352 lines
12 KiB
Python
352 lines
12 KiB
Python
"""M4 LLM 投资事件抽取测试。
|
|
|
|
不依赖真实 LLM API,所有调用通过 mock 注入响应。
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import asyncio
|
|
import json
|
|
from datetime import datetime
|
|
from pathlib import Path
|
|
from unittest.mock import AsyncMock, MagicMock
|
|
|
|
import pytest
|
|
|
|
from extractor import Article
|
|
from llm import (
|
|
EVENT_TYPES,
|
|
EventExtraction,
|
|
ExtractedEvent,
|
|
LLMCallError,
|
|
PromptTemplate,
|
|
Sentiment,
|
|
extract_event,
|
|
extract_event_async,
|
|
load_llm_config,
|
|
parse_event_json,
|
|
)
|
|
from llm.client import LLMConfig
|
|
from llm.extractor import _extract_json_object
|
|
|
|
# --------------------------------------------------------------------------- #
|
|
# fixtures
|
|
# --------------------------------------------------------------------------- #
|
|
|
|
def _article(
|
|
*,
|
|
title: str = "宁德时代签订 100GWh 长期供货协议",
|
|
content: str = "宁德时代(300750)与某车企签 5 年 100GWh 协议,涉及金额超 1500 亿。" * 3,
|
|
publish_time: datetime | None = datetime(2026, 6, 16, 10, 0),
|
|
) -> Article:
|
|
return Article(
|
|
source_id="cls",
|
|
url="https://www.cls.cn/detail/1",
|
|
url_hash="abc1234567890000",
|
|
title=title,
|
|
content=content,
|
|
publish_time=publish_time,
|
|
word_count=len(content),
|
|
)
|
|
|
|
|
|
@pytest.fixture
|
|
def fake_config() -> LLMConfig:
|
|
return LLMConfig(
|
|
provider="deepseek",
|
|
model="deepseek-chat",
|
|
api_key="sk-fake",
|
|
base_url="https://api.deepseek.com",
|
|
)
|
|
|
|
|
|
def _mock_completion(content: str, prompt_tokens: int = 100, completion_tokens: int = 50) -> MagicMock:
|
|
"""构造与 openai SDK 返回兼容的 mock 对象。"""
|
|
msg = MagicMock()
|
|
msg.content = content
|
|
choice = MagicMock()
|
|
choice.message = msg
|
|
usage = MagicMock()
|
|
usage.prompt_tokens = prompt_tokens
|
|
usage.completion_tokens = completion_tokens
|
|
resp = MagicMock()
|
|
resp.choices = [choice]
|
|
resp.usage = usage
|
|
return resp
|
|
|
|
|
|
# --------------------------------------------------------------------------- #
|
|
# EventExtraction 模型校验
|
|
# --------------------------------------------------------------------------- #
|
|
|
|
def test_event_extraction_minimal_valid() -> None:
|
|
e = EventExtraction(sentiment="positive", importance=4, event_type="重大合同")
|
|
assert e.sentiment == Sentiment.POSITIVE
|
|
assert e.importance == 4
|
|
|
|
|
|
def test_event_extraction_normalizes_stock_codes() -> None:
|
|
e = EventExtraction(
|
|
stock_codes=["300750.sz", " 300750.SZ ", "abc", "12345", "600519"],
|
|
sentiment="positive", importance=3, event_type="其他",
|
|
)
|
|
# 大小写 / 空白被规范;非法被过滤;去重
|
|
assert e.stock_codes == ["300750.SZ", "600519"]
|
|
|
|
|
|
def test_event_extraction_filters_empty_lists() -> None:
|
|
e = EventExtraction(
|
|
stock_codes=[], company_names=["", " ", "宁德时代", "宁德时代"],
|
|
industries=[],
|
|
sentiment="neutral", importance=1, event_type="其他",
|
|
)
|
|
assert e.company_names == ["宁德时代"]
|
|
assert e.stock_codes == []
|
|
|
|
|
|
def test_event_extraction_rejects_importance_out_of_range() -> None:
|
|
with pytest.raises(Exception): # noqa: B017
|
|
EventExtraction(sentiment="positive", importance=0, event_type="其他")
|
|
with pytest.raises(Exception): # noqa: B017
|
|
EventExtraction(sentiment="positive", importance=6, event_type="其他")
|
|
|
|
|
|
def test_event_extraction_normalizes_blank_event_type() -> None:
|
|
e = EventExtraction(sentiment="neutral", importance=1, event_type=" ")
|
|
assert e.event_type == "其他"
|
|
|
|
|
|
def test_event_types_constant_includes_common() -> None:
|
|
for must in ["业绩预告", "合作签约", "监管处罚", "其他"]:
|
|
assert must in EVENT_TYPES
|
|
|
|
|
|
# --------------------------------------------------------------------------- #
|
|
# JSON 提取与解析
|
|
# --------------------------------------------------------------------------- #
|
|
|
|
def test_extract_json_object_strips_fence() -> None:
|
|
s = '```json\n{"a": 1}\n```'
|
|
assert _extract_json_object(s) == '{"a": 1}'
|
|
|
|
|
|
def test_extract_json_object_picks_first_object() -> None:
|
|
s = '前置说明\n{"a": 1}\n更多文字'
|
|
assert _extract_json_object(s) == '{"a": 1}'
|
|
|
|
|
|
def test_extract_json_object_handles_nested() -> None:
|
|
s = '{"a": {"b": 2}}'
|
|
assert _extract_json_object(s) == '{"a": {"b": 2}}'
|
|
|
|
|
|
def test_parse_event_json_ok() -> None:
|
|
raw = json.dumps({
|
|
"stock_codes": ["300750.SZ"],
|
|
"company_names": ["宁德时代"],
|
|
"industries": ["动力电池"],
|
|
"sentiment": "positive",
|
|
"importance": 5,
|
|
"event_type": "重大合同",
|
|
"summary": "签订长期供货协议",
|
|
})
|
|
e = parse_event_json(raw)
|
|
assert e.sentiment == Sentiment.POSITIVE
|
|
assert e.stock_codes == ["300750.SZ"]
|
|
|
|
|
|
def test_parse_event_json_invalid_json_raises() -> None:
|
|
with pytest.raises(LLMCallError):
|
|
parse_event_json("not a json")
|
|
|
|
|
|
def test_parse_event_json_non_object_raises() -> None:
|
|
with pytest.raises(LLMCallError):
|
|
parse_event_json('["array"]')
|
|
|
|
|
|
def test_parse_event_json_schema_invalid_raises() -> None:
|
|
with pytest.raises(LLMCallError):
|
|
parse_event_json('{"importance": 99}') # 缺 sentiment + event_type 且 importance 越界
|
|
|
|
|
|
# --------------------------------------------------------------------------- #
|
|
# PromptTemplate
|
|
# --------------------------------------------------------------------------- #
|
|
|
|
def test_prompt_template_renders_placeholders(tmp_path: Path) -> None:
|
|
tpl_file = tmp_path / "tpl.md"
|
|
tpl_file.write_text(
|
|
"标题:{title}\n时间:{publish_time}\n源:{source_name}\n内容:\n{content}\nEND",
|
|
encoding="utf-8",
|
|
)
|
|
tpl = PromptTemplate(tpl_file)
|
|
art = _article()
|
|
rendered = tpl.render(art)
|
|
assert "标题:" + art.title in rendered
|
|
assert "2026-06-16" in rendered
|
|
assert "源:财联社" in rendered or "源:cls" in rendered # source_name 默认空,落到 source_id
|
|
assert art.content[:30] in rendered
|
|
|
|
|
|
def test_prompt_template_truncates_long_content(tmp_path: Path) -> None:
|
|
tpl_file = tmp_path / "tpl.md"
|
|
tpl_file.write_text("{content}", encoding="utf-8")
|
|
tpl = PromptTemplate(tpl_file)
|
|
art = _article(content="字" * 20000)
|
|
rendered = tpl.render(art)
|
|
assert "[正文过长已截断]" in rendered
|
|
assert len(rendered) < 20000
|
|
|
|
|
|
def test_prompt_template_default_path_loads() -> None:
|
|
"""项目内置 prompts/event_extraction.md 必须可加载,作为回归保护。"""
|
|
real = Path("prompts/event_extraction.md")
|
|
if not real.is_file():
|
|
pytest.skip("prompts/event_extraction.md 未找到")
|
|
tpl = PromptTemplate(real)
|
|
out = tpl.render(_article())
|
|
assert "{title}" not in out
|
|
assert "{content}" not in out
|
|
|
|
|
|
# --------------------------------------------------------------------------- #
|
|
# extract_event(同步,带重试)
|
|
# --------------------------------------------------------------------------- #
|
|
|
|
def test_extract_event_succeeds_first_try(fake_config: LLMConfig) -> None:
|
|
client = MagicMock()
|
|
raw = json.dumps({
|
|
"stock_codes": ["300750.SZ"],
|
|
"company_names": ["宁德时代"],
|
|
"industries": ["动力电池"],
|
|
"sentiment": "positive",
|
|
"importance": 5,
|
|
"event_type": "重大合同",
|
|
"summary": "100GWh 合作",
|
|
})
|
|
client.chat.completions.create = MagicMock(return_value=_mock_completion(raw))
|
|
|
|
result = extract_event(client, fake_config, _article())
|
|
assert isinstance(result, ExtractedEvent)
|
|
assert result.attempts == 1
|
|
assert result.event.sentiment == Sentiment.POSITIVE
|
|
assert result.provider == "deepseek"
|
|
assert result.prompt_tokens == 100
|
|
|
|
|
|
def test_extract_event_retries_on_invalid_json(fake_config: LLMConfig) -> None:
|
|
"""第 1/2 次返回非法 JSON,第 3 次成功。"""
|
|
valid = json.dumps({
|
|
"sentiment": "neutral", "importance": 1, "event_type": "其他",
|
|
})
|
|
client = MagicMock()
|
|
client.chat.completions.create = MagicMock(side_effect=[
|
|
_mock_completion("not a json"),
|
|
_mock_completion('{"sentiment":"???"}'), # schema 校验失败
|
|
_mock_completion(valid),
|
|
])
|
|
result = extract_event(client, fake_config, _article(), max_attempts=3)
|
|
assert result.attempts == 3
|
|
assert client.chat.completions.create.call_count == 3
|
|
|
|
|
|
def test_extract_event_gives_up_after_max(fake_config: LLMConfig) -> None:
|
|
client = MagicMock()
|
|
client.chat.completions.create = MagicMock(
|
|
return_value=_mock_completion("not a json")
|
|
)
|
|
with pytest.raises(LLMCallError) as exc:
|
|
extract_event(client, fake_config, _article(), max_attempts=2)
|
|
assert exc.value.attempts == 2
|
|
assert client.chat.completions.create.call_count == 2
|
|
|
|
|
|
def test_extract_event_handles_network_exception(fake_config: LLMConfig) -> None:
|
|
client = MagicMock()
|
|
valid = json.dumps({
|
|
"sentiment": "negative", "importance": 3, "event_type": "监管处罚",
|
|
})
|
|
client.chat.completions.create = MagicMock(side_effect=[
|
|
TimeoutError("net hang"),
|
|
_mock_completion(valid),
|
|
])
|
|
result = extract_event(client, fake_config, _article(), max_attempts=2)
|
|
assert result.attempts == 2
|
|
assert result.event.sentiment == Sentiment.NEGATIVE
|
|
|
|
|
|
# --------------------------------------------------------------------------- #
|
|
# extract_event_async
|
|
# --------------------------------------------------------------------------- #
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_extract_event_async_succeeds(fake_config: LLMConfig) -> None:
|
|
client = MagicMock()
|
|
valid = json.dumps({
|
|
"stock_codes": ["600519"],
|
|
"company_names": ["贵州茅台"],
|
|
"sentiment": "neutral",
|
|
"importance": 2,
|
|
"event_type": "财报披露",
|
|
"summary": "披露半年报",
|
|
})
|
|
client.chat.completions.create = AsyncMock(return_value=_mock_completion(valid))
|
|
|
|
sem = asyncio.Semaphore(2)
|
|
result = await extract_event_async(
|
|
client, fake_config, _article(), semaphore=sem,
|
|
)
|
|
assert result.event.stock_codes == ["600519"]
|
|
assert result.event.event_type == "财报披露"
|
|
|
|
|
|
@pytest.mark.asyncio
|
|
async def test_extract_event_async_retries(fake_config: LLMConfig) -> None:
|
|
valid = json.dumps({"sentiment": "positive", "importance": 4, "event_type": "其他"})
|
|
client = MagicMock()
|
|
client.chat.completions.create = AsyncMock(side_effect=[
|
|
ValueError("transient"),
|
|
_mock_completion(valid),
|
|
])
|
|
result = await extract_event_async(
|
|
client, fake_config, _article(), max_attempts=2,
|
|
)
|
|
assert result.attempts == 2
|
|
|
|
|
|
# --------------------------------------------------------------------------- #
|
|
# load_llm_config
|
|
# --------------------------------------------------------------------------- #
|
|
|
|
def test_load_llm_config_deepseek_from_env(monkeypatch: pytest.MonkeyPatch) -> None:
|
|
monkeypatch.setenv("LLM_PROVIDER", "deepseek")
|
|
monkeypatch.setenv("DEEPSEEK_API_KEY", "sk-test-deepseek")
|
|
monkeypatch.delenv("LLM_MODEL", raising=False)
|
|
cfg = load_llm_config()
|
|
assert cfg.provider == "deepseek"
|
|
assert cfg.api_key == "sk-test-deepseek"
|
|
assert cfg.model.startswith("deepseek")
|
|
assert "deepseek" in cfg.base_url
|
|
|
|
|
|
def test_load_llm_config_qwen_from_env(monkeypatch: pytest.MonkeyPatch) -> None:
|
|
monkeypatch.setenv("LLM_PROVIDER", "qwen")
|
|
monkeypatch.setenv("DASHSCOPE_API_KEY", "sk-test-qwen")
|
|
monkeypatch.delenv("LLM_MODEL", raising=False)
|
|
cfg = load_llm_config()
|
|
assert cfg.provider == "qwen"
|
|
assert cfg.api_key == "sk-test-qwen"
|
|
assert "dashscope" in cfg.base_url or "aliyuncs" in cfg.base_url
|
|
|
|
|
|
def test_load_llm_config_unknown_provider_raises(monkeypatch: pytest.MonkeyPatch) -> None:
|
|
with pytest.raises(ValueError):
|
|
load_llm_config(provider="anthropic")
|
|
|
|
|
|
def test_load_llm_config_missing_key_raises(monkeypatch: pytest.MonkeyPatch) -> None:
|
|
monkeypatch.delenv("DEEPSEEK_API_KEY", raising=False)
|
|
with pytest.raises(ValueError):
|
|
load_llm_config(provider="deepseek")
|