151 lines
5.7 KiB
Python
151 lines
5.7 KiB
Python
"""Fraud amount assessment and inquiry suggestion generation."""
|
||
import logging
|
||
from uuid import UUID
|
||
|
||
import httpx
|
||
from sqlalchemy import select
|
||
from sqlalchemy.ext.asyncio import AsyncSession
|
||
|
||
from app.core.config import settings
|
||
from app.models.transaction import TransactionRecord
|
||
from app.models.assessment import FraudAssessment, ConfidenceLevel, ReviewStatus
|
||
from app.rules.assessment_rules import classify_transaction
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
|
||
async def assess_case(case_id: UUID, db: AsyncSession) -> list[FraudAssessment]:
|
||
"""Run rule-based assessment on all non-duplicate transactions and generate reasons."""
|
||
result = await db.execute(
|
||
select(TransactionRecord)
|
||
.where(TransactionRecord.case_id == case_id)
|
||
.where(TransactionRecord.is_duplicate.is_(False))
|
||
.order_by(TransactionRecord.trade_time.asc())
|
||
)
|
||
transactions = list(result.scalars().all())
|
||
|
||
assessments: list[FraudAssessment] = []
|
||
for tx in transactions:
|
||
level, reason, exclude_reason = classify_transaction(tx)
|
||
|
||
fa = FraudAssessment(
|
||
case_id=case_id,
|
||
transaction_id=tx.id,
|
||
confidence_level=level,
|
||
assessed_amount=float(tx.amount) if level != ConfidenceLevel.low else 0,
|
||
reason=reason,
|
||
exclude_reason=exclude_reason,
|
||
review_status=ReviewStatus.pending,
|
||
)
|
||
db.add(fa)
|
||
assessments.append(fa)
|
||
|
||
await db.flush()
|
||
|
||
# try to enhance reasons via LLM
|
||
if settings.LLM_API_KEY and settings.LLM_API_URL:
|
||
for fa in assessments:
|
||
try:
|
||
enhanced = await _enhance_reason_via_llm(fa, transactions)
|
||
if enhanced:
|
||
fa.reason = enhanced
|
||
except Exception as e:
|
||
logger.debug("LLM reason enhancement skipped: %s", e)
|
||
await db.flush()
|
||
|
||
return assessments
|
||
|
||
|
||
async def generate_inquiry_suggestions(case_id: UUID, db: AsyncSession) -> list[str]:
|
||
"""Generate interview / inquiry suggestions based on assessment results."""
|
||
result = await db.execute(
|
||
select(FraudAssessment)
|
||
.where(FraudAssessment.case_id == case_id)
|
||
.order_by(FraudAssessment.created_at.asc())
|
||
)
|
||
assessments = list(result.scalars().all())
|
||
|
||
if not assessments:
|
||
return ["暂无分析结果，请先执行案件分析。"]
|
||
|
||
# try LLM generation
|
||
if settings.LLM_API_KEY and settings.LLM_API_URL:
|
||
try:
|
||
return await _generate_suggestions_via_llm(assessments)
|
||
except Exception as e:
|
||
logger.debug("LLM suggestions skipped: %s", e)
|
||
|
||
return _generate_suggestions_rule_based(assessments)
|
||
|
||
|
||
def _generate_suggestions_rule_based(assessments: list[FraudAssessment]) -> list[str]:
|
||
suggestions: list[str] = []
|
||
pending = [a for a in assessments if a.review_status == ReviewStatus.pending]
|
||
medium = [a for a in assessments if a.confidence_level == ConfidenceLevel.medium]
|
||
|
||
if pending:
|
||
suggestions.append(
|
||
f"有 {len(pending)} 笔交易尚未确认，建议逐笔向受害人核实是否受到诱导操作。"
|
||
)
|
||
if medium:
|
||
suggestions.append(
|
||
"部分交易置信度为中等，建议追问受害人交易的具体背景和对方的诱导话术。"
|
||
)
|
||
suggestions.append("是否还有其他未截图的转账记录或 APP 需要补充？")
|
||
suggestions.append("涉案金额中是否有已部分追回或返还的款项？")
|
||
suggestions.append(
|
||
"除了截图所示的 APP 外，是否还存在银行柜台、ATM、其他支付平台等转账渠道？"
|
||
)
|
||
return suggestions
|
||
|
||
|
||
async def _enhance_reason_via_llm(fa: FraudAssessment, all_tx: list) -> str | None:
|
||
prompt = (
|
||
f"这笔交易金额{fa.assessed_amount}元，置信等级{fa.confidence_level.value}。"
|
||
f"原始认定理由：{fa.reason}。"
|
||
"请用简洁中文优化认定理由表述，使之适合出现在办案文书中。只返回优化后的理由文字。"
|
||
)
|
||
async with httpx.AsyncClient(timeout=15) as client:
|
||
resp = await client.post(
|
||
settings.LLM_API_URL,
|
||
headers={"Authorization": f"Bearer {settings.LLM_API_KEY}"},
|
||
json={
|
||
"model": settings.LLM_MODEL,
|
||
"messages": [{"role": "user", "content": prompt}],
|
||
"max_tokens": 300,
|
||
},
|
||
)
|
||
resp.raise_for_status()
|
||
return resp.json()["choices"][0]["message"]["content"].strip()
|
||
|
||
|
||
async def _generate_suggestions_via_llm(assessments: list[FraudAssessment]) -> list[str]:
|
||
summary_lines = []
|
||
for a in assessments:
|
||
summary_lines.append(
|
||
f"- 金额{a.assessed_amount}元, 置信{a.confidence_level.value}, "
|
||
f"状态{a.review_status.value}, 理由: {a.reason[:60]}"
|
||
)
|
||
summary = "\n".join(summary_lines)
|
||
|
||
prompt = (
|
||
"你是一名反诈案件办案助手。以下是某诈骗案件的交易认定摘要：\n"
|
||
f"{summary}\n\n"
|
||
"请生成5条笔录辅助问询建议，帮助民警追问受害人以完善证据链。"
|
||
"只返回JSON数组格式的5个字符串。"
|
||
)
|
||
import json
|
||
async with httpx.AsyncClient(timeout=20) as client:
|
||
resp = await client.post(
|
||
settings.LLM_API_URL,
|
||
headers={"Authorization": f"Bearer {settings.LLM_API_KEY}"},
|
||
json={
|
||
"model": settings.LLM_MODEL,
|
||
"messages": [{"role": "user", "content": prompt}],
|
||
"max_tokens": 600,
|
||
},
|
||
)
|
||
resp.raise_for_status()
|
||
text = resp.json()["choices"][0]["message"]["content"].strip()
|
||
return json.loads(text.strip().strip("`").removeprefix("json").strip())
|