阶段4(上):LangGraph、Function Calling、MCP(SDK 2.x)、Agent 评测与运维、Agent 设计专题;run_examples 设置 RUN_EXAMPLES=1
This commit is contained in:
1 parent
6650f7eb6d
commit
caaec830b3
38 files changed
+3515
-64
No files matched your search
@@ -0,0 +1,146 @@
|
||||
# 只用标准库
|
||||
"""三个 Agent 设计模式的最小实现(离线可运行):
|
||||
|
||||
1. Plan-and-Execute + 失败重新规划
|
||||
2. Skills 渐进式加载:启动时只把技能的名字和描述放进上下文,需要时再加载全文
|
||||
3. 上下文预算:超过预算时把早期对话压缩成摘要,保留系统提示词和最近几轮
|
||||
|
||||
模型部分用规则模拟,接口形状和真实做法一致;真实项目中 plan / summarize 由大模型完成。
|
||||
运行:python patterns.py
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import re
|
||||
from dataclasses import dataclass, field
|
||||
from pathlib import Path
|
||||
|
||||
# ---------------- 1. Plan-and-Execute ----------------
|
||||
TOOLS = {
|
||||
"query_sales": lambda a: {"rows": [("华北", 120), ("华东", 200)]} if a.get("month") else {"error": "缺少 month 参数"},
|
||||
"make_chart": lambda a: {"chart": f"bar({len(a['rows'])} bars)"},
|
||||
"send_report": lambda a: {"sent_to": a["to"]},
|
||||
}
|
||||
|
||||
|
||||
@dataclass
|
||||
class Step:
|
||||
tool: str
|
||||
args: dict
|
||||
status: str = "pending" # pending / done / failed
|
||||
result: dict | None = None
|
||||
|
||||
|
||||
def plan(goal: str, feedback: str | None = None) -> list[Step]:
|
||||
"""规划器(真实项目中由大模型输出结构化计划)。收到失败反馈时修正计划。"""
|
||||
month = re.search(r"(\d{4}-\d{2})", goal)
|
||||
args = {"month": month.group(1)} if (month and feedback) else {}
|
||||
return [Step("query_sales", args), Step("make_chart", {}), Step("send_report", {"to": "ops@example.com"})]
|
||||
|
||||
|
||||
def plan_and_execute(goal: str, max_replans: int = 2) -> tuple[list[Step], list[str]]:
|
||||
log, feedback = [], None
|
||||
for attempt in range(max_replans + 1):
|
||||
steps = plan(goal, feedback)
|
||||
log.append(f"第 {attempt + 1} 次规划:{[s.tool for s in steps]}")
|
||||
ctx: dict = {}
|
||||
for s in steps:
|
||||
args = {**s.args, **({"rows": ctx["rows"]} if s.tool == "make_chart" else {})}
|
||||
s.result = TOOLS[s.tool](args)
|
||||
if "error" in s.result:
|
||||
s.status = "failed"
|
||||
feedback = f"{s.tool} 失败:{s.result['error']}"
|
||||
log.append(feedback)
|
||||
break
|
||||
s.status = "done"
|
||||
ctx.update(s.result)
|
||||
else:
|
||||
return steps, log
|
||||
return steps, log + ["超过最大重新规划次数,转人工"]
|
||||
|
||||
|
||||
# ---------------- 2. Skills 渐进式加载 ----------------
|
||||
@dataclass
|
||||
class Skill:
|
||||
name: str
|
||||
description: str
|
||||
path: Path
|
||||
|
||||
def body(self) -> str:
|
||||
text = self.path.read_text(encoding="utf-8")
|
||||
return text.split("---", 2)[2].strip() # 去掉 frontmatter
|
||||
|
||||
|
||||
def discover_skills(root: Path) -> list[Skill]:
|
||||
skills = []
|
||||
for f in sorted(root.glob("*/SKILL.md")):
|
||||
meta = dict(re.findall(r"^(name|description):\s*(.+)$", f.read_text(encoding="utf-8"), re.M))
|
||||
skills.append(Skill(meta["name"], meta["description"], f))
|
||||
return skills
|
||||
|
||||
|
||||
def skills_index(skills: list[Skill]) -> str:
|
||||
"""放进系统提示词的只有这一小段索引。"""
|
||||
return "\n".join(f"- {s.name}: {s.description}" for s in skills)
|
||||
|
||||
|
||||
def pick_skill(question: str, skills: list[Skill]) -> Skill | None:
|
||||
"""真实项目中由模型根据索引决定加载哪个技能(或通过一个 load_skill 工具)。"""
|
||||
keywords = {"refund-policy": ["退款", "退货"], "sql-report": ["统计", "报表"]}
|
||||
return next((s for s in skills if any(k in question for k in keywords.get(s.name, []))), None)
|
||||
|
||||
|
||||
# ---------------- 3. 上下文预算 ----------------
|
||||
def n_tokens(text: str) -> int:
|
||||
return len(text) # 粗略估算:中文约 1 字 1 token;真实项目用模型的分词器
|
||||
|
||||
|
||||
@dataclass
|
||||
class Context:
|
||||
system: str
|
||||
budget: int
|
||||
keep_recent: int = 4
|
||||
messages: list[tuple[str, str]] = field(default_factory=list)
|
||||
summary: str = ""
|
||||
|
||||
def add(self, role: str, text: str) -> None:
|
||||
self.messages.append((role, text))
|
||||
while self.size() > self.budget and len(self.messages) > self.keep_recent:
|
||||
old = self.messages[: len(self.messages) - self.keep_recent]
|
||||
self.messages = self.messages[len(self.messages) - self.keep_recent:]
|
||||
self.summary = summarize(self.summary, old)
|
||||
|
||||
def size(self) -> int:
|
||||
return n_tokens(self.system) + n_tokens(self.summary) + sum(n_tokens(t) for _, t in self.messages)
|
||||
|
||||
def render(self) -> list[tuple[str, str]]:
|
||||
head = [("system", self.system)] + ([("system", "之前对话摘要:" + self.summary)] if self.summary else [])
|
||||
return head + self.messages
|
||||
|
||||
|
||||
def summarize(prev: str, msgs: list[tuple[str, str]]) -> str:
|
||||
"""真实项目中由模型生成摘要,要求保留:用户目标、已确认的事实、未完成事项。"""
|
||||
facts = [t for r, t in msgs if r == "user" and re.search(r"\d|地址|改", t)]
|
||||
if not facts:
|
||||
return prev
|
||||
return (prev + ";" if prev else "") + "用户提供的信息:" + ";".join(facts)[:60]
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
steps, log = plan_and_execute("生成 2026-09 各大区销售报表并发给运营")
|
||||
print("\n".join(log))
|
||||
print("最终步骤状态:", [(s.tool, s.status) for s in steps])
|
||||
assert [s.status for s in steps] == ["done", "done", "done"] and len(log) == 3
|
||||
|
||||
skills = discover_skills(Path(__file__).parent / "skills")
|
||||
print("系统提示词中的技能索引:\n" + skills_index(skills))
|
||||
s = pick_skill("我上周买的耳机想退款,800 元", skills)
|
||||
print("按需加载技能:", s.name, "->", s.body().splitlines()[0])
|
||||
assert s.name == "refund-policy" and "500" in s.body()
|
||||
|
||||
ctx = Context(system="你是售后助手。", budget=40)
|
||||
for i, t in enumerate(["我的订单号是 A1001", "好的", "收货地址改到上海", "嗯",
|
||||
"什么时候发货", "预计明天", "能开发票吗", "可以"]):
|
||||
ctx.add("user" if i % 2 == 0 else "assistant", t)
|
||||
print("压缩后上下文:", ctx.render())
|
||||
assert ctx.size() <= 40 or len(ctx.messages) == ctx.keep_recent
|
||||
assert "A1001" in ctx.summary
|
||||
@@ -0,0 +1,10 @@
|
||||
---
|
||||
name: refund-policy
|
||||
description: 处理退款、退货、换货相关的问题时使用。包含退款规则、审批阈值和回复模板。
|
||||
---
|
||||
|
||||
# 退款处理规则
|
||||
|
||||
1. 签收 7 天内无理由退货;超过 7 天只处理质量问题。
|
||||
2. 退款金额 >= 500 元需要主管审批,先告知用户“已提交审批”。
|
||||
3. 不承诺具体到账时间,只说明“原路退回,以支付渠道为准”。
|
||||
@@ -0,0 +1,9 @@
|
||||
---
|
||||
name: sql-report
|
||||
description: 用户要求统计数据、生成报表或查询数据库时使用。包含可用的表、只读查询规范。
|
||||
---
|
||||
|
||||
# 报表查询规范
|
||||
|
||||
- 只允许 SELECT;必须带时间范围条件;单次最多返回 1000 行。
|
||||
- 订单表 orders(id, user_id, amount, status, created_at)。
|
||||
@@ -0,0 +1,29 @@
|
||||
# requires: pytest
|
||||
from pathlib import Path
|
||||
|
||||
import patterns as p
|
||||
|
||||
|
||||
def test_replan_after_failure():
|
||||
steps, log = p.plan_and_execute("生成 2026-09 报表")
|
||||
assert "query_sales 失败" in log[1] and all(s.status == "done" for s in steps)
|
||||
|
||||
|
||||
def test_give_up_after_max_replans():
|
||||
steps, log = p.plan_and_execute("生成报表(没有月份)", max_replans=1)
|
||||
assert log[-1].startswith("超过最大重新规划次数")
|
||||
|
||||
|
||||
def test_skill_index_is_small_and_body_loaded_on_demand():
|
||||
skills = p.discover_skills(Path(p.__file__).parent / "skills")
|
||||
idx = p.skills_index(skills)
|
||||
assert "500" not in idx and len(skills) == 2 # 索引里没有技能全文
|
||||
assert p.pick_skill("帮我统计上月订单", skills).name == "sql-report"
|
||||
assert p.pick_skill("今天天气", skills) is None
|
||||
|
||||
|
||||
def test_context_keeps_recent_and_summary():
|
||||
ctx = p.Context(system="s", budget=20, keep_recent=2)
|
||||
for t in ["订单 A1", "好", "地址 上海 1 号", "嗯", "发票", "可以"]:
|
||||
ctx.add("user", t)
|
||||
assert len(ctx.messages) == 2 and "A1" in ctx.summary
|
||||
Reference in new issue
Block a user