mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 04:11:58 +00:00
ThearchyHelios
7db9760121
验收：评测汇总里助手用量的模型是 "unknown"，写作用量的模型却是对的。写作用量来自运行的 llm.end（节点配置里 就有模型名）；助手用量来自 providers/usage.Meter，它只从回调的 invocation_params 取 model / model_name，而演示 模型 MockChatModel 没有 _identifying_params，调用参数里只有 _type 和 stop，于是落到 UNKNOWN_MODEL。 改法两处： - Meter：调用参数里取不到时，看 LangChain 照模型对象给回调的 metadata 里的 ls_model_name（调用时传的 model， 否则对象上的 model / model_name），再取不到才看回复、才记 unknown。不认得调用参数的模型类也记得对。 - MockChatModel 补上 _identifying_params（model_name），和 OpenAI 兼容、Anthropic 一样在调用参数里报模型名。 测试（test_usage_meter）：调用参数里不报模型名的模型、演示模型都按对象上的名字记；三种接入（OpenAI 兼容、 Anthropic、演示模型）由 factory 建出来，照 LangChain 回调时给的参数喂给计量，记的都是接入里配的模型 id， 调用参数里本身也带着；评测作答（eval/answer.Driver）走演示模型时助手用量记在 mock-fast 名下。两处修复都 去掉回跑挂 3 条，只留计量那处时挂「调用参数里本身带着」那一条。
276 lines
12 KiB
Python
276 lines
12 KiB
Python
from __future__ import annotations
|
||
|
||
import asyncio
|
||
import hashlib
|
||
import json
|
||
import random
|
||
import re
|
||
from typing import Any, AsyncIterator, Iterator, Sequence
|
||
|
||
from langchain_core.callbacks import AsyncCallbackManagerForLLMRun, CallbackManagerForLLMRun
|
||
from langchain_core.language_models.chat_models import BaseChatModel
|
||
from langchain_core.messages import AIMessage, AIMessageChunk, BaseMessage, ToolMessage
|
||
from langchain_core.outputs import ChatGeneration, ChatGenerationChunk, ChatResult
|
||
from langchain_core.tools import BaseTool
|
||
from pydantic import Field
|
||
|
||
|
||
def _text_of(message: BaseMessage) -> str:
|
||
content = message.content
|
||
if isinstance(content, str):
|
||
return content
|
||
if isinstance(content, list):
|
||
return "".join(
|
||
b.get("text", "") if isinstance(b, dict) else str(b) for b in content
|
||
)
|
||
return str(content)
|
||
|
||
|
||
def _fake_value(schema: dict[str, Any], key: str = "") -> Any:
|
||
"""按 JSON Schema 造一个形状正确的假值，让下游校验节点也能跑通。"""
|
||
typ = schema.get("type")
|
||
if "enum" in schema:
|
||
return schema["enum"][0]
|
||
if typ == "integer":
|
||
return 42
|
||
if typ == "number":
|
||
return 3.14
|
||
if typ == "boolean":
|
||
return True
|
||
if typ == "array":
|
||
item = schema.get("items") or {"type": "string"}
|
||
return [_fake_value(item, key)]
|
||
if typ == "object" or "properties" in schema:
|
||
props = schema.get("properties") or {}
|
||
return {k: _fake_value(v, k) for k, v in props.items()}
|
||
return f"[mock:{key or 'value'}]"
|
||
|
||
|
||
#: 助手按操作流写图时 system 里一定有这一句（api/copilot.py 的 _STREAM_PROTOCOL / APP_PROTOCOL_PROMPT）
|
||
_OPS_MARK = "输出格式（严格遵守）"
|
||
#: 看起来是在问数据的请求：演示模型据此决定要不要先 look 一次。闲聊、改提示词不看
|
||
_DATA_WORDS = re.compile(r"查|多少|几个|统计|数据|表|库|sql|金额|数量|排名|趋势|汇总|明细|报表|周报|月报|指标|销售|"
|
||
r"订单|人数|人次|收入|占比|同比|环比|top", re.I)
|
||
_QUERY_TOOL = re.compile(r"- db_query__(\S+?)（数据库）")
|
||
|
||
|
||
def _question_of(human: str) -> str:
|
||
"""助手请求里用户的原话：问数据是「问题：…」那一行，画布新建是「请设计一个工作流：…」，改图是最后一段。"""
|
||
for mark in ("问题：", "请设计一个工作流："):
|
||
if mark in human:
|
||
return human.rsplit(mark, 1)[1].split("\n", 1)[0].strip()
|
||
return human.strip().rsplit("\n", 1)[-1].strip()
|
||
|
||
|
||
def _ops(*ops: dict[str, Any]) -> str:
|
||
return "\n".join(json.dumps(op, ensure_ascii=False) for op in ops) + "\n"
|
||
|
||
|
||
def _copilot_script(messages: list[BaseMessage]) -> str | None:
|
||
"""助手（画布、问数据、应用润色）的演示脚本：没配模型接入时也能看到「查看 → 接着写」的完整过程。
|
||
|
||
- 数据类请求、有数据源：先 plan，再 look 一次（只写 kind 和数据源，按需求挑表），停下；
|
||
- 收到查看结果（对话里已经有一条自己的输出）：接着写一张「输入 → 查库的 Agent → 成果」的图；
|
||
- 闲聊：一行 reply，不看；
|
||
- 自查修正、应用润色：只说明，不改。
|
||
"""
|
||
system = next((_text_of(m) for m in messages if m.type == "system"), "")
|
||
if _OPS_MARK not in system:
|
||
return None
|
||
humans = [_text_of(m) for m in messages if m.type == "human"]
|
||
first, last = (humans[0], humans[-1]) if humans else ("", "")
|
||
if "app_add" in system:
|
||
return _ops({"op": "plan", "summary": "演示模型不改应用"},
|
||
{"op": "done", "explanation": "（演示模型）没有修改应用。如需真实的润色，请在「设置 → 模型接入」中添加模型接入"})
|
||
if "上线前自查" in last:
|
||
return _ops({"op": "done", "explanation": "演示模型不做修正"})
|
||
question = _question_of(first)
|
||
sources = _QUERY_TOOL.findall(system)
|
||
if any(m.type == "ai" for m in messages):
|
||
tool = f"db_query__{sources[0]}" if sources else None
|
||
answering = "用户问了下面这个问题" in first
|
||
return _ops(
|
||
{"op": "add_node", "node": {"id": "start", "type": "input", "label": "输入",
|
||
"config": {} if answering else {"fields": [{"name": "question"}]}}},
|
||
{"op": "add_node", "node": {"id": "ask", "type": "agent", "label": "查库回答", "config": {
|
||
"prompt": "{{ input.question }}", "tools": [tool] if tool else [], "assign_to": "answer"}}},
|
||
{"op": "add_edge", "edge": {"source": "start", "target": "ask"}},
|
||
{"op": "add_node", "node": {"id": "out", "type": "output", "label": "成果", "config": {
|
||
"fields": [{"name": "answer", "value": "{{ vars.answer }}"}]}}},
|
||
{"op": "add_edge", "edge": {"source": "ask", "target": "out"}},
|
||
{"op": "done", "explanation": "（演示模型）按查看结果搭了一张查库回答的图", "run": True},
|
||
)
|
||
if sources and '"op":"look"' in system and _DATA_WORDS.search(question):
|
||
return _ops({"op": "plan", "summary": "先看一下相关的表，再搭查库回答的图"},
|
||
{"op": "look", "need": question[:80] or "这个问题要用到的表",
|
||
"targets": [{"kind": "table", "source": sources[0]}]})
|
||
topic = re.sub(r"\s+", " ", question).strip()[:60] or "这个问题"
|
||
return _ops({"op": "reply", "text": f"（演示模型的回复）收到「{topic}」。如需真实的回答，请在「设置 → 模型接入」中添加模型接入"})
|
||
|
||
|
||
class MockChatModel(BaseChatModel):
|
||
"""不联网的假模型。
|
||
|
||
目的不是"能用"，而是让人在没有 API Key 的情况下就能把整条编排链路跑通：
|
||
它会真的产出 tool_call、真的按 schema 返回结构化 JSON、真的流式吐字，
|
||
所以画布上的高亮、工具节点、校验节点全都会被触发。
|
||
"""
|
||
|
||
model_name: str = "mock-fast"
|
||
latency: float = 0.012 # 每个 chunk 的间隔，模拟真实流式手感
|
||
seed: int | None = None
|
||
tools: list[dict[str, Any]] = Field(default_factory=list)
|
||
response_format: dict[str, Any] | None = None
|
||
|
||
@property
|
||
def _llm_type(self) -> str:
|
||
return "mock"
|
||
|
||
@property
|
||
def _identifying_params(self) -> dict[str, Any]:
|
||
"""调用参数里带上模型名（同 OpenAI 兼容、Anthropic 的 model）：用量计量（providers/usage）按它记在哪个模型名下，
|
||
不报的话评测汇总里助手用量的模型就成了 unknown。"""
|
||
return {"model_name": self.model_name}
|
||
|
||
def bind_tools(self, tools: Sequence[Any], **kwargs: Any) -> "MockChatModel":
|
||
normalized: list[dict[str, Any]] = []
|
||
for tool in tools:
|
||
if isinstance(tool, BaseTool):
|
||
normalized.append(
|
||
{
|
||
"name": tool.name,
|
||
"description": tool.description,
|
||
"schema": tool.args_schema.model_json_schema()
|
||
if getattr(tool, "args_schema", None)
|
||
else {},
|
||
}
|
||
)
|
||
elif isinstance(tool, dict):
|
||
fn = tool.get("function", tool)
|
||
normalized.append(
|
||
{
|
||
"name": fn.get("name", "tool"),
|
||
"description": fn.get("description", ""),
|
||
"schema": fn.get("parameters") or fn.get("input_schema") or {},
|
||
}
|
||
)
|
||
return self.model_copy(update={"tools": normalized})
|
||
|
||
def with_structured_output(self, schema: Any, **kwargs: Any) -> Any: # type: ignore[override]
|
||
json_schema = schema
|
||
if hasattr(schema, "model_json_schema"):
|
||
json_schema = schema.model_json_schema()
|
||
bound = self.model_copy(update={"response_format": json_schema})
|
||
if hasattr(schema, "model_validate"):
|
||
return bound | (lambda msg: schema.model_validate(json.loads(_text_of(msg))))
|
||
return bound | (lambda msg: json.loads(_text_of(msg)))
|
||
|
||
# ---------------- 内部 ----------------
|
||
|
||
def _decide(self, messages: list[BaseMessage]) -> AIMessage:
|
||
prompt = "\n".join(_text_of(m) for m in messages[-4:])
|
||
rng = random.Random(
|
||
self.seed
|
||
if self.seed is not None
|
||
else int(hashlib.sha1(prompt.encode()).hexdigest()[:8], 16)
|
||
)
|
||
|
||
# 要结构化输出就按 schema 造数据
|
||
if self.response_format:
|
||
payload = _fake_value(self.response_format)
|
||
return AIMessage(content=json.dumps(payload, ensure_ascii=False, indent=2))
|
||
|
||
# 助手按操作流写图：照脚本走一遍（数据类请求先 look 一次）
|
||
if not self.tools and (script := _copilot_script(messages)) is not None:
|
||
return AIMessage(content=script)
|
||
|
||
# 绑了工具、而且这一轮还没有工具返回过 —— 那就调一次工具，
|
||
# 让 agent 循环和画布上的工具节点真的被走到
|
||
already_called = any(isinstance(m, ToolMessage) for m in messages)
|
||
if self.tools and not already_called:
|
||
tool = rng.choice(self.tools)
|
||
args = _fake_value(tool.get("schema") or {"type": "object"})
|
||
if not isinstance(args, dict):
|
||
args = {}
|
||
return AIMessage(
|
||
content=f"需要先调用 `{tool['name']}` 获取信息。",
|
||
tool_calls=[
|
||
{"name": tool["name"], "args": args, "id": f"call_{rng.randrange(1 << 30):08x}"}
|
||
],
|
||
)
|
||
|
||
last_user = next(
|
||
(_text_of(m) for m in reversed(messages) if m.type == "human"), ""
|
||
)
|
||
topic = re.sub(r"\s+", " ", last_user).strip()[:60] or "这个问题"
|
||
body = (
|
||
f"（演示模型的回复）关于「{topic}」，示例内容分三点：\n\n"
|
||
f"1. 这是演示模型生成的示例内容，用于验证工作流能否正常运行。\n"
|
||
f"2. 当前节点收到 {len(messages)} 条消息、{len(self.tools)} 个工具。\n"
|
||
f"3. 如需真实结果，请在「设置 → 模型接入」中添加模型接入，并为节点换用对应模型。"
|
||
)
|
||
return AIMessage(content=body)
|
||
|
||
def _generate(
|
||
self,
|
||
messages: list[BaseMessage],
|
||
stop: list[str] | None = None,
|
||
run_manager: CallbackManagerForLLMRun | None = None,
|
||
**kwargs: Any,
|
||
) -> ChatResult:
|
||
message = self._decide(messages)
|
||
message.usage_metadata = {
|
||
"input_tokens": sum(len(_text_of(m)) // 4 for m in messages),
|
||
"output_tokens": len(_text_of(message)) // 4,
|
||
"total_tokens": 0,
|
||
}
|
||
return ChatResult(generations=[ChatGeneration(message=message)])
|
||
|
||
async def _agenerate(
|
||
self,
|
||
messages: list[BaseMessage],
|
||
stop: list[str] | None = None,
|
||
run_manager: AsyncCallbackManagerForLLMRun | None = None,
|
||
**kwargs: Any,
|
||
) -> ChatResult:
|
||
await asyncio.sleep(self.latency)
|
||
return self._generate(messages, stop, **kwargs)
|
||
|
||
def _stream(
|
||
self,
|
||
messages: list[BaseMessage],
|
||
stop: list[str] | None = None,
|
||
run_manager: CallbackManagerForLLMRun | None = None,
|
||
**kwargs: Any,
|
||
) -> Iterator[ChatGenerationChunk]:
|
||
message = self._decide(messages)
|
||
if message.tool_calls:
|
||
yield ChatGenerationChunk(
|
||
message=AIMessageChunk(
|
||
content=message.content, tool_calls=message.tool_calls
|
||
)
|
||
)
|
||
return
|
||
for piece in re.findall(r".{1,6}", _text_of(message), re.DOTALL):
|
||
yield ChatGenerationChunk(message=AIMessageChunk(content=piece))
|
||
|
||
async def _astream(
|
||
self,
|
||
messages: list[BaseMessage],
|
||
stop: list[str] | None = None,
|
||
run_manager: AsyncCallbackManagerForLLMRun | None = None,
|
||
**kwargs: Any,
|
||
) -> AsyncIterator[ChatGenerationChunk]:
|
||
message = self._decide(messages)
|
||
if message.tool_calls:
|
||
await asyncio.sleep(self.latency)
|
||
yield ChatGenerationChunk(
|
||
message=AIMessageChunk(
|
||
content=message.content, tool_calls=message.tool_calls
|
||
)
|
||
)
|
||
return
|
||
for piece in re.findall(r".{1,6}", _text_of(message), re.DOTALL):
|
||
await asyncio.sleep(self.latency)
|
||
yield ChatGenerationChunk(message=AIMessageChunk(content=piece))
|