Spaces:
Sleeping
Sleeping
File size: 29,366 Bytes
ccb470a |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 |
"""
Fraud Detector Analyst β LangChain + (optional) MCP
Advanced βprototype-firstβ build:
- Chat uses chat-completion models (LangChain ChatHuggingFace).
- AI Summary shows a notice when no inference is connected.
LLM env (serverless friendly):
HF_TOKEN (or HF_SPACES)
LC_CHAT_MODEL (default: "Qwen/Qwen2.5-0.5B-Instruct")
LC_CHAT_MODEL_FALLBACK (default: "mistralai/Mistral-7B-Instruct")
Summary behavior:
If no working inference/token -> summary fields display:
"π Please connect to an inference point to generate summary."
Optional MCP:
ENABLE_MCP=1
MCP_SANCTIONS_URL, MCP_HIGH_RISK_MCC_URL
MCP_AUTH_HEADER="Authorization: Bearer <token>"
Run:
pip install -r requirements.txt
python app.py
On Spaces:
Add secret HF_TOKEN (or HF_SPACES). Launch.
"""
from __future__ import annotations
import os, io, re, json, math, unicodedata, logging
from typing import Optional, Tuple, List, Dict
import numpy as np
import pandas as pd
import gradio as gr
from dotenv import load_dotenv
# LangChain
from langchain.tools import tool
from langchain_core.tools import Tool
from langchain.agents import initialize_agent, AgentType
from langchain.schema import HumanMessage, SystemMessage
from pydantic import BaseModel, Field
from langchain_huggingface import HuggingFaceEndpoint, ChatHuggingFace
# Phone normalization
try:
import phonenumbers
HAVE_PHONENUM = True
except Exception:
HAVE_PHONENUM = False
# ------------------------
# Setup
# ------------------------
load_dotenv()
logging.basicConfig(level=logging.INFO)
log = logging.getLogger("fraud-analyst")
HF_TOKEN = os.getenv("HF_TOKEN") or os.getenv("HF_SPACES")
# Chat models (chat-completions)
DEFAULT_CHAT_MODEL = os.getenv("LC_CHAT_MODEL", "Qwen/Qwen2.5-0.5B-Instruct")
FALLBACK_CHAT_MODEL = os.getenv("LC_CHAT_MODEL_FALLBACK", "mistralai/Mistral-7B-Instruct")
SUMMARY_NOTICE = "π Please connect to an inference point to generate summary."
CHAT_NOTICE = "π Chat model not configured. Set HF_TOKEN and LC_CHAT_MODEL to enable chat."
# ------------------------
# LLM builders
# ------------------------
def _mk_chat_llm(model_id: str) -> ChatHuggingFace:
"""
ChatHuggingFace uses HF Inference under the hood.
Although the backend task is 'text-generation', this wrapper handles chat-style messages.
"""
base = HuggingFaceEndpoint(
repo_id=model_id,
task="text-generation",
huggingfacehub_api_token=HF_TOKEN,
max_new_tokens=256,
temperature=0.2,
repetition_penalty=1.05,
timeout=60,
)
return ChatHuggingFace(llm=base)
def _heartbeat_chat(model_id: str) -> bool:
try:
chat = _mk_chat_llm(model_id)
_ = chat.invoke([HumanMessage(content="ok")])
return True
except Exception as e:
log.warning(f"Heartbeat failed for {model_id}: {str(e)[:160]}")
return False
def build_chat_llm() -> Optional[ChatHuggingFace]:
"""
Returns a working ChatHuggingFace or None (if token/permissions missing).
"""
log.info(f"HF token present: {bool(HF_TOKEN)} len={len(HF_TOKEN) if HF_TOKEN else 0}")
if HF_TOKEN and _heartbeat_chat(DEFAULT_CHAT_MODEL):
log.info(f"Using chat model: {DEFAULT_CHAT_MODEL}")
return _mk_chat_llm(DEFAULT_CHAT_MODEL)
if HF_TOKEN and _heartbeat_chat(FALLBACK_CHAT_MODEL):
log.info(f"Using fallback chat model: {FALLBACK_CHAT_MODEL}")
return _mk_chat_llm(FALLBACK_CHAT_MODEL)
log.warning("No working chat model; chat will show a notice.")
return None
CHAT_LLM = build_chat_llm()
# ------------------------
# Normalization helpers
# ------------------------
def _norm_colname(c: str) -> str:
c = c.strip().lower()
c = re.sub(r"\s+", "_", c)
c = re.sub(r"[^\w]+", "_", c)
return c.strip("_")
def _nfkc(s: str) -> str:
return unicodedata.normalize("NFKC", s)
def _collapse_ws(s: str) -> str:
return re.sub(r"\s+", " ", s).strip()
def _clean_str(x):
if pd.isna(x): return x
return _collapse_ws(_nfkc(str(x)))
def _is_email(s: str) -> bool:
return bool(re.match(r"^[A-Za-z0-9._%+\-]+@[A-Za-z0-9.\-]+\.[A-Za-z]{2,}$", s or ""))
def _clean_phone(s: str, default_region: str = "IN"):
if s is None or str(s).strip() == "":
return None, "missing_phone"
raw = re.sub(r"[^\d+]", "", str(s))
if HAVE_PHONENUM:
try:
pn = phonenumbers.parse(raw, default_region)
if phonenumbers.is_possible_number(pn) and phonenumbers.is_valid_number(pn):
return phonenumbers.format_number(pn, phonenumbers.PhoneNumberFormat.E164), None
return raw, "invalid_phone"
except Exception:
return raw, "invalid_phone"
digits = re.sub(r"\D", "", raw)
return (digits, None) if 8 <= len(digits) <= 15 else (digits, "invalid_phone")
def _parse_datetime(s):
try:
return pd.to_datetime(s, errors="coerce", utc=True)
except Exception:
return pd.NaT
def _to_numeric(series: pd.Series):
coerced = pd.to_numeric(series, errors="coerce")
return coerced, (coerced.isna() & series.notna())
def _read_csv_any(file_obj) -> pd.DataFrame:
if file_obj is None:
raise ValueError("No file uploaded.")
if hasattr(file_obj, "name"):
p = file_obj.name
try: return pd.read_csv(p)
except Exception: return pd.read_csv(p, encoding="latin-1")
try: return pd.read_csv(file_obj)
except Exception:
file_obj.seek(0)
return pd.read_csv(file_obj, encoding="latin-1")
def _standardize_df(df: pd.DataFrame) -> pd.DataFrame:
df = df.copy()
df.columns = [_norm_colname(c) for c in df.columns]
for c in df.select_dtypes(include=["object"]).columns:
df[c] = df[c].apply(_clean_str)
return df
def _prepare_generic(df: pd.DataFrame, expected: Dict[str, List[str]]):
issues = []
df0 = _standardize_df(df)
# Synonym mapping
colmap = {}
cols = set(df0.columns)
for canon, syns in expected.items():
found = None
for s in [canon] + syns:
s = _norm_colname(s)
if s in cols:
found = s; break
if found: colmap[canon] = found
# Email/phone quality
for c in list(df0.columns):
if "email" in c:
df0[c] = df0[c].apply(lambda x: str(x).lower().strip() if pd.notna(x) else x)
for idx, v in df0[c].items():
if pd.isna(v) or str(v).strip()=="":
issues.append({"row": idx, "field": c, "issue":"missing_email","value":""})
elif not _is_email(v):
issues.append({"row": idx, "field": c, "issue":"invalid_email","value":str(v)})
if "phone" in c or "mobile" in c:
vals = []
for idx, v in df0[c].items():
e164, prob = _clean_phone(v)
vals.append(e164)
if prob: issues.append({"row": idx, "field": c, "issue":prob, "value":str(v)})
df0[c] = vals
# Datetime parsing
for c in df0.columns:
if any(k in c for k in ["date","time","timestamp","created_at","updated_at"]):
parsed = _parse_datetime(df0[c])
bad = parsed.isna() & df0[c].notna()
for idx in df0.index[bad]:
issues.append({"row": int(idx), "field": c, "issue":"unparseable_timestamp", "value":str(df0.loc[idx, c])})
df0[c] = parsed
# Numeric coercions for common fields
for nc in ["amount","credit_score","utilization","dti","recent_defaults","income"]:
for c in df0.columns:
if c == nc or c.endswith("_"+nc) or nc in c:
coerced, badmask = _to_numeric(df0[c])
for idx in df0.index[badmask]:
issues.append({"row": int(idx), "field": c, "issue":"non_numeric", "value":str(df0.loc[idx, c])})
df0[c] = coerced
issues_df = pd.DataFrame(issues, columns=["row","field","issue","value"])
missing = [k for k in expected.keys() if k not in colmap]
quality_summary = f"Rows={len(df0)}, Cols={len(df0.columns)}; Missing required fields: {missing if missing else 'None'}"
return df0, issues_df, quality_summary, colmap
# ------------------------
# Modules & Rules
# ------------------------
TX_EXPECTED = {
"transaction_id":["txn_id","transactionid","id","tx_id"],
"customer_id":["cust_id","user_id","client_id"],
"amount":["amt","amount_inr","value"],
"timestamp":["date","event_time","created_at","tx_time"],
"merchant_category":["mcc","merchant_cat","category"]
}
def prepare_transactions(df): return _prepare_generic(df, TX_EXPECTED)
def detect_transactions(clean_df, colmap, high_risk_mcc: Optional[List[str]]=None):
high_risk = set(["HIGH_RISK","GAMBLING","CRYPTO_EXCHANGE","ESCORTS","CASINO"])
if high_risk_mcc:
high_risk.update([_nfkc(x).strip().upper().replace(" ","_") for x in high_risk_mcc])
if not all(k in colmap for k in ["customer_id","amount"]):
return pd.DataFrame(), "Required columns missing for detection (need at least customer_id, amount)."
df = clean_df.copy()
reasons = []
amtcol = colmap.get("amount")
if amtcol is not None:
reasons.append(("large_amount>10k", df[amtcol] > 10000))
reasons.append(("negative_amount", df[amtcol] < 0))
if "merchant_category" in colmap:
mcc = colmap["merchant_category"]
high = df[mcc].astype(str).str.upper().str.replace(" ","_", regex=False).isin(high_risk)
reasons.append(("merchant_category_high_risk", high))
if all(k in colmap for k in ["customer_id","timestamp","amount"]):
cid, ts, amt = colmap["customer_id"], colmap["timestamp"], colmap["amount"]
daily = df.groupby([cid, df[ts].dt.date])[amt].transform("sum")
reasons.append(("daily_sum_per_customer>50k", daily > 50000))
mask = None
for _, m in reasons:
mask = m if mask is None else (mask | m)
flagged = df[mask] if mask is not None else pd.DataFrame()
if not flagged.empty:
rr=[]
for _, row in flagged.iterrows():
hits=[]
a = row[amtcol] if amtcol in flagged.columns else None
if pd.notna(a) and a>10000: hits.append("large_amount")
if pd.notna(a) and a<0: hits.append("negative_amount")
if "merchant_category" in colmap:
val = str(row[colmap["merchant_category"]]).upper().replace(" ","_")
if val in high_risk: hits.append("mcc_high_risk")
# daily sum check reconstructed
try:
if all(k in colmap for k in ["customer_id","timestamp","amount"]):
sub = df[(df[colmap["customer_id"]]==row[colmap["customer_id"]]) &
(df[colmap["timestamp"]].dt.date==pd.to_datetime(row[colmap["timestamp"]], errors="coerce").date())]
if sub[colmap["amount"]].sum() > 50000: hits.append("daily_sum>50k")
except Exception: pass
rr.append(", ".join(sorted(set(hits))) or "rule_hit")
flagged = flagged.assign(risk_reason=rr)
stats = f"Transactions flagged: {len(flagged)} of {len(df)}."
return flagged, stats
KYC_EXPECTED = {
"customer_id":["cust_id","user_id","client_id"],
"name":["full_name","customer_name"],
"email":["email_address","mail"],
"phone":["phone_number","mobile","contact"],
"dob":["date_of_birth","birthdate"]
}
def prepare_kyc(df): return _prepare_generic(df, KYC_EXPECTED)
def _age_years(dob: pd.Series) -> pd.Series:
now = pd.Timestamp.utcnow()
return (now - dob).dt.days / 365.25
def detect_kyc(clean_df, colmap):
if not all(k in colmap for k in ["customer_id","name"]):
return pd.DataFrame(), "Required columns missing for KYC (need at least customer_id, name)."
df = clean_df.copy()
reasons=[]
if "email" in colmap:
dupe_email = df.duplicated(subset=[colmap["email"]], keep=False) & df[colmap["email"]].notna()
reasons.append(("duplicate_email", dupe_email))
if "phone" in colmap:
dupe_phone = df.duplicated(subset=[colmap["phone"]], keep=False) & df[colmap["phone"]].notna()
reasons.append(("duplicate_phone", dupe_phone))
if "dob" in colmap:
age = _age_years(df[colmap["dob"]])
invalid = (df[colmap["dob"]].isna()) | (df[colmap["dob"]] > pd.Timestamp.utcnow()) | (age > 120)
reasons.append(("invalid_dob", invalid))
if "name" in colmap:
name = df[colmap["name"]].astype(str)
susp = name.str.isupper() | name.str.contains(r"\d") | (name.str.len()<3)
reasons.append(("suspicious_name", susp))
mask = None
for _, m in reasons:
mask = m if mask is None else (mask | m)
flagged = df[mask] if mask is not None else pd.DataFrame()
if not flagged.empty:
flagged = flagged.assign(risk_reason="kyc_rule_hit")
stats = f"KYC flagged: {len(flagged)} of {len(df)}."
return flagged, stats
SAN_EXPECTED = {"customer_id":["cust_id","user_id","client_id"], "name":["full_name","customer_name"]}
def prepare_sanctions(df): return _prepare_generic(df, SAN_EXPECTED)
DEMO_SANCTIONS = pd.DataFrame({"name":["Ivan Petrov","Global Terror Org","Acme Front LLC","John Doe (PEP)","Shadow Brokers"]})
def token_overlap(a: str, b: str) -> int:
at = set(re.findall(r"[A-Za-z0-9]+", a.lower()))
bt = set(re.findall(r"[A-Za-z0-9]+", b.lower()))
return len(at & bt)
def detect_sanctions(clean_df, colmap, sanctions_df: Optional[pd.DataFrame]=None):
if "name" not in colmap:
return pd.DataFrame(), "Required column missing for Sanctions (need name)."
df = clean_df.copy()
sanc = sanctions_df if sanctions_df is not None else DEMO_SANCTIONS.copy()
sanc = _standardize_df(sanc)
if "name" not in sanc.columns:
for c in sanc.columns:
if "name" in c: sanc = sanc.rename(columns={c:"name"}); break
sanc_names = sanc["name"].dropna().astype(str).tolist()
matches=[]
for idx, row in df.iterrows():
nm = str(row[colmap["name"]] or "").strip()
if not nm: continue
if any(nm.lower()==s.lower() for s in sanc_names):
matches.append((idx,"exact")); continue
if any(token_overlap(nm, s) >= 2 for s in sanc_names):
matches.append((idx,"fuzzy"))
flagged = df.loc[[i for i,_ in matches]].copy() if matches else pd.DataFrame()
if not flagged.empty:
mt = {i:t for i,t in matches}
flagged = flagged.assign(match_type=[mt.get(i,"") for i in flagged.index])
stats = f"Sanctions matches: {len(flagged)} of {len(df)}. (Using {'uploaded/MCP' if sanctions_df is not None else 'demo'} list)"
return flagged, stats
CR_EXPECTED = {
"customer_id":["cust_id","user_id","client_id"],
"credit_score":["creditscore","score"],
"utilization":["util","credit_utilization","utilization_ratio"],
"dti":["debt_to_income","debt_to_income_ratio"],
"recent_defaults":["defaults","recentdefaults"],
"income":["annual_income","salary"]
}
def prepare_credit(df): return _prepare_generic(df, CR_EXPECTED)
def detect_credit(clean_df, colmap):
needed = ["credit_score","utilization","dti","recent_defaults","income"]
if not any(k in colmap for k in needed):
return pd.DataFrame(), "Required columns missing for Credit Risk."
df = clean_df.copy()
cs = df[colmap.get("credit_score","credit_score")] if "credit_score" in colmap else pd.Series([np.nan]*len(df))
util= df[colmap.get("utilization","utilization")] if "utilization" in colmap else pd.Series([np.nan]*len(df))
dti = df[colmap.get("dti","dti")] if "dti" in colmap else pd.Series([np.nan]*len(df))
rde = df[colmap.get("recent_defaults","recent_defaults")] if "recent_defaults" in colmap else pd.Series([np.nan]*len(df))
inc = df[colmap.get("income","income")] if "income" in colmap else pd.Series([np.nan]*len(df))
out=[]
for i in range(len(df)):
hits=0; reasons=[]
if pd.notna(cs.iloc[i]) and cs.iloc[i] < 600: hits+=1; reasons.append("credit_score<600")
if pd.notna(util.iloc[i]) and util.iloc[i] > 0.8: hits+=1; reasons.append("utilization>0.8")
if pd.notna(dti.iloc[i]) and dti.iloc[i] > 0.4: hits+=1; reasons.append("DTI>0.4")
if pd.notna(rde.iloc[i]) and rde.iloc[i] > 0: hits+=1; reasons.append("recent_defaults>0")
if pd.notna(inc.iloc[i]) and inc.iloc[i] < 30000: hits+=1; reasons.append("income<30000")
level = "High" if hits>=3 else ("Medium" if hits==2 else ("Low" if hits==1 else "None"))
out.append((hits, level, ", ".join(reasons)))
risk_score=[x[0] for x in out]; risk_level=[x[1] for x in out]; reason=[x[2] for x in out]
res = df.assign(risk_score=risk_score, risk_level=risk_level, risk_reason=reason)
flagged = res[res["risk_level"].isin(["High","Medium","Low"]) & (res["risk_level"]!="None")]
stats = f"Credit Risk flagged: {len(flagged)} of {len(df)}. Distribution: High={(res['risk_level']=='High').sum()}, Medium={(res['risk_level']=='Medium').sum()}, Low={(res['risk_level']=='Low').sum()}."
return flagged, stats
# ------------------------
# Summarizer (notice-first)
# ------------------------
SUMMARY_SYS = "You are a helpful Fraud/Risk analyst. Be concise (<120 words), list key counts, drivers, and data quality caveats."
def summarize_ai(context: str) -> str:
"""
If chat LLM is available, use it to generate a short summary.
Otherwise return the prototype notice string.
"""
if CHAT_LLM is None:
return SUMMARY_NOTICE
try:
out = CHAT_LLM.invoke([SystemMessage(content=SUMMARY_SYS), HumanMessage(content=context[:4000])])
if hasattr(out, "content"): return out.content
return str(out)
except Exception as e:
msg = str(e)
if "401" in msg or "403" in msg:
return SUMMARY_NOTICE
return SUMMARY_NOTICE
# ------------------------
# Optional MCP
# ------------------------
from urllib.request import Request, urlopen
def _mcp_get_json(url: str, auth_header: Optional[str]):
try:
req = Request(url)
if auth_header:
k, v = auth_header.split(":", 1)
req.add_header(k.strip(), v.strip())
with urlopen(req, timeout=10) as r:
return json.loads(r.read().decode("utf-8"))
except Exception as e:
log.warning(f"MCP fetch failed: {e}")
return None
def mcp_fetch_sanctions() -> Optional[pd.DataFrame]:
if os.getenv("ENABLE_MCP","0") not in ("1","true","TRUE"): return None
url = os.getenv("MCP_SANCTIONS_URL")
if not url: return None
data = _mcp_get_json(url, os.getenv("MCP_AUTH_HEADER"))
if not data: return None
if isinstance(data, list):
if all(isinstance(x, dict) for x in data):
rows = [{"name": x.get("name") or x.get("Name")} for x in data if x.get("name") or x.get("Name")]
return pd.DataFrame(rows) if rows else None
if all(isinstance(x, str) for x in data):
return pd.DataFrame({"name": data})
return None
def mcp_fetch_high_risk_mcc() -> Optional[List[str]]:
if os.getenv("ENABLE_MCP","0") not in ("1","true","TRUE"): return None
url = os.getenv("MCP_HIGH_RISK_MCC_URL")
if not url: return None
data = _mcp_get_json(url, os.getenv("MCP_AUTH_HEADER"))
return [str(x) for x in data] if isinstance(data, list) else None
# ------------------------
# Pipelines (per tab)
# ------------------------
def run_transactions(file):
try:
df = _read_csv_any(file)
clean, issues, quality, colmap = prepare_transactions(df)
mcc = mcp_fetch_high_risk_mcc()
flagged, stats = detect_transactions(clean, colmap, mcc)
ctx = f"[Transactions]\n{stats}\nQuality: {quality}\nHead:\n{clean.head(5).to_csv(index=False)}\nFlagged:\n{flagged.head(5).to_csv(index=False)}"
ai = summarize_ai(ctx)
return ai, stats, flagged, issues
except Exception as e:
return f"Error: {e}", "Validation failed.", pd.DataFrame(), pd.DataFrame()
def run_kyc(file):
try:
df = _read_csv_any(file)
clean, issues, quality, colmap = prepare_kyc(df)
flagged, stats = detect_kyc(clean, colmap)
ctx = f"[KYC]\n{stats}\nQuality: {quality}\nHead:\n{clean.head(5).to_csv(index=False)}\nFlagged:\n{flagged.head(5).to_csv(index=False)}"
ai = summarize_ai(ctx)
return ai, stats, flagged, issues
except Exception as e:
return f"Error: {e}", "Validation failed.", pd.DataFrame(), pd.DataFrame()
def run_sanctions(customers_file, sanctions_file):
try:
df = _read_csv_any(customers_file)
clean, issues, quality, colmap = prepare_sanctions(df)
sanc_df = mcp_fetch_sanctions()
if sanc_df is None and sanctions_file is not None:
sanc_df = _read_csv_any(sanctions_file)
flagged, stats = detect_sanctions(clean, colmap, sanc_df)
ctx = f"[Sanctions]\n{stats}\nQuality: {quality}\nHead:\n{clean.head(5).to_csv(index=False)}\nMatches:\n{flagged.head(5).to_csv(index=False)}"
ai = summarize_ai(ctx)
return ai, stats, flagged, issues
except Exception as e:
return f"Error: {e}", "Validation failed.", pd.DataFrame(), pd.DataFrame()
def run_credit(file):
try:
df = _read_csv_any(file)
clean, issues, quality, colmap = prepare_credit(df)
flagged, stats = detect_credit(clean, colmap)
ctx = f"[Credit]\n{stats}\nQuality: {quality}\nHead:\n{clean.head(5).to_csv(index=False)}\nFlagged:\n{flagged.head(5).to_csv(index=False)}"
ai = summarize_ai(ctx)
return ai, stats, flagged, issues
except Exception as e:
return f"Error: {e}", "Validation failed.", pd.DataFrame(), pd.DataFrame()
# ------------------------
# Tools (CSV text in β concise text out)
# ------------------------
def _csv_text_to_df(csv_text: str) -> pd.DataFrame:
return pd.read_csv(io.StringIO(csv_text))
class TransactionCSVInput(BaseModel):
csv_text: str = Field(..., description="Transactions CSV text")
@tool("transactions_fraud_tool", args_schema=TransactionCSVInput)
def transactions_fraud_tool(csv_text: str) -> str:
df = _csv_text_to_df(csv_text)
clean, issues, quality, colmap = prepare_transactions(df)
flagged, stats = detect_transactions(clean, colmap)
return f"{stats}\nDQ issues: {len(issues)}\nFirst flagged:\n{flagged.head(5).to_csv(index=False)}"[:2800]
class KYCCSVInput(BaseModel):
csv_text: str = Field(..., description="KYC CSV text")
@tool("kyc_fraud_tool", args_schema=KYCCSVInput)
def kyc_fraud_tool(csv_text: str) -> str:
df = _csv_text_to_df(csv_text)
clean, issues, quality, colmap = prepare_kyc(df)
flagged, stats = detect_kyc(clean, colmap)
return f"{stats}\nDQ issues: {len(issues)}\nFirst flagged:\n{flagged.head(5).to_csv(index=False)}"[:2800]
class SanctionsCSVInput(BaseModel):
csv_text: str = Field(..., description="Customers CSV text with a 'name' column")
@tool("sanctions_pep_tool", args_schema=SanctionsCSVInput)
def sanctions_pep_tool(csv_text: str) -> str:
df = _csv_text_to_df(csv_text)
clean, issues, quality, colmap = prepare_sanctions(df)
flagged, stats = detect_sanctions(clean, colmap)
return f"{stats}\nDQ issues: {len(issues)}\nFirst matches:\n{flagged.head(5).to_csv(index=False)}"[:2800]
class CreditCSVInput(BaseModel):
csv_text: str = Field(..., description="Credit CSV text")
@tool("credit_risk_tool", args_schema=CreditCSVInput)
def credit_risk_tool(csv_text: str) -> str:
df = _csv_text_to_df(csv_text)
clean, issues, quality, colmap = prepare_credit(df)
flagged, stats = detect_credit(clean, colmap)
return f"{stats}\nDQ issues: {len(issues)}\nFirst flagged:\n{flagged.head(5).to_csv(index=False)}"[:2800]
TOOLS: List[Tool] = [
transactions_fraud_tool,
kyc_fraud_tool,
sanctions_pep_tool,
credit_risk_tool,
]
# ------------------------
# Agent (chat-completions)
# ------------------------
AGENT_SYSTEM = """You are an AI Consultant for Fraud/Risk.
You have tools for Transactions, KYC, Sanctions/PEP, and Credit Risk.
If the user pastes a small CSV snippet, pick the relevant tool and analyze it.
Be concise and actionable."""
def build_agent():
if CHAT_LLM is None:
class Stub:
def invoke(self, prompt): return CHAT_NOTICE
return Stub()
return initialize_agent(
TOOLS,
CHAT_LLM,
agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION,
verbose=False,
agent_kwargs={"system_message": AGENT_SYSTEM},
handle_parsing_errors=True,
)
AGENT = build_agent()
def agent_reply(history: List[Dict], user_msg: str):
try:
looks_like_csv = ("," in user_msg) and ("\n" in user_msg) and (user_msg.count(",") >= 2)
prompt = f"CSV snippet detected. Decide tool and analyze:\n\n{user_msg}" if looks_like_csv else user_msg
res = AGENT.invoke(prompt)
if isinstance(res, dict) and "output" in res: return res["output"]
return str(res)
except Exception as e:
return f"Agent error: {e}"
# ------------------------
# UI
# ------------------------
with gr.Blocks(title="Fraud Detector Analyst β LangChain + MCP", theme=gr.themes.Soft()) as demo:
gr.Markdown("# π‘οΈ Fraud Detector Analyst β LangChain + MCP")
gr.Markdown(
"This prototype runs **rules & data checks locally**. "
"Chat + AI summaries require a remote inference provider (HF Inference)."
)
with gr.Tabs():
with gr.Tab("Transactions"):
gr.Markdown("Upload a **transactions** CSV.")
tx_file = gr.File(file_types=[".csv"], label="Transactions CSV", type="binary")
tx_ai = gr.Textbox(label="AI Summary (requires inference)", value=SUMMARY_NOTICE, lines=6)
tx_stats = gr.Textbox(label="Stats", lines=3)
tx_flagged = gr.Dataframe(label="Flagged Transactions")
tx_issues = gr.Dataframe(label="Data Quality Issues (row, field, issue, value)")
tx_file.upload(run_transactions, inputs=[tx_file], outputs=[tx_ai, tx_stats, tx_flagged, tx_issues])
with gr.Tab("KYC"):
gr.Markdown("Upload a **KYC** CSV.")
kyc_file = gr.File(file_types=[".csv"], label="KYC CSV", type="binary")
kyc_ai = gr.Textbox(label="AI Summary (requires inference)", value=SUMMARY_NOTICE, lines=6)
kyc_stats = gr.Textbox(label="Stats", lines=3)
kyc_flagged = gr.Dataframe(label="Flagged KYC Rows")
kyc_issues = gr.Dataframe(label="Data Quality Issues")
kyc_file.upload(run_kyc, inputs=[kyc_file], outputs=[kyc_ai, kyc_stats, kyc_flagged, kyc_issues])
with gr.Tab("Sanctions/PEP"):
gr.Markdown("Upload **customers** CSV (+ optional sanctions CSV).")
san_customers = gr.File(file_types=[".csv"], label="Customers CSV", type="binary")
san_list = gr.File(file_types=[".csv"], label="Sanctions/PEP CSV (optional)", type="binary")
san_ai = gr.Textbox(label="AI Summary (requires inference)", value=SUMMARY_NOTICE, lines=6)
san_stats = gr.Textbox(label="Stats", lines=3)
san_flagged = gr.Dataframe(label="Matches")
san_issues = gr.Dataframe(label="Data Quality Issues")
san_customers.upload(run_sanctions, inputs=[san_customers, san_list], outputs=[san_ai, san_stats, san_flagged, san_issues])
san_list.upload(run_sanctions, inputs=[san_customers, san_list], outputs=[san_ai, san_stats, san_flagged, san_issues])
with gr.Tab("Credit Risk"):
gr.Markdown("Upload a **credit** CSV.")
cr_file = gr.File(file_types=[".csv"], label="Credit CSV", type="binary")
cr_ai = gr.Textbox(label="AI Summary (requires inference)", value=SUMMARY_NOTICE, lines=6)
cr_stats = gr.Textbox(label="Stats", lines=3)
cr_flagged = gr.Dataframe(label="Flagged Applicants")
cr_issues = gr.Dataframe(label="Data Quality Issues")
cr_file.upload(run_credit, inputs=[cr_file], outputs=[cr_ai, cr_stats, cr_flagged, cr_issues])
with gr.Tab("AI Consultant (Agent)"):
gr.Markdown("Paste a small CSV snippet or ask questions. Uses chat-completions when configured.")
chatbot = gr.Chatbot(type="messages", label="Fraud AI Consultant")
user_in = gr.Textbox(label="Message or CSV snippet")
send_btn = gr.Button("Send")
def _chat_fn(history, msg):
reply = agent_reply(history, msg)
history = (history or []) + [{"role":"user","content":msg}, {"role":"assistant","content":reply}]
return history, ""
send_btn.click(_chat_fn, inputs=[chatbot, user_in], outputs=[chatbot, user_in])
gr.Markdown(
"### βοΈ Enable inference\n"
"- Set **HF_TOKEN** (or HF_SPACES on Spaces)\n"
"- Optional: **LC_CHAT_MODEL** (default Qwen 0.5B Instruct), **LC_CHAT_MODEL_FALLBACK** (default Mistral 7B Instruct)\n"
"- Optional MCP: `ENABLE_MCP=1`, `MCP_SANCTIONS_URL`, `MCP_HIGH_RISK_MCC_URL`, `MCP_AUTH_HEADER`"
)
if __name__ == "__main__":
demo.launch(server_name="0.0.0.0", server_port=7860) |