feat: add persisted batch text testing
This commit is contained in:
Xin Wang
386
backend/test_schemas.py
Normal file
386
backend/test_schemas.py
Normal file
@@ -0,0 +1,386 @@
|
||||
"""Contracts for persisted text test cases and batch execution results."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import re
|
||||
from datetime import datetime
|
||||
from typing import Annotated, Any, Literal, Union
|
||||
|
||||
from pydantic import Field, field_validator, model_validator
|
||||
|
||||
from schemas import CamelModel
|
||||
|
||||
|
||||
TestCaseInputMode = Literal[
|
||||
"fixed_script_text",
|
||||
"fixed_script_turn_voice",
|
||||
"fixed_script_continuous_voice",
|
||||
"user_sim_text",
|
||||
"user_sim_voice",
|
||||
]
|
||||
ContextRole = Literal["agent", "user", "tool_call", "tool_result"]
|
||||
EvaluationKind = Literal["reply", "tool_call", "overall"]
|
||||
EvaluationStatus = Literal["pass", "fail"]
|
||||
CaseStatus = Literal["waiting", "running", "pass", "fail", "error", "skipped"]
|
||||
RunStatus = Literal["queued", "running", "completed", "cancelled"]
|
||||
ErrorStage = Literal["pipeline", "model", "tool", "evaluation"]
|
||||
StopReason = Literal["manual", "assertion_failure", "execution_error"]
|
||||
|
||||
|
||||
class ContextTurn(CamelModel):
|
||||
role: ContextRole
|
||||
content: str = Field(max_length=20_000)
|
||||
tool_name: str | None = Field(default=None, max_length=128)
|
||||
tool_call_id: str | None = Field(default=None, max_length=128)
|
||||
is_error: bool | None = None
|
||||
|
||||
@model_validator(mode="after")
|
||||
def validate_content(self):
|
||||
self.content = self.content.strip()
|
||||
self.tool_call_id = (self.tool_call_id or "").strip() or None
|
||||
if not self.content:
|
||||
raise ValueError("上下文内容不能为空")
|
||||
if self.role in {"tool_call", "tool_result"}:
|
||||
self.tool_name = (self.tool_name or "").strip()
|
||||
if not self.tool_name:
|
||||
raise ValueError("工具上下文必须填写工具名称")
|
||||
try:
|
||||
payload = json.loads(self.content)
|
||||
except json.JSONDecodeError as exc:
|
||||
raise ValueError("工具上下文必须是有效 JSON") from exc
|
||||
if self.role == "tool_call" and not isinstance(payload, dict):
|
||||
raise ValueError("Tool Call 参数必须是 JSON 对象")
|
||||
return self
|
||||
|
||||
|
||||
class ToolParamAssertion(CamelModel):
|
||||
name: str = Field(min_length=1, max_length=128)
|
||||
match_mode: Literal["exact", "regex", "llm"]
|
||||
value: str = Field(min_length=1, max_length=4_000)
|
||||
|
||||
@field_validator("name", "value")
|
||||
@classmethod
|
||||
def strip_required_text(cls, value: str) -> str:
|
||||
value = value.strip()
|
||||
if not value:
|
||||
raise ValueError("参数名称和值不能为空")
|
||||
return value
|
||||
|
||||
@model_validator(mode="after")
|
||||
def validate_regex(self):
|
||||
if self.match_mode == "regex":
|
||||
try:
|
||||
re.compile(self.value)
|
||||
except re.error as exc:
|
||||
raise ValueError(f"正则表达式无效: {exc}") from exc
|
||||
return self
|
||||
|
||||
|
||||
class ToolMockResponse(CamelModel):
|
||||
outcome: Literal["success", "error"] = "success"
|
||||
body: str = Field(default='{"status":"ok"}', max_length=100_000)
|
||||
delay_ms: int = Field(default=0, ge=0, le=60_000)
|
||||
|
||||
@model_validator(mode="after")
|
||||
def validate_json_body(self):
|
||||
if not self.body.strip():
|
||||
raise ValueError("Mock 工具返回值不能为空")
|
||||
try:
|
||||
json.loads(self.body)
|
||||
except json.JSONDecodeError as exc:
|
||||
raise ValueError("Mock 工具返回值必须是有效 JSON") from exc
|
||||
return self
|
||||
|
||||
|
||||
class ReplyExpectedBehavior(CamelModel):
|
||||
id: str = Field(min_length=1, max_length=128)
|
||||
type: Literal["reply"] = "reply"
|
||||
assertion_type: Literal["keyword", "llm"]
|
||||
keywords: list[str] = Field(default_factory=list, max_length=50)
|
||||
keyword_match_mode: Literal["any", "all"] = "any"
|
||||
negate_keywords: bool = False
|
||||
llm_criteria: str = Field(default="", max_length=4_000)
|
||||
|
||||
@model_validator(mode="after")
|
||||
def validate_assertion(self):
|
||||
self.keywords = [item.strip() for item in self.keywords if item.strip()]
|
||||
self.llm_criteria = self.llm_criteria.strip()
|
||||
if self.assertion_type == "keyword" and not self.keywords:
|
||||
raise ValueError("关键词断言至少需要一个关键词")
|
||||
if self.assertion_type == "llm" and not self.llm_criteria:
|
||||
raise ValueError("LLM 判断要求不能为空")
|
||||
return self
|
||||
|
||||
|
||||
class ToolCallExpectedBehavior(CamelModel):
|
||||
id: str = Field(min_length=1, max_length=128)
|
||||
type: Literal["tool_call"] = "tool_call"
|
||||
tool_id: str = Field(min_length=1, max_length=40)
|
||||
function_name: str = Field(min_length=1, max_length=128)
|
||||
expectation: Literal["called", "not_called"]
|
||||
min_calls: int = Field(default=1, ge=0, le=100)
|
||||
max_calls: int | None = Field(default=None, ge=0, le=100)
|
||||
param_assertions: list[ToolParamAssertion] = Field(default_factory=list, max_length=50)
|
||||
mock_response: ToolMockResponse = Field(default_factory=ToolMockResponse)
|
||||
|
||||
@model_validator(mode="after")
|
||||
def validate_call_expectation(self):
|
||||
self.tool_id = self.tool_id.strip()
|
||||
self.function_name = self.function_name.strip()
|
||||
if not self.tool_id or not self.function_name:
|
||||
raise ValueError("必须选择有效的工具")
|
||||
if self.expectation == "called" and self.min_calls < 1:
|
||||
raise ValueError("应调用工具时最少调用次数必须大于 0")
|
||||
if self.max_calls is not None and self.max_calls < self.min_calls:
|
||||
raise ValueError("最多调用次数不能小于最少调用次数")
|
||||
return self
|
||||
|
||||
|
||||
ExpectedBehavior = Annotated[
|
||||
Union[ReplyExpectedBehavior, ToolCallExpectedBehavior],
|
||||
Field(discriminator="type"),
|
||||
]
|
||||
|
||||
|
||||
class FixedInputTurn(CamelModel):
|
||||
id: str = Field(min_length=1, max_length=128)
|
||||
user_input: str = Field(min_length=1, max_length=20_000)
|
||||
behaviors: list[ExpectedBehavior] = Field(default_factory=list, max_length=100)
|
||||
|
||||
@field_validator("user_input")
|
||||
@classmethod
|
||||
def strip_user_input(cls, value: str) -> str:
|
||||
value = value.strip()
|
||||
if not value:
|
||||
raise ValueError("用户输入不能为空")
|
||||
return value
|
||||
|
||||
|
||||
class OverallCriterion(CamelModel):
|
||||
id: str = Field(min_length=1, max_length=128)
|
||||
type: Literal["llm"] = "llm"
|
||||
name: str = Field(min_length=1, max_length=80)
|
||||
criteria: str = Field(min_length=1, max_length=2_000)
|
||||
|
||||
@field_validator("name", "criteria")
|
||||
@classmethod
|
||||
def strip_criterion(cls, value: str) -> str:
|
||||
value = value.strip()
|
||||
if not value:
|
||||
raise ValueError("整体评估标准不能为空")
|
||||
return value
|
||||
|
||||
|
||||
class TestCaseDefinition(CamelModel):
|
||||
context_turns: list[ContextTurn] = Field(default_factory=list, max_length=100)
|
||||
turns: list[FixedInputTurn] = Field(min_length=1, max_length=100)
|
||||
overall_criteria: list[OverallCriterion] = Field(default_factory=list, max_length=50)
|
||||
|
||||
@model_validator(mode="after")
|
||||
def validate_has_expectation(self):
|
||||
if not any(turn.behaviors for turn in self.turns) and not self.overall_criteria:
|
||||
raise ValueError("至少配置一项有效的预期行为或整体评估标准")
|
||||
return self
|
||||
|
||||
@model_validator(mode="after")
|
||||
def validate_tool_context_pairs(self):
|
||||
pending: list[tuple[str | None, str]] = []
|
||||
seen_ids: set[str] = set()
|
||||
for index, turn in enumerate(self.context_turns):
|
||||
if turn.role == "tool_call":
|
||||
call_id = (turn.tool_call_id or "").strip() or None
|
||||
if call_id and call_id in seen_ids:
|
||||
raise ValueError(f"上下文第 {index + 1} 条 Tool Call ID 重复")
|
||||
if call_id:
|
||||
seen_ids.add(call_id)
|
||||
pending.append((call_id, turn.tool_name or ""))
|
||||
continue
|
||||
if turn.role != "tool_result":
|
||||
continue
|
||||
|
||||
result_id = (turn.tool_call_id or "").strip() or None
|
||||
matched_index = -1
|
||||
for pending_index in range(len(pending) - 1, -1, -1):
|
||||
call_id, tool_name = pending[pending_index]
|
||||
if result_id:
|
||||
matches = call_id == result_id
|
||||
else:
|
||||
matches = tool_name == turn.tool_name
|
||||
if matches:
|
||||
matched_index = pending_index
|
||||
break
|
||||
if matched_index < 0:
|
||||
raise ValueError(
|
||||
f"上下文第 {index + 1} 条 Tool Result 没有匹配的 Tool Call"
|
||||
)
|
||||
_call_id, tool_name = pending.pop(matched_index)
|
||||
if tool_name != turn.tool_name:
|
||||
raise ValueError(
|
||||
f"上下文第 {index + 1} 条 Tool Result 的工具名称不匹配"
|
||||
)
|
||||
|
||||
if pending:
|
||||
names = "、".join(tool_name for _call_id, tool_name in pending)
|
||||
raise ValueError(f"上下文 Tool Call 缺少对应的 Tool Result：{names}")
|
||||
return self
|
||||
|
||||
|
||||
class TestSuiteCreate(CamelModel):
|
||||
name: str = Field(min_length=1, max_length=128)
|
||||
description: str = Field(default="", max_length=2_048)
|
||||
|
||||
@field_validator("name")
|
||||
@classmethod
|
||||
def strip_name(cls, value: str) -> str:
|
||||
value = value.strip()
|
||||
if not value:
|
||||
raise ValueError("测试集名称不能为空")
|
||||
return value
|
||||
|
||||
|
||||
class TestSuiteUpdate(TestSuiteCreate):
|
||||
pass
|
||||
|
||||
|
||||
class TestSuiteOut(CamelModel):
|
||||
id: str
|
||||
name: str
|
||||
description: str
|
||||
case_count: int = 0
|
||||
passed_count: int = 0
|
||||
run_count: int = 0
|
||||
updated_at: datetime
|
||||
|
||||
|
||||
class TestCaseWrite(CamelModel):
|
||||
name: str = Field(min_length=1, max_length=128)
|
||||
description: str = Field(default="", max_length=2_048)
|
||||
input_mode: TestCaseInputMode = "fixed_script_text"
|
||||
context_turns: list[ContextTurn] = Field(default_factory=list, max_length=100)
|
||||
turns: list[FixedInputTurn] = Field(min_length=1, max_length=100)
|
||||
overall_criteria: list[OverallCriterion] = Field(default_factory=list, max_length=50)
|
||||
|
||||
@field_validator("name")
|
||||
@classmethod
|
||||
def strip_name(cls, value: str) -> str:
|
||||
value = value.strip()
|
||||
if not value:
|
||||
raise ValueError("测试用例名称不能为空")
|
||||
return value
|
||||
|
||||
@model_validator(mode="after")
|
||||
def validate_mvp_definition(self):
|
||||
if self.input_mode != "fixed_script_text":
|
||||
raise ValueError("第一版只支持固定脚本 · 文字")
|
||||
TestCaseDefinition(
|
||||
context_turns=self.context_turns,
|
||||
turns=self.turns,
|
||||
overall_criteria=self.overall_criteria,
|
||||
)
|
||||
return self
|
||||
|
||||
def definition(self) -> TestCaseDefinition:
|
||||
return TestCaseDefinition(
|
||||
context_turns=self.context_turns,
|
||||
turns=self.turns,
|
||||
overall_criteria=self.overall_criteria,
|
||||
)
|
||||
|
||||
|
||||
class TestCaseOut(TestCaseWrite):
|
||||
id: str
|
||||
suite_id: str
|
||||
last_result: Literal["pass", "fail", "not_run"]
|
||||
sort_order: int
|
||||
updated_at: datetime
|
||||
|
||||
|
||||
class TestCaseOrderIn(CamelModel):
|
||||
case_ids: list[str] = Field(min_length=1, max_length=1_000)
|
||||
|
||||
|
||||
class TestCaseBulkDeleteIn(CamelModel):
|
||||
case_ids: list[str] = Field(min_length=1, max_length=1_000)
|
||||
|
||||
|
||||
class BatchRunConfig(CamelModel):
|
||||
concurrency: int = Field(default=3, ge=1, le=20)
|
||||
timeout_secs: int = Field(default=60, ge=1, le=600)
|
||||
failure_strategy: Literal["continue", "stop_on_fail"] = "continue"
|
||||
error_retry_count: int = Field(default=0, ge=0, le=3)
|
||||
error_strategy: Literal["continue", "stop_on_error"] = "continue"
|
||||
|
||||
|
||||
class BatchRunCreate(CamelModel):
|
||||
assistant_id: str = Field(min_length=1, max_length=40)
|
||||
evaluator_model_resource_id: str = Field(min_length=1, max_length=40)
|
||||
case_ids: list[str] = Field(min_length=1, max_length=1_000)
|
||||
config: BatchRunConfig = Field(default_factory=BatchRunConfig)
|
||||
title: str | None = Field(default=None, max_length=256)
|
||||
|
||||
@field_validator("case_ids")
|
||||
@classmethod
|
||||
def unique_case_ids(cls, values: list[str]) -> list[str]:
|
||||
normalized = list(dict.fromkeys(value.strip() for value in values if value.strip()))
|
||||
if not normalized:
|
||||
raise ValueError("至少选择一个测试用例")
|
||||
return normalized
|
||||
|
||||
|
||||
class BatchExecutionError(CamelModel):
|
||||
code: str
|
||||
message: str
|
||||
stage: ErrorStage
|
||||
retryable: bool
|
||||
|
||||
|
||||
class BatchEvaluationResult(CamelModel):
|
||||
id: str
|
||||
label: str
|
||||
kind: EvaluationKind
|
||||
status: EvaluationStatus
|
||||
expected: str
|
||||
actual: str
|
||||
reason: str = ""
|
||||
|
||||
|
||||
class BatchToolCallRecord(CamelModel):
|
||||
id: str
|
||||
function_name: str
|
||||
arguments_json: str
|
||||
result_json: str
|
||||
outcome: Literal["success", "error"]
|
||||
duration_ms: int = 0
|
||||
|
||||
|
||||
class BatchTurnResult(CamelModel):
|
||||
id: str
|
||||
index: int
|
||||
user_input: str
|
||||
assistant_reply: str
|
||||
tool_calls: list[BatchToolCallRecord] = Field(default_factory=list)
|
||||
evaluations: list[BatchEvaluationResult] = Field(default_factory=list)
|
||||
|
||||
|
||||
class BatchRunCaseOut(CamelModel):
|
||||
id: str
|
||||
name: str
|
||||
status: CaseStatus
|
||||
turns: list[BatchTurnResult] = Field(default_factory=list)
|
||||
overall_criteria: list[BatchEvaluationResult] = Field(default_factory=list)
|
||||
attempt_count: int
|
||||
max_attempts: int
|
||||
execution_error: BatchExecutionError | None = None
|
||||
|
||||
|
||||
class BatchRunSnapshotOut(CamelModel):
|
||||
id: str
|
||||
status: RunStatus
|
||||
title: str
|
||||
assistant_name: str
|
||||
config: dict[str, Any]
|
||||
cases: list[BatchRunCaseOut]
|
||||
started_at: datetime
|
||||
finished_at: datetime | None
|
||||
stop_reason: StopReason | None
|
||||
Reference in New Issue
Block a user