |
for k in sorted(d.keys()):
|
| 78 |
-
if k not in
|
| 79 |
lines.append(f"{k}: {d.get(k)}")
|
| 80 |
-
|
| 81 |
return "\n".join(lines)
|
| 82 |
|
| 83 |
|
| 84 |
-
# =========================
|
| 85 |
-
# Load CSV
|
| 86 |
-
# =========================
|
| 87 |
def load_customer_csv(csv_path: str) -> pd.DataFrame:
|
| 88 |
df_all = pd.read_csv(csv_path)
|
| 89 |
df_all.columns = [c.strip() for c in df_all.columns]
|
|
@@ -124,10 +126,9 @@ def build_mock_systems(df_all: pd.DataFrame):
|
|
| 124 |
return df_credit, df_account, df_gov
|
| 125 |
|
| 126 |
|
| 127 |
-
def get_customer_profile(customer_id: str
|
| 128 |
customer_id = str(customer_id).strip()
|
| 129 |
-
|
| 130 |
-
credit_rec = df_credit[df_credit["ID"].astype(str) == customer_id]
|
| 131 |
if credit_rec.empty:
|
| 132 |
return None
|
| 133 |
|
|
@@ -135,13 +136,13 @@ def get_customer_profile(customer_id: str, df_credit, df_account, df_gov):
|
|
| 135 |
email = credit_rec.iloc[0]["Email"]
|
| 136 |
credit_score = int(credit_rec.iloc[0]["Credit_Score"])
|
| 137 |
|
| 138 |
-
acct_rec =
|
| 139 |
nationality = acct_rec.iloc[0]["Nationality"] if not acct_rec.empty else None
|
| 140 |
account_status = acct_rec.iloc[0]["Account_Status"] if not acct_rec.empty else None
|
| 141 |
|
| 142 |
pr_status = None
|
| 143 |
if nationality and str(nationality).strip().lower() == "non-singaporean":
|
| 144 |
-
gov_rec =
|
| 145 |
pr_status = bool(gov_rec.iloc[0]["PR_Status"]) if not gov_rec.empty else None
|
| 146 |
|
| 147 |
return {
|
|
@@ -155,9 +156,9 @@ def get_customer_profile(customer_id: str, df_credit, df_account, df_gov):
|
|
| 155 |
}
|
| 156 |
|
| 157 |
|
| 158 |
-
# =========================
|
| 159 |
-
#
|
| 160 |
-
# =========================
|
| 161 |
def extract_pdf_text(pdf_path: str) -> str:
|
| 162 |
from pypdf import PdfReader
|
| 163 |
reader = PdfReader(pdf_path)
|
|
@@ -167,80 +168,63 @@ def extract_pdf_text(pdf_path: str) -> str:
|
|
| 167 |
return "\n".join(pages)
|
| 168 |
|
| 169 |
|
| 170 |
-
def
|
| 171 |
-
|
| 172 |
-
interest_policy_text = extract_pdf_text(interest_pdf_path)
|
| 173 |
-
|
| 174 |
rate_matches = re.findall(
|
| 175 |
r"\b(Low|Medium|High)\b\s+([0-9]+\.[0-9]+)\s*%?",
|
| 176 |
-
|
| 177 |
flags=re.IGNORECASE
|
| 178 |
)
|
| 179 |
interest_rates = {k.capitalize(): float(v) for k, v in rate_matches}
|
| 180 |
|
|
|
|
| 181 |
risk_rows = re.findall(
|
| 182 |
r"(\d{3})\s*(?:-|β|β)?\s*(\d{3})\s+(Delinquent|Closed|Good-standing)\s+(High|Medium|Low)",
|
| 183 |
-
|
| 184 |
flags=re.IGNORECASE,
|
| 185 |
)
|
| 186 |
|
| 187 |
risk_mapping = {}
|
| 188 |
for lo, hi, status, risk in risk_rows:
|
| 189 |
band = (int(lo), int(hi))
|
| 190 |
-
|
| 191 |
-
risk_mapping[key] = risk.capitalize()
|
| 192 |
|
| 193 |
-
return
|
| 194 |
|
| 195 |
|
| 196 |
-
# =========================================================
|
| 197 |
-
# Deterministic rules (same as Step 5)
|
| 198 |
-
# =========================================================
|
| 199 |
def _score_band(score: int):
|
| 200 |
-
if 300 <= score <= 674:
|
| 201 |
-
|
| 202 |
-
if
|
| 203 |
-
|
| 204 |
-
if 750 <= score <= 850:
|
| 205 |
-
return (750, 850)
|
| 206 |
-
if score < 300:
|
| 207 |
-
return (300, 674)
|
| 208 |
return (750, 850)
|
| 209 |
|
| 210 |
|
| 211 |
-
def determine_overall_risk(score: int, account_status: str
|
| 212 |
band = _score_band(score)
|
| 213 |
status = str(account_status).strip().lower()
|
| 214 |
-
|
| 215 |
-
if key not in risk_mapping:
|
| 216 |
-
return "High"
|
| 217 |
-
return risk_mapping[key]
|
| 218 |
|
| 219 |
|
| 220 |
-
def determine_interest_rate(overall_risk: str
|
| 221 |
-
return float(
|
| 222 |
|
| 223 |
|
| 224 |
-
def is_non_singaporean_no_pr(customer_id: str
|
| 225 |
-
|
| 226 |
-
|
| 227 |
-
|
| 228 |
-
nationality = nat_row.iloc[0]["Nationality"] if not nat_row.empty else None
|
| 229 |
|
| 230 |
-
|
| 231 |
-
|
| 232 |
|
| 233 |
-
|
| 234 |
-
except Exception:
|
| 235 |
-
return False
|
| 236 |
|
| 237 |
|
| 238 |
-
def apply_mandatory_exception_to_report(report_text: str, customer_id: str
|
| 239 |
-
if report_text is None
|
| 240 |
-
report_text = ""
|
| 241 |
-
text = str(report_text)
|
| 242 |
|
| 243 |
-
if not is_non_singaporean_no_pr(customer_id
|
| 244 |
return text
|
| 245 |
|
| 246 |
text = re.sub(
|
|
@@ -264,12 +248,45 @@ def apply_mandatory_exception_to_report(report_text: str, customer_id: str, df_a
|
|
| 264 |
return text
|
| 265 |
|
| 266 |
|
| 267 |
-
# =========================
|
| 268 |
-
#
|
| 269 |
-
# =========================
|
| 270 |
-
|
| 271 |
-
|
| 272 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 273 |
|
| 274 |
CUSTOMER DATA:
|
| 275 |
{customer_data}
|
|
@@ -282,24 +299,22 @@ USER QUESTION:
|
|
| 282 |
|
| 283 |
ANSWER:
|
| 284 |
"""
|
| 285 |
-
qa_prompt = PromptTemplate(
|
| 286 |
-
input_variables=["customer_data", "policy_rules", "question"],
|
| 287 |
-
template=qa_prompt_template
|
| 288 |
)
|
| 289 |
|
| 290 |
-
|
|
|
|
|
|
|
| 291 |
|
| 292 |
-
Write in THIRD PERSON about the customer.
|
| 293 |
-
Always use the customer's Name and possessive.
|
| 294 |
Never address the reader as 'you' or 'your'.
|
| 295 |
|
| 296 |
-
You must provide ADVICE/RECOMMENDATION.
|
| 297 |
Use ONLY the provided Customer Data and Policy Rules.
|
| 298 |
|
| 299 |
REQUIREMENTS:
|
| 300 |
-
- Provide 3-5 actionable advice points.
|
| 301 |
-
- Include a final recommendation: APPROVE or NOT RECOMMEND / REJECT.
|
| 302 |
-
- If customer is Non-Singaporean and PR_Status is False, you MUST recommend NOT RECOMMEND / REJECT.
|
| 303 |
- Keep it concise.
|
| 304 |
|
| 305 |
CUSTOMER DATA:
|
|
@@ -313,17 +328,14 @@ USER QUESTION:
|
|
| 313 |
|
| 314 |
ANSWER:
|
| 315 |
"""
|
| 316 |
-
advice_prompt = PromptTemplate(
|
| 317 |
-
input_variables=["customer_data", "policy_rules", "question"],
|
| 318 |
-
template=advice_prompt_template
|
| 319 |
)
|
| 320 |
|
| 321 |
-
|
| 322 |
-
|
| 323 |
-
|
| 324 |
-
|
| 325 |
-
|
| 326 |
-
|
| 327 |
|
| 328 |
CUSTOMER DATA:
|
| 329 |
{customer_data}
|
|
@@ -333,250 +345,200 @@ POLICY RULES:
|
|
| 333 |
|
| 334 |
REPORT:
|
| 335 |
"""
|
| 336 |
-
report_prompt = PromptTemplate(
|
| 337 |
-
input_variables=["customer_data", "policy_rules"],
|
| 338 |
-
template=report_prompt_template
|
| 339 |
)
|
| 340 |
|
| 341 |
|
| 342 |
-
# =========================
|
| 343 |
-
#
|
| 344 |
-
# =========================
|
| 345 |
-
|
| 346 |
-
|
| 347 |
-
|
| 348 |
-
INT_PDF = os.path.join(BASE, "Interest_Rate_Policy.pdf")
|
| 349 |
|
| 350 |
-
|
| 351 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 352 |
|
| 353 |
-
|
| 354 |
-
api_key = os.getenv("OPENAI_API_KEY", "")
|
|
|
|
|
|
|
| 355 |
if not api_key:
|
| 356 |
raise RuntimeError("Missing OPENAI_API_KEY. Set it in Space Secrets.")
|
| 357 |
|
| 358 |
-
|
| 359 |
-
|
| 360 |
-
|
| 361 |
|
| 362 |
-
# Load
|
| 363 |
-
|
| 364 |
-
|
|
|
|
| 365 |
|
| 366 |
-
#
|
| 367 |
-
|
| 368 |
-
policy_full_text = risk_text + "\n\n" + interest_text
|
| 369 |
|
| 370 |
# LLM
|
| 371 |
model_name = os.getenv("OPENAI_MODEL", "gpt-4o")
|
| 372 |
-
|
| 373 |
|
| 374 |
# Chains (LCEL)
|
| 375 |
-
|
| 376 |
-
|
| 377 |
-
|
| 378 |
-
|
| 379 |
-
# For dropdown convenience
|
| 380 |
-
df_names = df_credit[["ID", "Name"]].copy()
|
| 381 |
-
df_names["label"] = df_names["Name"] + " (ID " + df_names["ID"].astype(str) + ")"
|
| 382 |
-
all_choices = df_names["label"].tolist()
|
| 383 |
-
|
| 384 |
-
return (
|
| 385 |
-
df_credit, df_account, df_gov,
|
| 386 |
-
risk_mapping, interest_rates,
|
| 387 |
-
risk_text, interest_text,
|
| 388 |
-
policy_full_text,
|
| 389 |
-
qa_chain, advice_chain, report_chain,
|
| 390 |
-
df_names, all_choices
|
| 391 |
-
)
|
| 392 |
-
|
| 393 |
|
| 394 |
-
try:
|
| 395 |
-
(
|
| 396 |
-
DF_CREDIT, DF_ACCOUNT, DF_GOV,
|
| 397 |
-
RISK_MAP, RATE_MAP,
|
| 398 |
-
RISK_TEXT, INT_TEXT,
|
| 399 |
-
POLICY_FULL,
|
| 400 |
-
QA_CHAIN, ADVICE_CHAIN, REPORT_CHAIN,
|
| 401 |
-
DF_NAMES, ALL_CHOICES
|
| 402 |
-
) = init()
|
| 403 |
-
INIT_ERROR = None
|
| 404 |
except Exception as e:
|
| 405 |
INIT_ERROR = str(e)
|
| 406 |
|
| 407 |
|
| 408 |
-
def
|
| 409 |
-
global
|
| 410 |
-
if
|
| 411 |
-
return
|
| 412 |
|
| 413 |
-
|
| 414 |
-
|
| 415 |
-
|
| 416 |
-
|
| 417 |
-
|
| 418 |
-
|
| 419 |
-
|
| 420 |
-
|
| 421 |
-
|
| 422 |
-
|
| 423 |
-
|
|
|
|
| 424 |
|
| 425 |
|
| 426 |
-
def
|
| 427 |
if not use_rag:
|
| 428 |
return POLICY_FULL
|
| 429 |
|
| 430 |
-
retriever, err = _build_retriever_if_needed()
|
| 431 |
-
if retriever is None:
|
| 432 |
-
return POLICY_FULL + f"\n\n[Note] RAG disabled due to embeddings error: {err}"
|
| 433 |
-
|
| 434 |
try:
|
| 435 |
-
|
| 436 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 437 |
return ctx if ctx else POLICY_FULL
|
| 438 |
except Exception as e:
|
| 439 |
-
return POLICY_FULL + f"\n\n[Note] RAG
|
| 440 |
-
|
| 441 |
-
|
| 442 |
-
def find_matches(name_or_id: str):
|
| 443 |
-
if INIT_ERROR:
|
| 444 |
-
return [], f"β Initialization error:\n{INIT_ERROR}"
|
| 445 |
-
|
| 446 |
-
s = (name_or_id or "").strip()
|
| 447 |
-
if not s:
|
| 448 |
-
return [], "Type a Name or ID, then click Find."
|
| 449 |
-
|
| 450 |
-
# ID
|
| 451 |
-
if s.isdigit():
|
| 452 |
-
prof = get_customer_profile(s, DF_CREDIT, DF_ACCOUNT, DF_GOV)
|
| 453 |
-
if prof:
|
| 454 |
-
label = prof["Name"] + " (ID " + str(prof["ID"]) + ")"
|
| 455 |
-
return [label], f"β
Found ID {s}"
|
| 456 |
-
return [], f"β ID {s} not found."
|
| 457 |
-
|
| 458 |
-
# Name contains
|
| 459 |
-
results = DF_CREDIT[DF_CREDIT["Name"].astype(str).str.contains(s, case=False, na=False)]
|
| 460 |
-
if results.empty:
|
| 461 |
-
return [], f"β No customer matched '{s}'."
|
| 462 |
-
|
| 463 |
-
if len(results) == 1:
|
| 464 |
-
row = results.iloc[0]
|
| 465 |
-
label = row["Name"] + " (ID " + str(row["ID"]) + ")"
|
| 466 |
-
return [label], f"β
Found Name '{row['Name']}'"
|
| 467 |
-
|
| 468 |
-
labels = []
|
| 469 |
-
for _, r in results.iterrows():
|
| 470 |
-
labels.append(r["Name"] + " (ID " + str(r["ID"]) + ")")
|
| 471 |
-
return labels, f"β οΈ Multiple matches for '{s}'. Please select one."
|
| 472 |
-
|
| 473 |
-
|
| 474 |
-
def _resolve_id_from_label(label: str) -> str:
|
| 475 |
-
row = DF_NAMES[DF_NAMES["label"] == label]
|
| 476 |
-
if row.empty:
|
| 477 |
-
return ""
|
| 478 |
-
return str(row.iloc[0]["ID"])
|
| 479 |
|
| 480 |
|
| 481 |
-
|
|
|
|
|
|
|
|
|
|
| 482 |
if INIT_ERROR:
|
| 483 |
-
return f"β Initialization error:\n{INIT_ERROR}"
|
| 484 |
|
| 485 |
-
|
| 486 |
-
|
| 487 |
-
|
| 488 |
-
|
| 489 |
-
|
| 490 |
-
|
| 491 |
-
|
| 492 |
-
|
| 493 |
-
|
| 494 |
-
|
| 495 |
-
|
| 496 |
-
|
| 497 |
-
|
| 498 |
-
|
| 499 |
-
|
| 500 |
-
|
| 501 |
-
|
| 502 |
-
|
| 503 |
-
|
| 504 |
-
"
|
| 505 |
-
|
| 506 |
-
|
| 507 |
-
|
| 508 |
-
|
| 509 |
-
|
| 510 |
-
|
| 511 |
-
|
| 512 |
-
|
| 513 |
-
|
| 514 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 515 |
})
|
| 516 |
-
|
| 517 |
-
|
| 518 |
-
# Option 3: Full report (+ mandatory exception enforcement)
|
| 519 |
-
customer_data_with_exception = prof_text + (
|
| 520 |
-
"\n\nMANDATORY EXCEPTION (must follow): Non-Singaporean with PR_Status = False => NOT RECOMMENDED / REJECTED."
|
| 521 |
-
if is_non_singaporean_no_pr(cid, DF_ACCOUNT, DF_GOV)
|
| 522 |
-
else ""
|
| 523 |
-
)
|
| 524 |
|
| 525 |
-
|
| 526 |
-
"
|
| 527 |
-
"policy_rules": policy_context
|
| 528 |
-
})
|
| 529 |
-
|
| 530 |
-
report_text = apply_mandatory_exception_to_report(full_report.content, cid, DF_ACCOUNT, DF_GOV)
|
| 531 |
-
return one_sentence_per_line(report_text)
|
| 532 |
|
| 533 |
|
| 534 |
-
# =========================
|
| 535 |
-
# Gradio UI (
|
| 536 |
-
# =========================
|
| 537 |
with gr.Blocks(title="Bank Loan Officer System") as demo:
|
| 538 |
-
gr.Markdown("# π¦
|
| 539 |
-
gr.Markdown("Type Applicant Name
|
| 540 |
|
| 541 |
if INIT_ERROR:
|
| 542 |
gr.Markdown(f"## β Initialization error\n\n```\n{INIT_ERROR}\n```")
|
| 543 |
|
| 544 |
-
|
| 545 |
-
|
| 546 |
-
|
|
|
|
| 547 |
|
| 548 |
-
|
| 549 |
-
|
| 550 |
-
|
| 551 |
-
|
| 552 |
-
|
| 553 |
-
|
| 554 |
-
|
| 555 |
-
|
| 556 |
-
|
| 557 |
-
|
| 558 |
-
gr.
|
| 559 |
-
|
| 560 |
-
|
| 561 |
-
|
| 562 |
-
choices=[
|
| 563 |
-
"1) Check Risk & Interest",
|
| 564 |
-
"2) Advice / Recommendation",
|
| 565 |
-
"3) FULL Formal Loan Report"
|
| 566 |
-
],
|
| 567 |
-
value="1) Check Risk & Interest"
|
| 568 |
-
)
|
| 569 |
-
use_rag = gr.Checkbox(
|
| 570 |
-
label="Use RAG (FAISS embeddings + retrieval). If it fails, auto fallback.",
|
| 571 |
-
value=True
|
| 572 |
-
)
|
| 573 |
|
| 574 |
run_btn = gr.Button("π Run")
|
| 575 |
-
output = gr.Textbox(label="Output", lines=
|
|
|
|
|
|
|
|
|
|
| 576 |
|
| 577 |
-
run_btn.click(fn=run_action, inputs=[matches, action, use_rag], outputs=[output])
|
| 578 |
|
| 579 |
-
|
|
|
|
|
|
|
| 580 |
PORT = int(os.environ.get("PORT", 7860))
|
| 581 |
demo.queue().launch(
|
| 582 |
server_name="0.0.0.0",
|
|
|
|
| 1 |
import os
|
| 2 |
import re
|
| 3 |
import textwrap
|
| 4 |
+
import traceback
|
| 5 |
import pandas as pd
|
| 6 |
import gradio as gr
|
| 7 |
|
| 8 |
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
|
| 9 |
+
from langchain_community.vectorstores import FAISS
|
| 10 |
from langchain_core.documents import Document
|
| 11 |
from langchain_core.prompts import PromptTemplate
|
|
|
|
| 12 |
|
| 13 |
|
| 14 |
+
# =========================
|
| 15 |
+
# Files expected in repo root
|
| 16 |
+
# =========================
|
| 17 |
+
BASE_DIR = os.path.dirname(__file__)
|
| 18 |
+
CSV_FILE = os.path.join(BASE_DIR, "Customer records.csv")
|
| 19 |
+
RISK_PDF = os.path.join(BASE_DIR, "Risk_Policy.pdf")
|
| 20 |
+
INT_PDF = os.path.join(BASE_DIR, "Interest_Rate_Policy.pdf")
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
# =========================
|
| 24 |
+
# Helpers: formatting
|
| 25 |
+
# =========================
|
| 26 |
def one_sentence_per_line(text: str, width: int = 110) -> str:
|
| 27 |
if text is None:
|
| 28 |
return ""
|
|
|
|
| 33 |
prefix, body = prefix_match.group(1), prefix_match.group(2)
|
| 34 |
wrapped = textwrap.wrap(body, width=max(20, width - len(prefix))) or [""]
|
| 35 |
return [prefix + wrapped[0]] + [(" " * len(prefix)) + w for w in wrapped[1:]]
|
| 36 |
+
return textwrap.wrap(line, width=width) or [""]
|
|
|
|
| 37 |
|
| 38 |
out_lines = []
|
| 39 |
for raw_line in str(text).splitlines():
|
|
|
|
| 44 |
parts = re.split(r"(?<=[.!?])\s+", line)
|
| 45 |
for part in parts:
|
| 46 |
part = part.strip()
|
| 47 |
+
if part:
|
| 48 |
+
out_lines.extend(_wrap_line(part))
|
|
|
|
|
|
|
| 49 |
return "\n".join(out_lines)
|
| 50 |
|
| 51 |
|
|
|
|
| 62 |
return text
|
| 63 |
|
| 64 |
|
| 65 |
+
def format_customer_profile(profile: dict) -> str:
|
| 66 |
+
if not profile:
|
| 67 |
return ""
|
|
|
|
|
|
|
|
|
|
|
|
|
| 68 |
d = dict(profile)
|
| 69 |
+
|
| 70 |
+
preferred = ["ID", "Name", "Email", "Credit_Score", "Nationality", "Account_Status", "PR_Status"]
|
| 71 |
|
| 72 |
nat = str(d.get("Nationality", "")).strip().lower()
|
| 73 |
if nat == "singaporean":
|
| 74 |
d.pop("PR_Status", None)
|
| 75 |
|
| 76 |
lines = []
|
| 77 |
+
for k in preferred:
|
| 78 |
if k in d:
|
| 79 |
lines.append(f"{k}: {d.get(k)}")
|
|
|
|
| 80 |
for k in sorted(d.keys()):
|
| 81 |
+
if k not in preferred:
|
| 82 |
lines.append(f"{k}: {d.get(k)}")
|
|
|
|
| 83 |
return "\n".join(lines)
|
| 84 |
|
| 85 |
|
| 86 |
+
# =========================
|
| 87 |
+
# Load CSV + build mock systems
|
| 88 |
+
# =========================
|
| 89 |
def load_customer_csv(csv_path: str) -> pd.DataFrame:
|
| 90 |
df_all = pd.read_csv(csv_path)
|
| 91 |
df_all.columns = [c.strip() for c in df_all.columns]
|
|
|
|
| 126 |
return df_credit, df_account, df_gov
|
| 127 |
|
| 128 |
|
| 129 |
+
def get_customer_profile(customer_id: str):
|
| 130 |
customer_id = str(customer_id).strip()
|
| 131 |
+
credit_rec = DF_CREDIT[DF_CREDIT["ID"].astype(str) == customer_id]
|
|
|
|
| 132 |
if credit_rec.empty:
|
| 133 |
return None
|
| 134 |
|
|
|
|
| 136 |
email = credit_rec.iloc[0]["Email"]
|
| 137 |
credit_score = int(credit_rec.iloc[0]["Credit_Score"])
|
| 138 |
|
| 139 |
+
acct_rec = DF_ACCOUNT[DF_ACCOUNT["ID"].astype(str) == customer_id]
|
| 140 |
nationality = acct_rec.iloc[0]["Nationality"] if not acct_rec.empty else None
|
| 141 |
account_status = acct_rec.iloc[0]["Account_Status"] if not acct_rec.empty else None
|
| 142 |
|
| 143 |
pr_status = None
|
| 144 |
if nationality and str(nationality).strip().lower() == "non-singaporean":
|
| 145 |
+
gov_rec = DF_GOV[DF_GOV["ID"].astype(str) == customer_id]
|
| 146 |
pr_status = bool(gov_rec.iloc[0]["PR_Status"]) if not gov_rec.empty else None
|
| 147 |
|
| 148 |
return {
|
|
|
|
| 156 |
}
|
| 157 |
|
| 158 |
|
| 159 |
+
# =========================
|
| 160 |
+
# PDF ingest + parse policies
|
| 161 |
+
# =========================
|
| 162 |
def extract_pdf_text(pdf_path: str) -> str:
|
| 163 |
from pypdf import PdfReader
|
| 164 |
reader = PdfReader(pdf_path)
|
|
|
|
| 168 |
return "\n".join(pages)
|
| 169 |
|
| 170 |
|
| 171 |
+
def parse_policies(risk_text: str, interest_text: str):
|
| 172 |
+
# Interest rates
|
|
|
|
|
|
|
| 173 |
rate_matches = re.findall(
|
| 174 |
r"\b(Low|Medium|High)\b\s+([0-9]+\.[0-9]+)\s*%?",
|
| 175 |
+
interest_text,
|
| 176 |
flags=re.IGNORECASE
|
| 177 |
)
|
| 178 |
interest_rates = {k.capitalize(): float(v) for k, v in rate_matches}
|
| 179 |
|
| 180 |
+
# Risk table rows
|
| 181 |
risk_rows = re.findall(
|
| 182 |
r"(\d{3})\s*(?:-|β|β)?\s*(\d{3})\s+(Delinquent|Closed|Good-standing)\s+(High|Medium|Low)",
|
| 183 |
+
risk_text,
|
| 184 |
flags=re.IGNORECASE,
|
| 185 |
)
|
| 186 |
|
| 187 |
risk_mapping = {}
|
| 188 |
for lo, hi, status, risk in risk_rows:
|
| 189 |
band = (int(lo), int(hi))
|
| 190 |
+
risk_mapping[(band, status.strip().lower())] = risk.capitalize()
|
|
|
|
| 191 |
|
| 192 |
+
return risk_mapping, interest_rates
|
| 193 |
|
| 194 |
|
|
|
|
|
|
|
|
|
|
| 195 |
def _score_band(score: int):
|
| 196 |
+
if 300 <= score <= 674: return (300, 674)
|
| 197 |
+
if 675 <= score <= 749: return (675, 749)
|
| 198 |
+
if 750 <= score <= 850: return (750, 850)
|
| 199 |
+
if score < 300: return (300, 674)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 200 |
return (750, 850)
|
| 201 |
|
| 202 |
|
| 203 |
+
def determine_overall_risk(score: int, account_status: str) -> str:
|
| 204 |
band = _score_band(score)
|
| 205 |
status = str(account_status).strip().lower()
|
| 206 |
+
return RISK_MAPPING.get((band, status), "High")
|
|
|
|
|
|
|
|
|
|
| 207 |
|
| 208 |
|
| 209 |
+
def determine_interest_rate(overall_risk: str) -> float:
|
| 210 |
+
return float(INTEREST_RATES[overall_risk])
|
| 211 |
|
| 212 |
|
| 213 |
+
def is_non_singaporean_no_pr(customer_id: str) -> bool:
|
| 214 |
+
cid = str(customer_id).strip()
|
| 215 |
+
nat_row = DF_ACCOUNT[DF_ACCOUNT["ID"].astype(str).str.strip() == cid]
|
| 216 |
+
nationality = nat_row.iloc[0]["Nationality"] if not nat_row.empty else None
|
|
|
|
| 217 |
|
| 218 |
+
pr_row = DF_GOV[DF_GOV["ID"].astype(str).str.strip() == cid]
|
| 219 |
+
pr_status = bool(pr_row.iloc[0]["PR_Status"]) if not pr_row.empty else False
|
| 220 |
|
| 221 |
+
return (str(nationality).strip().lower() != "singaporean") and (pr_status is False)
|
|
|
|
|
|
|
| 222 |
|
| 223 |
|
| 224 |
+
def apply_mandatory_exception_to_report(report_text: str, customer_id: str) -> str:
|
| 225 |
+
text = "" if report_text is None else str(report_text)
|
|
|
|
|
|
|
| 226 |
|
| 227 |
+
if not is_non_singaporean_no_pr(customer_id):
|
| 228 |
return text
|
| 229 |
|
| 230 |
text = re.sub(
|
|
|
|
| 248 |
return text
|
| 249 |
|
| 250 |
|
| 251 |
+
# =========================
|
| 252 |
+
# Unstructured resolver (Colab-style)
|
| 253 |
+
# =========================
|
| 254 |
+
def resolve_customer_id(unstructured_text: str):
|
| 255 |
+
s = (unstructured_text or "").strip()
|
| 256 |
+
if not s:
|
| 257 |
+
return None, "β Please enter Applicant Name or ID."
|
| 258 |
+
|
| 259 |
+
# 1) Extract an ID from any sentence
|
| 260 |
+
m = re.search(r"\b(\d{3,})\b", s)
|
| 261 |
+
if m:
|
| 262 |
+
cid = m.group(1)
|
| 263 |
+
prof = get_customer_profile(cid)
|
| 264 |
+
if prof:
|
| 265 |
+
return cid, f"β
Found ID {cid}: {prof['Name']}"
|
| 266 |
+
return None, f"β No such customer ID: {cid}"
|
| 267 |
+
|
| 268 |
+
# 2) Otherwise treat as name search (contains)
|
| 269 |
+
results = DF_CREDIT[DF_CREDIT["Name"].astype(str).str.contains(s, case=False, na=False)]
|
| 270 |
+
if results.empty:
|
| 271 |
+
return None, f"β No such customer: '{s}'"
|
| 272 |
+
|
| 273 |
+
if len(results) == 1:
|
| 274 |
+
cid = str(results.iloc[0]["ID"])
|
| 275 |
+
nm = str(results.iloc[0]["Name"])
|
| 276 |
+
return cid, f"β
Found Name '{nm}' -> ID {cid}"
|
| 277 |
+
|
| 278 |
+
# Multiple matches: no dropdown, just tell user to type ID
|
| 279 |
+
opts = ", ".join([f"{r['Name']} (ID {r['ID']})" for _, r in results.iterrows()])
|
| 280 |
+
return None, f"β οΈ Multiple customers match '{s}'. Please enter ID. Matches: {opts}"
|
| 281 |
+
|
| 282 |
+
|
| 283 |
+
# =========================
|
| 284 |
+
# Prompts / Chains
|
| 285 |
+
# =========================
|
| 286 |
+
QA_PROMPT = PromptTemplate(
|
| 287 |
+
input_variables=["customer_data", "policy_rules", "question"],
|
| 288 |
+
template="""
|
| 289 |
+
You are a helpful banking assistant. Answer the user's question based strictly on the provided Customer Data and Policy Rules.
|
| 290 |
|
| 291 |
CUSTOMER DATA:
|
| 292 |
{customer_data}
|
|
|
|
| 299 |
|
| 300 |
ANSWER:
|
| 301 |
"""
|
|
|
|
|
|
|
|
|
|
| 302 |
)
|
| 303 |
|
| 304 |
+
ADVICE_PROMPT = PromptTemplate(
|
| 305 |
+
input_variables=["customer_data", "policy_rules", "question"],
|
| 306 |
+
template="""You are a helpful loan officer assistant.
|
| 307 |
|
| 308 |
+
Write in THIRD PERSON about the customer. Always use the customer's Name and possessive.
|
|
|
|
| 309 |
Never address the reader as 'you' or 'your'.
|
| 310 |
|
| 311 |
+
You must provide ADVICE/RECOMMENDATION (not just restating risk and rate).
|
| 312 |
Use ONLY the provided Customer Data and Policy Rules.
|
| 313 |
|
| 314 |
REQUIREMENTS:
|
| 315 |
+
- Provide 3-5 actionable advice points (short sentences).
|
| 316 |
+
- Include a clear final recommendation: APPROVE or NOT RECOMMEND / REJECT.
|
| 317 |
+
- If customer is Non-Singaporean and PR_Status is False, you MUST recommend NOT RECOMMEND / REJECT regardless of risk level.
|
| 318 |
- Keep it concise.
|
| 319 |
|
| 320 |
CUSTOMER DATA:
|
|
|
|
| 328 |
|
| 329 |
ANSWER:
|
| 330 |
"""
|
|
|
|
|
|
|
|
|
|
| 331 |
)
|
| 332 |
|
| 333 |
+
REPORT_PROMPT = PromptTemplate(
|
| 334 |
+
input_variables=["customer_data", "policy_rules"],
|
| 335 |
+
template="""
|
| 336 |
+
You are a senior loan officer. Generate a comprehensive loan assessment report based on the provided customer data and banking policies.
|
| 337 |
+
Analyze the customer's profile, determine overall risk, calculate interest rate, and provide a clear recommendation.
|
| 338 |
+
Follow any mandatory exceptions.
|
| 339 |
|
| 340 |
CUSTOMER DATA:
|
| 341 |
{customer_data}
|
|
|
|
| 345 |
|
| 346 |
REPORT:
|
| 347 |
"""
|
|
|
|
|
|
|
|
|
|
| 348 |
)
|
| 349 |
|
| 350 |
|
| 351 |
+
# =========================
|
| 352 |
+
# Global init (Spaces safe)
|
| 353 |
+
# =========================
|
| 354 |
+
RETRIEVER = None
|
| 355 |
+
POLICY_FULL = ""
|
| 356 |
+
LLM = None
|
|
|
|
| 357 |
|
| 358 |
+
INIT_ERROR = None
|
| 359 |
|
| 360 |
+
try:
|
| 361 |
+
# Check files
|
| 362 |
+
missing = [p for p in [CSV_FILE, RISK_PDF, INT_PDF] if not os.path.exists(p)]
|
| 363 |
+
if missing:
|
| 364 |
+
raise RuntimeError("Missing required files in repo root:\n" + "\n".join(missing))
|
| 365 |
|
| 366 |
+
# Clean key (fixes illegal header newline)
|
| 367 |
+
api_key = os.getenv("OPENAI_API_KEY", "")
|
| 368 |
+
api_key = api_key.strip()
|
| 369 |
+
os.environ["OPENAI_API_KEY"] = api_key
|
| 370 |
if not api_key:
|
| 371 |
raise RuntimeError("Missing OPENAI_API_KEY. Set it in Space Secrets.")
|
| 372 |
|
| 373 |
+
# Load CSV
|
| 374 |
+
DF_ALL = load_customer_csv(CSV_FILE)
|
| 375 |
+
DF_CREDIT, DF_ACCOUNT, DF_GOV = build_mock_systems(DF_ALL)
|
| 376 |
|
| 377 |
+
# Load PDFs
|
| 378 |
+
RISK_TEXT = extract_pdf_text(RISK_PDF)
|
| 379 |
+
INT_TEXT = extract_pdf_text(INT_PDF)
|
| 380 |
+
POLICY_FULL = RISK_TEXT + "\n\n" + INT_TEXT
|
| 381 |
|
| 382 |
+
# Parse policies
|
| 383 |
+
RISK_MAPPING, INTEREST_RATES = parse_policies(RISK_TEXT, INT_TEXT)
|
|
|
|
| 384 |
|
| 385 |
# LLM
|
| 386 |
model_name = os.getenv("OPENAI_MODEL", "gpt-4o")
|
| 387 |
+
LLM = ChatOpenAI(model=model_name, temperature=0, openai_api_key=api_key)
|
| 388 |
|
| 389 |
# Chains (LCEL)
|
| 390 |
+
QA_CHAIN = QA_PROMPT | LLM
|
| 391 |
+
ADVICE_CHAIN = ADVICE_PROMPT | LLM
|
| 392 |
+
REPORT_CHAIN = REPORT_PROMPT | LLM
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 393 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 394 |
except Exception as e:
|
| 395 |
INIT_ERROR = str(e)
|
| 396 |
|
| 397 |
|
| 398 |
+
def build_retriever_if_needed():
|
| 399 |
+
global RETRIEVER
|
| 400 |
+
if RETRIEVER is not None:
|
| 401 |
+
return RETRIEVER
|
| 402 |
|
| 403 |
+
api_key = os.getenv("OPENAI_API_KEY", "").strip()
|
| 404 |
+
if not api_key:
|
| 405 |
+
return None
|
| 406 |
+
|
| 407 |
+
docs = [
|
| 408 |
+
Document(page_content=RISK_TEXT, metadata={"source": "Risk_Policy.pdf"}),
|
| 409 |
+
Document(page_content=INT_TEXT, metadata={"source": "Interest_Rate_Policy.pdf"}),
|
| 410 |
+
]
|
| 411 |
+
embeddings = OpenAIEmbeddings(openai_api_key=api_key)
|
| 412 |
+
db = FAISS.from_documents(docs, embeddings)
|
| 413 |
+
RETRIEVER = db.as_retriever()
|
| 414 |
+
return RETRIEVER
|
| 415 |
|
| 416 |
|
| 417 |
+
def get_policy_context(use_rag: bool) -> str:
|
| 418 |
if not use_rag:
|
| 419 |
return POLICY_FULL
|
| 420 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 421 |
try:
|
| 422 |
+
retriever = build_retriever_if_needed()
|
| 423 |
+
if retriever is None:
|
| 424 |
+
return POLICY_FULL + "\n\n[Note] RAG unavailable (missing API key). Using full policy text."
|
| 425 |
+
|
| 426 |
+
docs = retriever.invoke("risk level interest rate PR status")
|
| 427 |
+
ctx = "\n\n".join([d.page_content for d in docs]).strip()
|
| 428 |
return ctx if ctx else POLICY_FULL
|
| 429 |
except Exception as e:
|
| 430 |
+
return POLICY_FULL + f"\n\n[Note] RAG failed, using full policy text: {e}"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 431 |
|
| 432 |
|
| 433 |
+
# =========================
|
| 434 |
+
# Main Run (Colab-style)
|
| 435 |
+
# =========================
|
| 436 |
+
def run_action(user_input: str, action: str, use_rag: bool):
|
| 437 |
if INIT_ERROR:
|
| 438 |
+
return f"β Initialization error:\n\n{INIT_ERROR}"
|
| 439 |
|
| 440 |
+
try:
|
| 441 |
+
customer_id, msg = resolve_customer_id(user_input)
|
| 442 |
+
if not customer_id:
|
| 443 |
+
return msg
|
| 444 |
+
|
| 445 |
+
profile = get_customer_profile(customer_id)
|
| 446 |
+
if not profile:
|
| 447 |
+
return f"β No such customer ID: {customer_id}"
|
| 448 |
+
|
| 449 |
+
profile_text = format_customer_profile(profile)
|
| 450 |
+
policy_context = get_policy_context(use_rag)
|
| 451 |
+
|
| 452 |
+
# Deterministic summary (optional but helps demo)
|
| 453 |
+
overall_risk = determine_overall_risk(profile["Credit_Score"], profile["Account_Status"])
|
| 454 |
+
rate = determine_interest_rate(overall_risk)
|
| 455 |
+
must_reject = is_non_singaporean_no_pr(customer_id)
|
| 456 |
+
|
| 457 |
+
det = [
|
| 458 |
+
msg,
|
| 459 |
+
"",
|
| 460 |
+
"Deterministic (Policy-based):",
|
| 461 |
+
f"- Overall risk: {overall_risk}",
|
| 462 |
+
f"- Interest rate: {rate:.3f}%",
|
| 463 |
+
]
|
| 464 |
+
if must_reject:
|
| 465 |
+
det.append("- Mandatory exception: NOT RECOMMEND / REJECT (Non-Singaporean without PR).")
|
| 466 |
+
det_text = "\n".join(det)
|
| 467 |
+
|
| 468 |
+
if action == "1) Check Risk & Interest":
|
| 469 |
+
question = f"What are the risk level and applicable interest rate for the customer {customer_id}?"
|
| 470 |
+
resp = QA_CHAIN.invoke({
|
| 471 |
+
"customer_data": profile_text,
|
| 472 |
+
"policy_rules": policy_context,
|
| 473 |
+
"question": question
|
| 474 |
+
})
|
| 475 |
+
return det_text + "\n\nAI Output:\n" + one_sentence_per_line(resp.content)
|
| 476 |
+
|
| 477 |
+
if action == "2) Advice / Recommendation":
|
| 478 |
+
question = f"What interest rate advice can be recommended for customer with Id {customer_id}?"
|
| 479 |
+
resp = ADVICE_CHAIN.invoke({
|
| 480 |
+
"customer_data": profile_text,
|
| 481 |
+
"policy_rules": policy_context,
|
| 482 |
+
"question": question
|
| 483 |
+
})
|
| 484 |
+
return det_text + "\n\nAI Output:\n" + one_sentence_per_line(enforce_third_person(resp.content, profile.get("Name", "Customer")))
|
| 485 |
+
|
| 486 |
+
# 3) FULL report
|
| 487 |
+
customer_data_with_exception = profile_text + (
|
| 488 |
+
"\n\nMANDATORY EXCEPTION (must follow): Non-Singaporean with PR_Status = False => NOT RECOMMENDED / REJECTED."
|
| 489 |
+
if must_reject else ""
|
| 490 |
+
)
|
| 491 |
+
full_report = REPORT_CHAIN.invoke({
|
| 492 |
+
"customer_data": customer_data_with_exception,
|
| 493 |
+
"policy_rules": policy_context
|
| 494 |
})
|
| 495 |
+
final_text = apply_mandatory_exception_to_report(full_report.content, customer_id)
|
| 496 |
+
return det_text + "\n\nFull Report:\n" + one_sentence_per_line(final_text)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 497 |
|
| 498 |
+
except Exception:
|
| 499 |
+
return "β Run failed:\n\n" + traceback.format_exc()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 500 |
|
| 501 |
|
| 502 |
+
# =========================
|
| 503 |
+
# Gradio UI (NO dropdown)
|
| 504 |
+
# =========================
|
| 505 |
with gr.Blocks(title="Bank Loan Officer System") as demo:
|
| 506 |
+
gr.Markdown("# π¦ Bank Loan Officer System (Unstructured Input Search)")
|
| 507 |
+
gr.Markdown("Type Applicant **Name / ID / sentence**. The system resolves and responds like your Colab notebook.")
|
| 508 |
|
| 509 |
if INIT_ERROR:
|
| 510 |
gr.Markdown(f"## β Initialization error\n\n```\n{INIT_ERROR}\n```")
|
| 511 |
|
| 512 |
+
user_input = gr.Textbox(
|
| 513 |
+
label="Applicant Name or ID (unstructured)",
|
| 514 |
+
placeholder="e.g. 3333 OR Hilda OR 'please check loan for 3333'"
|
| 515 |
+
)
|
| 516 |
|
| 517 |
+
action = gr.Radio(
|
| 518 |
+
label="Action",
|
| 519 |
+
choices=[
|
| 520 |
+
"1) Check Risk & Interest",
|
| 521 |
+
"2) Advice / Recommendation",
|
| 522 |
+
"3) FULL Formal Loan Report"
|
| 523 |
+
],
|
| 524 |
+
value="1) Check Risk & Interest"
|
| 525 |
+
)
|
| 526 |
+
|
| 527 |
+
use_rag = gr.Checkbox(
|
| 528 |
+
label="Use RAG (FAISS embeddings). If it fails, auto fallback.",
|
| 529 |
+
value=False
|
| 530 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 531 |
|
| 532 |
run_btn = gr.Button("π Run")
|
| 533 |
+
output = gr.Textbox(label="Output", lines=28)
|
| 534 |
+
|
| 535 |
+
run_btn.click(fn=run_action, inputs=[user_input, action, use_rag], outputs=[output])
|
| 536 |
+
user_input.submit(fn=run_action, inputs=[user_input, action, use_rag], outputs=[output])
|
| 537 |
|
|
|
|
| 538 |
|
| 539 |
+
# =========================
|
| 540 |
+
# HF Spaces launch (the add-on at bottom)
|
| 541 |
+
# =========================
|
| 542 |
PORT = int(os.environ.get("PORT", 7860))
|
| 543 |
demo.queue().launch(
|
| 544 |
server_name="0.0.0.0",
|