shangfangjian/Profeto
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

安全加固 + 数据管线修复 + 质量改进(第2批)

Browse Source 
安全加固:
- /api/v1/predict 限流改用 get_client_ip 防 X-Forwarded-For 伪造
- 新增 TRUST_PROXY_HEADERS/REQUIRE_ADMIN_AUTH 配置,默认 fail-closed
- 生产环境管理接口未配置鉴权时拒绝(503),不再放行

数据管线修复:
- injuries IntegrityError 改用 begin_nested(SAVEPOINT)隔离批次
- injuries 空名单 vs 未配置语义区分(has_data 精确标记)
- bzzoiro 统计字段映射(shots/possession/cards) + 入库条件放宽
- available_at 语义收紧(回测防泄漏)
- team_names NFKD 去变音双重查找 + 补齐变体键

预测系统改进:
- multi-agent 全专家失败时 status=degraded 跳过终裁
- agent_weights 独立持久化到 predictions 表

新增迁移:
- 0014_predictions_agent_weights.py

新增测试(8个文件):
- test_ip_spoofing.py: 限流防伪造
- test_require_admin_fail_closed.py: 生产 fail-closed
- test_injuries_integrity_rollback.py: SAVEPOINT 隔离
- test_injuries_slice_semantic.py: 空名单 vs 未配置
- test_available_at.py: 回测防泄漏
- test_bzzoirot_stats.py: 统计字段映射
- test_team_names_normalize.py: NFKD 变体
- test_multi_agent_degraded.py: 全失败 degraded
- test_agent_weights_persist.py: 权重持久化
This commit is contained in:
Profeto Agent
2026-09-19 08:02:57 +00:00
parent bee330f31f
commit 6c24672e87
21 changed files with 1580 additions and 87 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llm/agents/orchestrator.py
+53 -20
View File
@@ -243,11 +243,35 @@ async def predict_match_multi(
# 2. 并行专家(各自独立配置,使用统一 cutoff)
reports = await run_specialists(header, version=version, before=cutoff)
# 3. 终裁
aggregator_provider = await _agent_provider("aggregator", tier="aggregator")
final, agg_prompt_tokens, agg_completion_tokens = await run_aggregator(
header, reports, provider=aggregator_provider, version=version
)
# 2.5 统计有效专家报告数量
ok_reports = [r for r in reports if r.status == "ok"]
has_valid_data = len(ok_reports) > 0
# 3. 终裁(仅当有有效专家报告时执行)
if has_valid_data:
aggregator_provider = await _agent_provider("aggregator", tier="aggregator")
final, agg_prompt_tokens, agg_completion_tokens = await run_aggregator(
header, reports, provider=aggregator_provider, version=version
)
else:
# 所有专家无数据/均失败:跳过终裁,标记 degraded
logger.warning(
"match %s: 所有 %d 位专家均无有效数据,跳过终裁,标记 degraded",
match_id, len(reports),
)
aggregator_provider = await _agent_provider("aggregator", tier="aggregator")
final = {
"pred_home_goals": None,
"pred_away_goals": None,
"alt_pred_home_goals": None,
"alt_pred_away_goals": None,
"pred_1x2": None,
"subjective_confidence": None,
"reasoning": f"所有 {len(reports)} 位专家均无有效数据或预测失败(状态: {','.join(r.status for r in reports)})",
"agent_weights": {},
}
agg_prompt_tokens = 0
agg_completion_tokens = 0
latency_ms = int((time.perf_counter() - start) * 1000)
@@ -262,15 +286,23 @@ async def predict_match_multi(
if m is None:
raise ValueError(f"match {match_id} not found")
# 严格校验终裁输出
# 根据是否有有效数据决定校验策略
from src.llm.validation import validate_agent_weights, validate_prediction_output
try:
validated = validate_prediction_output(final)
except Exception as e:
raise RuntimeError(f"终裁输出校验失败: {e}")
# agent_weights 同样必须过校验(旧实现直接取 raw 值落库,未做任何检查)
agent_weights = validate_agent_weights(final.get("agent_weights"))
if has_valid_data:
# 有有效报告:严格校验终裁输出
try:
validated = validate_prediction_output(final)
except Exception as e:
raise RuntimeError(f"终裁输出校验失败: {e}")
agent_weights = validate_agent_weights(final.get("agent_weights"))
pred_status = "success"
else:
# 无有效报告:跳过严格校验,直接构造降级结果
validated = None # type: ignore
agent_weights = {}
pred_status = "degraded"
pred = await _upsert_prediction(
session,
match_id=match_id,
@@ -283,16 +315,17 @@ async def predict_match_multi(
"prompt_tokens": sum(r.prompt_tokens or 0 for r in reports) + agg_prompt_tokens,
"completion_tokens": sum(r.completion_tokens or 0 for r in reports) + agg_completion_tokens,
"latency_ms": latency_ms,
"pred_home_goals": validated.pred_home_goals,
"pred_away_goals": validated.pred_away_goals,
"alt_pred_home_goals": validated.alt_pred_home_goals,
"alt_pred_away_goals": validated.alt_pred_away_goals,
"pred_1x2": validated.pred_1x2,
"subjective_confidence": validated.subjective_confidence,
"reasoning": validated.reasoning,
"pred_home_goals": validated.pred_home_goals if validated else None,
"pred_away_goals": validated.pred_away_goals if validated else None,
"alt_pred_home_goals": validated.alt_pred_home_goals if validated else None,
"alt_pred_away_goals": validated.alt_pred_away_goals if validated else None,
"pred_1x2": validated.pred_1x2 if validated else None,
"subjective_confidence": validated.subjective_confidence if validated else None,
"reasoning": validated.reasoning if validated else final.get("reasoning", ""),
"raw_response": final,
"agent_outputs": [r.to_dict() for r in reports],
"status": "success",
"agent_weights": agent_weights,
"status": pred_status,
"match_kickoff_at": match_kickoff_at,
"prediction_cutoff_at": prediction_cutoff_at,
"prediction_created_at": now,
src/llm/context_builder.py
+57 -17
View File
@@ -40,11 +40,23 @@ def _outcome(home_goals: int, away_goals: int, side: str) -> str:
def _is_stats_available(stats, before) -> bool:
"""检查统计数据在 cutoff 时间是否已可用。"""
"""检查统计数据在 cutoff 时间是否已可用。
available_at 语义:该条统计「对外可被使用」的最早时间,
至少不得早于比赛结束。用于回测防泄漏。
规则:
- before is None(实盘):available_at 为 None 时允许(兼容旧数据)
- before is not None(回测):available_at 为 None 视为不可用(保守)
- available_at > cutoff:不可用(数据在 cutoff 之后才生成)
"""
if before is None:
# 实盘模式:无时间信息时允许(兼容旧数据)
return True
# 回测模式(cutoff 不为 None):
# available_at 为 None → 无法确认是否在 cutoff 前可用,保守视为不可用
if stats.available_at is None:
return True # 无时间信息时保守处理:允许使用
return False
return stats.available_at <= before
@@ -330,34 +342,62 @@ async def injuries_slice(header: MatchHeader, *, before=None, db: AsyncSession |
before=cutoff: 只使用 cutoff 之前已采集的伤停数据,防回测泄漏。
db: 可选共享 session(见模块 docstring)。
语义区分:
- 查询成功 + 空结果 → has_data=True(明确知道「无人伤停」)
- 源未配置 / 查询失败 → has_data=False(无法判断,跳过 LLM)
"""
from src.data.injuries import get_injuries_for_match
from src.data.injuries import get_injuries_for_match, InjuryQueryResult
cutoff = before or header.match_dt
if db is not None:
home_injuries = await get_injuries_for_match(db, header.home_team_id, cutoff, as_of=cutoff)
away_injuries = await get_injuries_for_match(db, header.away_team_id, cutoff, as_of=cutoff)
home_result = await get_injuries_for_match(db, header.home_team_id, cutoff, as_of=cutoff)
away_result = await get_injuries_for_match(db, header.away_team_id, cutoff, as_of=cutoff)
else:
async with AsyncSessionLocal() as new_db:
home_injuries = await get_injuries_for_match(new_db, header.home_team_id, cutoff, as_of=cutoff)
away_injuries = await get_injuries_for_match(new_db, header.away_team_id, cutoff, as_of=cutoff)
home_result = await get_injuries_for_match(new_db, header.home_team_id, cutoff, as_of=cutoff)
away_result = await get_injuries_for_match(new_db, header.away_team_id, cutoff, as_of=cutoff)
# 判断是否有有效查询结果
# 两队都成功查询(即使为空) → has_data=True
# 任一查询失败或源未配置 → has_data=False
both_succeeded = (
home_result.query_status == "success"
and away_result.query_status == "success"
)
any_configured = (
home_result.query_status != "source_not_configured"
or away_result.query_status != "source_not_configured"
)
lines = ["── 阵容完整性 ──"]
n_records = 0
for label, injuries in (("主队", home_injuries), ("客队", away_injuries)):
if injuries:
n_records += len(injuries)
lines.append(f" {label}伤停({len(injuries)}人):")
for inj in injuries[:8]: # 最多显示 8 条
for label, result in (("主队", home_result), ("客队", away_result)):
if result.query_status == "source_not_configured":
lines.append(f" {label}: 伤停源未配置")
elif result.query_status == "query_error":
lines.append(f" {label}: 查询异常")
elif result.records:
n_records += len(result.records)
lines.append(f" {label}伤停({len(result.records)}人):")
for inj in result.records[:8]:
reason = inj.reason or inj.injury_type or "未知"
lines.append(f" - {inj.player_name}: {reason}")
if len(injuries) > 8:
lines.append(f" ...及其他 {len(injuries) - 8} 人")
if len(result.records) > 8:
lines.append(f" ...及其他 {len(result.records) - 8} 人")
else:
lines.append(f" {label}: 无伤停数据")
# 查询成功但无人伤停
lines.append(f" {label}: 当前无伤停记录")
if n_records == 0:
return SliceResult(text="── 阵容完整性 ──\n 无数据", has_data=False, n_records=0)
# 决定 has_data:
# - 两队都成功查询(即使为空) → True(明确知道名单)
# - 源未配置且无数据 → False
has_data = both_succeeded or (any_configured and n_records > 0)
if not has_data:
# 保留详细状态文案(伤停源未配置/查询异常),而非通用「无数据」
return SliceResult(text="\n".join(lines), has_data=False, n_records=0)
return SliceResult(text="\n".join(lines), has_data=True, n_records=n_records)