- features.py:gate() 纯函数——轮数/字符估算/意图黑名单/仓库级信号/长度门 -> t1_hard_ok(任一硬门不过即 False,D-G1) - grader.py:Grader.decide §8 时序(embed 降级检查 -> 特征门 -> LinearHead 概率 -> conformal 阈值:p1>=τ1 且 t1_hard_ok->T1,p3>=τ3 或 repo_signals->T3, 其余 T2 默认;collect/shadow 只写观察 executed=现行为,live 决策即执行; 全模式 observer.log);工件/阈值缓存 + invalidate;D-G4 规则门退化 - routes:/v1/route 契约(D-G6 不落 query 原文) - fix(observer):embedding list -> BLOB 转换(修 sqlite 绑定) - 测试 +6(门矩阵/shadow 不改流/live 决策/降级/保守阈值/写观察),全量 405 passed
156 lines
5.9 KiB
Python
156 lines
5.9 KiB
Python
"""分级决策组合(T-G5,§8 主时序):
|
||
|
||
embedder.embed(挂 -> fallback=T2+规则门,D-G4)
|
||
-> features.gate(t1_hard_ok)
|
||
-> LinearHead.predict -> {p1,p2,p3}
|
||
-> conformal 阈值:p1>=τ1 且 t1_hard_ok -> T1;p3>=τ3 或 repo_signals -> T3;其余 T2
|
||
-> mode 裁剪:collect/shadow 只写观察(decided≠executed),live 返回决策
|
||
-> observer.log(全模式必写)
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import time
|
||
import uuid
|
||
from dataclasses import dataclass, field
|
||
from typing import Any, Dict, List, Optional
|
||
|
||
from gateway.sense.calibrate import load_active
|
||
from gateway.sense.classifier import LinearHead
|
||
from gateway.sense.config import SenseConfig
|
||
from gateway.sense.embedder import embed
|
||
from gateway.sense.errors import EmbedderDown
|
||
from gateway.sense.features import gate
|
||
|
||
|
||
@dataclass
|
||
class TierDecision:
|
||
tier: str # 决策档(live 时即执行档的候选)
|
||
probs: Dict[str, float]
|
||
hard_gates: Dict[str, Any]
|
||
thresholds_version: str = ""
|
||
head_version: str = ""
|
||
mode: str = "collect"
|
||
fallback: bool = False # True = embedder/工件故障,规则门退化
|
||
executed_tier: str = "" # 现行为执行的档位(collect/shadow 记录用)
|
||
|
||
|
||
def _rule_tier(feats) -> str:
|
||
"""规则门兜底(无模型时的决策):仓库级 -> T3;t1 硬门过 -> T1;否则 T2。"""
|
||
if feats.repo_signals:
|
||
return "T3"
|
||
if feats.t1_hard_ok:
|
||
return "T1"
|
||
return "T2"
|
||
|
||
|
||
class Grader:
|
||
"""决策器(持有 active 工件缓存;工件/阈值切换后调 invalidate)。"""
|
||
|
||
def __init__(self, cfg: SenseConfig, store, observer=None):
|
||
self.cfg = cfg
|
||
self.store = store
|
||
self.observer = observer
|
||
self._head: Optional[LinearHead] = None
|
||
self._head_loaded = False
|
||
self._thresholds: Optional[Dict[str, Any]] = None
|
||
|
||
def _load_head(self):
|
||
if not self._head_loaded:
|
||
art = self.store.active_artifact("head")
|
||
if art:
|
||
try:
|
||
self._head = LinearHead.load(art["path"], version=art["version"])
|
||
except Exception: # noqa: BLE001
|
||
self._head = None
|
||
self._head_loaded = True
|
||
return self._head
|
||
|
||
def invalidate(self) -> None:
|
||
"""工件 promote 后调用(重载 active 工件与阈值)。"""
|
||
self._head = None
|
||
self._head_loaded = False
|
||
self._thresholds = None
|
||
|
||
def _thresholds_cached(self) -> Dict[str, Any]:
|
||
if self._thresholds is None:
|
||
self._thresholds = load_active(self.store, self.cfg.models_dir)
|
||
return self._thresholds
|
||
|
||
async def decide(self, query_or_messages, consumer: str,
|
||
domain: str = "", request_id: str = "",
|
||
executed_tier: str = "") -> TierDecision:
|
||
"""分级决策(§8 时序;全模式写观察)。"""
|
||
ts = time.time()
|
||
rid = request_id or ("rt" + uuid.uuid4().hex[:10])
|
||
feats = gate(query_or_messages, consumer, self.cfg)
|
||
probs: Dict[str, float] = {}
|
||
head_version = ""
|
||
fallback = False
|
||
|
||
vec: Optional[List[int]] = None
|
||
try:
|
||
vec = await embed(feats and (query_or_messages
|
||
if isinstance(query_or_messages, str)
|
||
else "\n".join(
|
||
str(m.get("content") or "")
|
||
for m in query_or_messages)),
|
||
self.cfg.embedder)
|
||
except EmbedderDown:
|
||
fallback = True
|
||
|
||
head = None if fallback else self._load_head()
|
||
if head is None:
|
||
fallback = True
|
||
|
||
if not fallback and vec is not None:
|
||
probs = head.predict([float(v) for v in vec])
|
||
head_version = head.version
|
||
|
||
th = self._thresholds_cached()
|
||
th_version = str(th.get("version") or "")
|
||
|
||
# ---- 决策 ----
|
||
if fallback:
|
||
tier = _rule_tier(feats) # D-G4 规则门退化
|
||
else:
|
||
t1_ok = (probs.get("t1", 0.0) >= float(th.get("t1", 0.9))
|
||
and feats.t1_hard_ok)
|
||
t3_ok = (probs.get("t3", 0.0) >= float(th.get("t3", 0.9))
|
||
or feats.repo_signals)
|
||
if t3_ok and not t1_ok:
|
||
tier = "T3"
|
||
elif t1_ok:
|
||
tier = "T1"
|
||
else:
|
||
tier = "T2"
|
||
|
||
# ---- mode 裁剪(D-G7)----
|
||
if self.cfg.mode == "live":
|
||
executed = tier # live:决策即执行
|
||
else:
|
||
# collect/shadow:现行为——规则门等价(T1 门/T3 信号)近似 v2 现状
|
||
executed = executed_tier or _rule_tier(feats)
|
||
|
||
decision = TierDecision(
|
||
tier=tier, probs=probs,
|
||
hard_gates={"turns": feats.turns, "est_tokens": feats.est_tokens,
|
||
"single_turn": feats.single_turn,
|
||
"intent_blocked": feats.intent_blocked,
|
||
"repo_signals": feats.repo_signals,
|
||
"over_length": feats.over_length,
|
||
"t1_hard_ok": feats.t1_hard_ok},
|
||
thresholds_version=th_version, head_version=head_version,
|
||
mode=self.cfg.mode, fallback=fallback,
|
||
executed_tier=executed)
|
||
|
||
# ---- 观察必写(全模式)----
|
||
if self.observer is not None:
|
||
self.observer.log(__import__("gateway.sense.observer",
|
||
fromlist=["Observation"]).Observation(
|
||
request_id=rid, consumer=consumer, decided_tier=tier,
|
||
executed_tier=executed, probs=probs,
|
||
policy_version=f"head:{head_version}|th:{th_version}",
|
||
features=decision.hard_gates, embedding=vec,
|
||
bucket="default", domain=domain, ts=ts))
|
||
return decision
|