Files
aiagent/backend/scripts/create_zhini_kefu_16.py

553 lines
18 KiB
Python
Raw Normal View History

#!/usr/bin/env python3
"""
知你客服15号复制为知你客服16号 **B 能力**画布 **Loop + 循环体内 LLM**
- **主线**Cache 等上游 `code-build` 注入 `loop_rounds` 数组 `loop` 体内`llm-subtask` `loop_end`
循环结束后 `code-merge` 将各段 LLM 输出折叠为 `right`/`reply`并合并 `user_profile` `llm-unified` 之后的节点 Cache End
- **弃用节点** `llm-unified` 从主链摘除并保留在画布上无连线避免与多段执行重复真正推理在 `zhini16-llm-subtask`
- **工具** 15 相同每段内 `max_tool_iterations` 可单独配置默认 12
- **轮数**环境变量 `ZHINI16_LOOP_ROUNDS`默认 3上限 8
引擎要求`loop` **第一条出边**必须指向循环体起点脚本会对 `zhini16-loop-main` 的出边排序保证
用法:
cd backend && .\\venv\\Scripts\\python.exe scripts/create_zhini_kefu_16.py
环境变量: PLATFORM_BASE_URL, PLATFORM_USERNAME, PLATFORM_PASSWORD,
SOURCE_AGENT_NAME默认 知你客服15号, TARGET_NAME默认 知你客服16号
ZHINI16_LOOP_ROUNDS, ZHINI16_SUBTASK_MAX_TOOL_ITERATIONS
"""
from __future__ import annotations
import copy
import json
import os
import sys
from collections import defaultdict
from typing import Any, Dict, List, Optional, Tuple
import requests
BASE = os.getenv("PLATFORM_BASE_URL", "http://127.0.0.1:8037").rstrip("/")
USER = os.getenv("PLATFORM_USERNAME", "admin")
PWD = os.getenv("PLATFORM_PASSWORD", "123456")
SOURCE_NAME = os.getenv("SOURCE_AGENT_NAME", "知你客服15号")
TARGET_NAME = os.getenv("TARGET_NAME", "知你客服16号")
LOOP_ROUNDS = max(1, min(int(os.getenv("ZHINI16_LOOP_ROUNDS", "3")), 8))
SUBTASK_MAX_TOOL_ITER = max(1, min(int(os.getenv("ZHINI16_SUBTASK_MAX_TOOL_ITERATIONS", "12")), 64))
TOOLS_V16: List[str] = [
"http_request",
"file_read",
"file_write",
"text_analyze",
"datetime",
"math_calculate",
"system_info",
"json_process",
"database_query",
"adb_log",
]
LLM_UNIFIED = "llm-unified"
N_CODE_BUILD = "zhini16-code-build-rounds"
N_LOOP = "zhini16-loop-main"
N_SUB_LLM = "zhini16-llm-subtask"
N_LOOP_END = "zhini16-loop-end"
N_CODE_MERGE = "zhini16-code-merge-rounds"
PROMPT_V16_MARKER = "【知你客服 16 号 · Loop 多段执行】"
PROMPT_V16_SUBTASK_EXTRA = f"""
{PROMPT_V16_MARKER}
执行方式当前用户请求会在**同一次 API 调用**内被拆成 **{{{{round_total}}}}** 段顺序执行你收到的是其中** {{{{round_index}}}} ** 0 起计轮次标量 `round`={{{{round}}}}每段都调用同一套工具与纪律但请**聚焦本段可推进的子目标**避免与前后段完全重复若本段仅需承接上文可简短小结并推进下一步
输出每段末尾仍输出**一行合法 JSON** `intent``reply``user_profile` 14/15 一致`reply` **本段**面向用户的可见说明最终给用户展示时会与各段 `reply` 合并请勿在段内假设用户已看到其他段的正文
纪律继承 15 多步工具链`database_query` SELECT勿编造工具返回勿刷屏 DSML
"""
CODE_MERGE_PYTHON = r"""
import json as _json
def _parse_tail_json_obj(s):
if not isinstance(s, str):
return None
t = s.strip()
if not t:
return None
last_nl = t.rfind("\n")
last_line = t[last_nl + 1 :].strip() if last_nl >= 0 else t
if not last_line.startswith("{"):
return None
try:
o = _json.loads(last_line)
return o if isinstance(o, dict) else None
except Exception:
return None
def _reply_from_segment(s):
if not isinstance(s, str):
return str(s)
o = _parse_tail_json_obj(s)
if o and isinstance(o.get("reply"), str) and o["reply"].strip():
return o["reply"].strip()
return s.strip()
parts = []
if isinstance(input_data, dict):
if isinstance(input_data.get("input"), list):
parts = input_data.get("input")
elif isinstance(input_data.get("right"), list):
parts = input_data.get("right")
elif isinstance(input_data, list):
parts = input_data
chunks = []
merged_profile = {}
for i, p in enumerate(parts):
if p is None:
continue
chunks.append("【第%d段】\n%s" % (i + 1, _reply_from_segment(p)))
o = _parse_tail_json_obj(p) if isinstance(p, str) else None
if o and isinstance(o.get("user_profile"), dict):
merged_profile.update(o["user_profile"])
merged_text = "\n\n".join(chunks) if chunks else "(循环未产生有效输出)"
out = dict(input_data) if isinstance(input_data, dict) else {}
out.pop("input", None)
out["reply"] = merged_text
out["right"] = {"right": merged_text}
if merged_profile:
out["user_profile_update"] = merged_profile
result = out
""".strip()
def _code_build_source(rounds: List[int]) -> str:
return (
"out = dict(input_data) if isinstance(input_data, dict) else {}\n"
f"out['loop_rounds'] = {rounds!r}\n"
"result = out\n"
)
def _sanitize_edges(edges: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
seen: set = set()
out: List[Dict[str, Any]] = []
for e in edges or []:
s, t = e.get("source"), e.get("target")
if not s or not t:
continue
if s == t:
continue
key = (s, t)
if key in seen:
continue
seen.add(key)
ne = dict(e)
ne["sourceHandle"] = "right"
ne["targetHandle"] = "left"
if not ne.get("id"):
ne["id"] = f"edge_{s}_{t}"
out.append(ne)
return out
def _find_start_node_ids(nodes: List[Dict[str, Any]]) -> List[str]:
ids: List[str] = []
for n in nodes or []:
nid = n.get("id") or ""
nt = (n.get("type") or (n.get("data") or {}).get("type") or "").lower()
if nt == "start" or nid in ("start", "start-1") or str(nid).startswith("start-"):
ids.append(nid)
return ids
def _compute_ranks(
nodes: List[Dict[str, Any]], edges: List[Dict[str, Any]]
) -> Dict[str, int]:
node_ids = [n["id"] for n in nodes if n.get("id")]
start_ids = _find_start_node_ids(nodes)
incoming: Dict[str, int] = {nid: 0 for nid in node_ids}
for e in edges:
s, t = e.get("source"), e.get("target")
if not s or not t or s == t:
continue
if t in incoming:
incoming[t] += 1
if not start_ids:
start_ids = [nid for nid in node_ids if incoming.get(nid, 0) == 0] or ([node_ids[0]] if node_ids else [])
rank: Dict[str, int] = {s: 0 for s in start_ids}
nmax = max(len(nodes), 8)
for _ in range(nmax + 5):
updated = False
for e in edges:
s, t = e.get("source"), e.get("target")
if not s or not t or s == t:
continue
if s not in rank:
continue
nv = rank[s] + 1
if t not in rank or rank[t] < nv:
rank[t] = nv
updated = True
if not updated:
break
max_r = max(rank.values(), default=0)
for nid in node_ids:
if nid not in rank:
rank[nid] = max_r + 1
max_r += 1
return rank
def _apply_layered_positions(nodes: List[Dict[str, Any]], ranks: Dict[str, int]) -> None:
layers: Dict[int, List[str]] = defaultdict(list)
for nid, r in ranks.items():
layers[r].append(nid)
for r in layers:
layers[r].sort()
x0, y0 = 80.0, 140.0
x_step = 300.0
y_step = 110.0
for r in sorted(layers.keys()):
ids = layers[r]
nlen = len(ids)
y_base = y0 - (nlen - 1) * y_step / 2.0
for j, nid in enumerate(ids):
for node in nodes:
if node.get("id") != nid:
continue
pos = node.setdefault("position", {})
pos["x"] = x0 + r * x_step
pos["y"] = y_base + j * y_step
break
def improve_workflow_layout_and_edges(wf: Dict[str, Any]) -> Tuple[int, int]:
nodes = wf.get("nodes") or []
raw_edges = wf.get("edges") or []
loops = sum(
1
for e in raw_edges
if e.get("source") and e.get("target") and e.get("source") == e.get("target")
)
clean = _sanitize_edges(raw_edges)
removed_dup = len(raw_edges) - len(clean) - loops
wf["edges"] = clean
ranks = _compute_ranks(nodes, clean)
_apply_layered_positions(nodes, ranks)
return loops, max(0, removed_dup)
def _sort_loop_out_edges(edges: List[Dict[str, Any]], loop_id: str, body_target: str) -> None:
loop_es = [e for e in edges if e.get("source") == loop_id]
non = [e for e in edges if e.get("source") != loop_id]
body = [e for e in loop_es if e.get("target") == body_target]
rest = [e for e in loop_es if e.get("target") != body_target]
edges[:] = non + body + rest
def _upsert_node(nodes: List[Dict[str, Any]], node: Dict[str, Any]) -> None:
nid = node.get("id")
for i, n in enumerate(nodes):
if n.get("id") == nid:
nodes[i] = node
return
nodes.append(node)
def _topology_already_applied(nodes: List[Dict[str, Any]]) -> bool:
return any(n.get("id") == N_LOOP for n in (nodes or []))
def _apply_loop_b_topology(wf: Dict[str, Any], rounds: List[int]) -> None:
nodes = wf.setdefault("nodes", [])
edges = wf.setdefault("edges", [])
pres: List[str] = []
posts: List[str] = []
seen_p: set = set()
seen_t: set = set()
for e in edges:
if e.get("target") == LLM_UNIFIED:
s = e.get("source")
if s and s not in seen_p:
seen_p.add(s)
pres.append(s)
if e.get("source") == LLM_UNIFIED:
t = e.get("target")
if t and t not in seen_t:
seen_t.add(t)
posts.append(t)
if not pres:
print("错误: 未找到指向 llm-unified 的边,无法从 15 号主线改造。", file=sys.stderr)
raise SystemExit(2)
if not posts:
print("错误: 未找到从 llm-unified 出发的边。", file=sys.stderr)
raise SystemExit(2)
edges[:] = [e for e in edges if e.get("target") != LLM_UNIFIED and e.get("source") != LLM_UNIFIED]
llm_template: Optional[Dict[str, Any]] = None
for n in nodes:
if n.get("id") == LLM_UNIFIED:
llm_template = copy.deepcopy(n)
break
sub_llm: Dict[str, Any] = {
"id": N_SUB_LLM,
"type": "llm",
"position": {"x": 0, "y": 0},
"data": {
"label": "16号·循环体内LLM",
"prompt": "你是知你客服多段执行中的子任务模型。\n" + PROMPT_V16_SUBTASK_EXTRA,
"enable_tools": True,
"tools": list(TOOLS_V16),
"selected_tools": list(TOOLS_V16),
"max_tool_iterations": SUBTASK_MAX_TOOL_ITER,
},
}
if llm_template:
td = llm_template.get("data") or {}
sd = sub_llm["data"]
for k in ("provider", "model", "temperature", "max_tokens"):
if k in td and td[k] is not None:
sd[k] = td[k]
base_p = (td.get("prompt") or "").strip()
if base_p and PROMPT_V16_MARKER not in base_p:
sd["prompt"] = base_p + "\n" + PROMPT_V16_SUBTASK_EXTRA.strip()
_upsert_node(
nodes,
{
"id": N_CODE_BUILD,
"type": "code",
"position": {"x": 0, "y": 0},
"data": {
"label": "16号·注入loop_rounds",
"language": "python",
"code": _code_build_source(rounds),
},
},
)
_upsert_node(
nodes,
{
"id": N_LOOP,
"type": "loop",
"position": {"x": 0, "y": 0},
"data": {
"label": "16号·多段循环",
"items_path": "right.loop_rounds",
"item_variable": "round",
"error_handling": "continue",
},
},
)
_upsert_node(nodes, sub_llm)
_upsert_node(
nodes,
{
"id": N_LOOP_END,
"type": "loop_end",
"position": {"x": 0, "y": 0},
"data": {"label": "16号·循环结束"},
},
)
_upsert_node(
nodes,
{
"id": N_CODE_MERGE,
"type": "code",
"position": {"x": 0, "y": 0},
"data": {
"label": "16号·合并各段输出",
"language": "python",
"code": CODE_MERGE_PYTHON,
},
},
)
new_edges: List[Dict[str, Any]] = []
for p in pres:
new_edges.append(
{
"id": f"e_{p}_{N_CODE_BUILD}",
"source": p,
"target": N_CODE_BUILD,
"sourceHandle": "right",
"targetHandle": "left",
}
)
new_edges.append(
{
"id": f"e_{N_CODE_BUILD}_{N_LOOP}",
"source": N_CODE_BUILD,
"target": N_LOOP,
"sourceHandle": "right",
"targetHandle": "left",
}
)
new_edges.append(
{
"id": f"e_{N_LOOP}_{N_SUB_LLM}",
"source": N_LOOP,
"target": N_SUB_LLM,
"sourceHandle": "right",
"targetHandle": "left",
}
)
new_edges.append(
{
"id": f"e_{N_SUB_LLM}_{N_LOOP_END}",
"source": N_SUB_LLM,
"target": N_LOOP_END,
"sourceHandle": "right",
"targetHandle": "left",
}
)
new_edges.append(
{
"id": f"e_{N_LOOP}_{N_CODE_MERGE}",
"source": N_LOOP,
"target": N_CODE_MERGE,
"sourceHandle": "right",
"targetHandle": "left",
}
)
for p in posts:
new_edges.append(
{
"id": f"e_{N_CODE_MERGE}_{p}",
"source": N_CODE_MERGE,
"target": p,
"sourceHandle": "right",
"targetHandle": "left",
}
)
edges.extend(new_edges)
_sort_loop_out_edges(edges, N_LOOP, N_SUB_LLM)
def _patch_zhini16_content_only(wf: Dict[str, Any], rounds: List[int]) -> None:
"""已存在 16 号拓扑时,仅刷新 code 与 sub-llm 参数。"""
for n in wf.get("nodes") or []:
nid = n.get("id")
if nid == N_CODE_BUILD:
n.setdefault("data", {})["code"] = _code_build_source(rounds)
elif nid == N_SUB_LLM:
d = n.setdefault("data", {})
d["tools"] = list(TOOLS_V16)
d["selected_tools"] = list(TOOLS_V16)
d["max_tool_iterations"] = SUBTASK_MAX_TOOL_ITER
elif nid == N_LOOP:
d = n.setdefault("data", {})
d["items_path"] = "right.loop_rounds"
d["item_variable"] = "round"
d["error_handling"] = "continue"
elif nid == N_CODE_MERGE:
n.setdefault("data", {})["code"] = CODE_MERGE_PYTHON
def _find_agent_id_by_name(h: Dict[str, str], name: str) -> Optional[str]:
r = requests.get(f"{BASE}/api/v1/agents", params={"search": name, "limit": 50}, headers=h, timeout=30)
if r.status_code != 200:
return None
for a in r.json() or []:
if a.get("name") == name:
return a.get("id")
return None
def main() -> int:
rounds = list(range(1, LOOP_ROUNDS + 1))
r = requests.post(
f"{BASE}/api/v1/auth/login",
data={"username": USER, "password": PWD},
headers={"Content-Type": "application/x-www-form-urlencoded"},
timeout=15,
)
if r.status_code != 200:
print("登录失败:", r.status_code, r.text[:500], file=sys.stderr)
return 1
token = r.json().get("access_token")
if not token:
print("无 access_token", file=sys.stderr)
return 1
h = {"Authorization": f"Bearer {token}", "Content-Type": "application/json"}
src_id = _find_agent_id_by_name(h, SOURCE_NAME)
if not src_id:
print(f"未找到源 Agent: {SOURCE_NAME}", file=sys.stderr)
return 1
existing = _find_agent_id_by_name(h, TARGET_NAME)
if existing:
print("已存在", TARGET_NAME, "-> 仅更新工作流", existing)
new_id = existing
g = requests.get(f"{BASE}/api/v1/agents/{new_id}", headers=h, timeout=30)
if g.status_code != 200:
print("读取失败:", g.text, file=sys.stderr)
return 1
agent = g.json()
else:
dup = requests.post(
f"{BASE}/api/v1/agents/{src_id}/duplicate",
headers=h,
json={"name": TARGET_NAME},
timeout=60,
)
if dup.status_code != 201:
print("复制失败:", dup.status_code, dup.text[:800], file=sys.stderr)
return 1
new_id = dup.json()["id"]
agent = dup.json()
print("已创建副本:", new_id, TARGET_NAME)
wf = copy.deepcopy(agent["workflow_config"])
if _topology_already_applied(wf.get("nodes") or []):
_patch_zhini16_content_only(wf, rounds)
else:
_apply_loop_b_topology(wf, rounds)
loops, dup_edges = improve_workflow_layout_and_edges(wf)
print(f"连线整理: 去掉自环 {loops} 条, 合并重复边 {dup_edges}")
desc = (
"知你客服16号B能力——Loop+循环体内LLM"
f"默认 {LOOP_ROUNDS} 段顺序执行,每段节点 zhini16-llm-subtask 工具迭代上限 {SUBTASK_MAX_TOOL_ITER}"
"原 llm-unified 已从主链摘除;各段 reply 经 zhini16-code-merge-rounds 合并后写入下游 Cache/End。"
)
up = requests.put(
f"{BASE}/api/v1/agents/{new_id}",
headers=h,
json={"description": desc, "workflow_config": wf},
timeout=120,
)
if up.status_code != 200:
print("更新失败:", up.status_code, up.text[:1200], file=sys.stderr)
return 1
print("loop_rounds:", rounds)
print("工具:", ", ".join(TOOLS_V16))
print("Agent ID:", new_id)
print(json.dumps({"id": new_id, "name": TARGET_NAME}, ensure_ascii=False))
return 0
if __name__ == "__main__":
raise SystemExit(main())