tzt/projectAIpopular
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(v2): 架构与算法优化二轮——推理机不变量外提、知识库匹配预编译、协作循环增量索引

Browse Source 
- inference.py（/chat/legacy 热路径）：kb.match 循环不变量外提（原每步全量重扫+重排序，
  最坏 O(steps×rules×patterns)）；fired 查重 list→set
- knowledge.py：Rule patterns 注册侧懒缓存小写副本（原每条规则每次匹配重复 lower）；
  match() 文本只 lower 一次（原逐规则重复）；load() 的 yaml 文件名集合提到循环外
- worker.py：本地端点生成器 httpx.AsyncClient 懒建复用（原每步新建/销毁连接，
  对齐 ArchitectClient 惯用法；协作循环最多 10 次生成免重复建连）
- pipeline.py（协作循环）：plan_by_id O(1) 步定义查找；done 集合增量维护
  （原每轮重建 progress+archive 扫描）；领域只解析一次（原 _artifact_name 每步
  全领域 kb.match）；_deps_done 支持传入预填集合（保持旧签名兼容）
- v2stats.py：回合数分布改增量聚合（sum/max/分桶计数），summary() O(n)→O(1)，
  不再持有无界 list（修长时运行内存增长）
- gateway/agent.py + api.py：AgentService 运行计数 O(1) 化（原 register 全量扫描），
  状态迁移收敛到 _transition_state 单一入口（api.py cancel/异常两处绕过点一并接入，
  消除计数与状态脱节隐患）；21 项 agent 测试全绿（两轮全量 230 passed 复核）
This commit is contained in:
tzt
2026-09-18 23:45:35 +08:00
parent 0767030f02
commit 8820e8da48
7 changed files with 103 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files
router_system/pipeline.py
+21 -10
View File
@@ -110,20 +110,24 @@ class CollaborativePipeline:
plan = brief.get("plan") or []
pending = [p.get("id") for p in plan]
plan_ids = {p.get("id") for p in plan} # 用于 _deps_done 过滤
plan_by_id = {p.get("id"): p for p in plan} # O(1) 步定义查找（原每轮线性扫描）
done_ids = self._done_ids(ws) # 增量维护的完成集（原每轮重建）
base_domain = self._guess_domain(query) # ws["query"] 恒定，领域只解析一次
while pending and not ws.exhausted():
progressed = False
for sid in list(pending):
step = next((p for p in plan if p.get("id") == sid), {})
step = plan_by_id.get(sid, {})
deps = step.get("deps") or []
# 只检查在 plan 中的依赖；不在 plan 的 ID 视为"不存在"→自动满足
relevant = [d for d in deps if d in plan_ids]
if not self._deps_done(ws, relevant):
if not all(d in done_ids for d in relevant):
continue
existing = self._read_artifact(request_id, self._artifact_name(sid, ws))
existing = self._read_artifact(request_id, self._artifact_name(sid, ws, domain=base_domain))
outcome = await self.worker.run_step(ws, sid, existing_artifact=existing,
hint=self._last_decision_for(ws, sid))
if outcome.status == "done":
pending.remove(sid)
done_ids.add(sid)
self._save_artifact(request_id, outcome.artifact_name, outcome.artifact_text)
ws.rollup()
route.append(f"step:{sid}:done")
@@ -230,9 +234,10 @@ class CollaborativePipeline:
return hits[0].domain
return "general"
def _artifact_name(self, sid: str, ws: Workspace) -> str:
def _artifact_name(self, sid: str, ws: Workspace, domain: Optional[str] = None) -> str:
from .worker import artifact_name_for
domain = self._guess_domain(ws["query"])
if domain is None:
domain = self._guess_domain(ws["query"])
# 用 brief.tags 优先
tags = (ws.get("brief") or {}).get("tags") or []
for t in tags:
@@ -241,8 +246,10 @@ class CollaborativePipeline:
break
return artifact_name_for(sid, domain)
def _deps_done(self, ws: Workspace, deps: List[str]) -> bool:
# progress 中 done 的条目；done 条目 rollup 后移到 archive（字符串格式如 "s1: ..."）
@staticmethod
def _done_ids(ws: Workspace) -> set:
"""收集已完成 step id 集合：progress 中 done 的条目 + archive 折叠行
（done 条目 rollup 后移到 archive，字符串格式如 "s1: ..."）。"""
done = {p["step"] for p in ws.get("progress", []) if p.get("status") == "done"}
for entry in ws.get("archive", []):
if isinstance(entry, dict):
@@ -251,9 +258,13 @@ class CollaborativePipeline:
sid = entry.split(":")[0].strip() if ":" in entry else ""
else:
sid = ""
if sid in deps:
done.add(sid)
return all(d in done for d in deps)
done.add(sid)
return done
def _deps_done(self, ws: Workspace, deps: List[str], done_ids: Optional[set] = None) -> bool:
if done_ids is None:
done_ids = self._done_ids(ws)
return all(d in done_ids for d in deps)
def _last_decision_for(self, ws: Workspace, sid: str) -> str:
"""取最近一条针对该 step 的决策 reply，作为 worker hint。"""