"""意图识别：JSON 解析、相对日期、关键词退路、以及模型不可用时 classify_and_extract 仍然给出可用结果。""" from __future__ import annotations import json from datetime import date, timedelta import pytest from backend import nlu class TestParseJson: def test_plain_object(self): assert nlu._parse_json('{"business_id": "a", "confidence": 0.9}') == {"business_id": "a", "confidence": 0.9} def test_fenced_block(self): assert nlu._parse_json('```json\n{"business_id": "a"}\n```') == {"business_id": "a"} def test_object_embedded_in_prose(self): # 模型偶尔会在 JSON 前后带一句话，取第一个到最后一个花括号之间的内容 assert nlu._parse_json('好的，结果如下：\n{"business_id": "a"}\n以上。') == {"business_id": "a"} @pytest.mark.parametrize("text", ["", " ", "根本没有 JSON", "{不是合法 JSON}", "[1, 2, 3]"]) def test_unusable_returns_none(self, text): assert nlu._parse_json(text) is None def test_top_level_array_rejected(self): # 只接受对象；数组不是调用方期望的形状 assert nlu._parse_json('[{"business_id": "a"}]') is None class TestRelativeDate: @pytest.mark.parametrize("word,offset", [("今天", 0), ("今日", 0), ("明天", 1), ("后天", 2), ("昨天", -1), ("前天", -2)]) def test_relative_words(self, word, offset): assert nlu._resolve_relative_date(f"帮我填{word}的查课日志") == (date.today() + timedelta(days=offset)).isoformat() @pytest.mark.parametrize("text,expected", [ ("听课时间 2026-09-17", "2026-09-17"), ("听课时间 2026/9/7", "2026-09-07"), ("听课时间 2026年9月7日", "2026-09-07"), ]) def test_absolute_dates_normalised(self, text, expected): assert nlu._resolve_relative_date(text) == expected def test_month_day_uses_current_year(self): assert nlu._resolve_relative_date("9月7日那节课") == f"{date.today().year}-09-07" def test_no_date_returns_none(self): assert nlu._resolve_relative_date("帮我看看有什么能办的") is None class TestPhraseScore: """退路打分用最长公共子串，不能要求整段短语原样出现。""" def test_partial_overlap_counts(self): # 知识里写「每月查课日志」，用户说「帮我填今天的查课日志」：整段不匹配，公共子串「查课日志」4 个字 assert nlu._phrase_score("每月查课日志", "帮我填今天的查课日志") == 4 def test_single_character_overlap_ignored(self): assert nlu._phrase_score("辅导员填写谈心谈话记录", "帮我填查课日志") == 0 def test_too_short_phrase_ignored(self): assert nlu._phrase_score("的", "帮我填今天的查课日志") == 0 def test_no_overlap(self): assert nlu._phrase_score("请假销假", "今天天气怎么样") == 0 class TestKeywordFallback: @pytest.mark.parametrize("message", [ "帮我填查课日志", "帮我填今天的查课日志，班级计科2301班", "我要填每月查课要求的日志", "听课记录还好还好", ]) def test_recognises_business_from_varied_phrasings(self, knowledge, message): # 回归：早先按整段短语匹配，这几句全都落到 generic_explore，模型不可用时等于没有保底 assert nlu._fallback(message, knowledge["businesses"], None)["business_id"] == "counselor_worklog" def test_matched_business_gets_higher_confidence(self, knowledge): assert nlu._fallback("帮我填查课日志", knowledge["businesses"], None)["confidence"] == 0.4 def test_unmatched_message_falls_back_to_generic_explore(self, knowledge): result = nlu._fallback("今天天气怎么样", knowledge["businesses"], None) assert result["business_id"] == "generic_explore" assert result["confidence"] == 0.2 def test_generic_explore_does_not_outscore_a_specific_business(self, knowledge): # generic_explore 的 intent 里有「查找某个入口」「有哪些字段」，不排除它就会抢走具体业务 assert nlu._fallback("帮我查一下课的日志入口", knowledge["businesses"], None)["business_id"] == "counselor_worklog" def test_fixed_business_id_wins_over_keywords(self, knowledge): result = nlu._fallback("帮我填查课日志", knowledge["businesses"], "generic_explore") assert result["business_id"] == "generic_explore" def test_extracts_class_and_date_without_model(self, knowledge): result = nlu._fallback("帮我填今天的查课日志，班级计科2301班", knowledge["businesses"], None) assert result["inputs"]["class_name"] == "计科2301班" assert result["inputs"]["observe_date"] == date.today().isoformat() def test_class_label_prefix_stripped(self, knowledge): # 「班级是计科2301班」里的「班级是」不能被当成班级名的一部分 result = nlu._fallback("班级是计科2301班", knowledge["businesses"], "counselor_worklog") assert result["inputs"]["class_name"] == "计科2301班" def test_only_declared_fields_extracted(self, knowledge): # generic_explore 没声明任何字段，即使原话里有班级也不抽 result = nlu._fallback("帮我看看计科2301班", knowledge["businesses"], "generic_explore") assert result["inputs"] == {} class TestClassifyAndExtract: async def test_uses_model_result_and_filters_undeclared_fields(self, knowledge, monkeypatch): async def fake_llm(prompt, cfg): return { "business_id": "counselor_worklog", "confidence": 0.95, "inputs": {"class_name": "计科2301班", "course_name": "高等数学", "不存在的字段": "x"}, } monkeypatch.setattr(nlu, "get_llm_config", lambda: {"api_key": "k", "model": "m", "base_url": "u"}) monkeypatch.setattr(nlu, "_call_llm", fake_llm) result = await nlu.classify_and_extract("帮我填查课日志", knowledge["businesses"]) assert result["source"] == "llm" assert result["business_id"] == "counselor_worklog" assert result["inputs"] == {"class_name": "计科2301班", "course_name": "高等数学"} async def test_model_failure_falls_back_to_keywords(self, knowledge, monkeypatch): async def boom(prompt, cfg): raise RuntimeError("Error code: 402 - Insufficient Balance") monkeypatch.setattr(nlu, "get_llm_config", lambda: {"api_key": "k", "model": "m", "base_url": "u"}) monkeypatch.setattr(nlu, "_call_llm", boom) result = await nlu.classify_and_extract("帮我填今天的查课日志，班级计科2301班", knowledge["businesses"]) assert result["source"] == "fallback" assert result["business_id"] == "counselor_worklog" assert result["inputs"]["class_name"] == "计科2301班" assert "402" in result["error"] async def test_no_api_key_skips_model_entirely(self, knowledge, monkeypatch): async def must_not_be_called(prompt, cfg): raise AssertionError("没有 Key 时不应该调用模型") monkeypatch.setattr(nlu, "get_llm_config", lambda: {"api_key": "", "model": "m", "base_url": "u"}) monkeypatch.setattr(nlu, "_call_llm", must_not_be_called) result = await nlu.classify_and_extract("帮我填查课日志", knowledge["businesses"]) assert result["source"] == "fallback" async def test_unknown_business_id_from_model_is_corrected(self, knowledge, monkeypatch): async def fake_llm(prompt, cfg): return {"business_id": "模型编出来的业务", "confidence": 0.9, "inputs": {}} monkeypatch.setattr(nlu, "get_llm_config", lambda: {"api_key": "k", "model": "m", "base_url": "u"}) monkeypatch.setattr(nlu, "_call_llm", fake_llm) result = await nlu.classify_and_extract("随便说点什么", knowledge["businesses"]) assert result["business_id"] == "generic_explore" @pytest.mark.parametrize("value", ["null", "None", "未提及", "", " "]) async def test_placeholder_values_are_dropped(self, knowledge, monkeypatch, value): async def fake_llm(prompt, cfg): return {"business_id": "counselor_worklog", "confidence": 0.9, "inputs": {"class_name": value}} monkeypatch.setattr(nlu, "get_llm_config", lambda: {"api_key": "k", "model": "m", "base_url": "u"}) monkeypatch.setattr(nlu, "_call_llm", fake_llm) result = await nlu.classify_and_extract("帮我填查课日志", knowledge["businesses"]) assert "class_name" not in result["inputs"] async def test_confidence_clamped_to_unit_range(self, knowledge, monkeypatch): async def fake_llm(prompt, cfg): return {"business_id": "counselor_worklog", "confidence": 9.9, "inputs": {}} monkeypatch.setattr(nlu, "get_llm_config", lambda: {"api_key": "k", "model": "m", "base_url": "u"}) monkeypatch.setattr(nlu, "_call_llm", fake_llm) result = await nlu.classify_and_extract("帮我填查课日志", knowledge["businesses"]) assert result["confidence"] == 1.0 async def test_existing_inputs_are_offered_to_the_prompt(self, knowledge, monkeypatch): seen = {} async def fake_llm(prompt, cfg): seen["prompt"] = prompt return {"business_id": "counselor_worklog", "confidence": 0.9, "inputs": {}} monkeypatch.setattr(nlu, "get_llm_config", lambda: {"api_key": "k", "model": "m", "base_url": "u"}) monkeypatch.setattr(nlu, "_call_llm", fake_llm) await nlu.classify_and_extract("补充一下", knowledge["businesses"], existing={"class_name": "计科2301班"}) assert "计科2301班" in seen["prompt"] class TestCallLlmKeyFallback: """主 Key 402 时 _call_llm 必须切到兜底 Key 重发，而不是让任务失败。""" @staticmethod def _stub_openai(monkeypatch, keys_seen: list[str]): import openai class FakeCompletions: def __init__(self, key): self._key = key async def create(self, **kwargs): keys_seen.append(self._key) if self._key == "PRIMARY": import httpx raise openai.APIStatusError( "Error code: 402 - Insufficient Balance", response=httpx.Response(402, request=httpx.Request("POST", "http://x")), body=None, ) return type("R", (), {"choices": [type("C", (), {"message": type("M", (), {"content": '{"ok": 1}'})()})()]})() class FakeClient: def __init__(self, api_key, base_url, timeout): self.chat = type("Chat", (), {"completions": FakeCompletions(api_key)})() monkeypatch.setattr(openai, "AsyncOpenAI", FakeClient) async def test_switches_to_fallback_on_402(self, monkeypatch): from backend.browser import llm_keys marked = {} monkeypatch.setattr(llm_keys, "mark_primary_exhausted", lambda reason="": marked.update(hit=True)) monkeypatch.setattr(llm_keys, "clear_primary_exhausted", lambda: None) keys_seen: list[str] = [] self._stub_openai(monkeypatch, keys_seen) cfg = {"api_key": "PRIMARY", "primary_api_key": "PRIMARY", "fallback_api_key": "FALLBACK", "base_url": "http://x", "model": "m"} assert await nlu._call_llm("hi", cfg) == {"ok": 1} assert keys_seen == ["PRIMARY", "FALLBACK"] assert marked.get("hit") is True async def test_without_fallback_the_402_propagates(self, monkeypatch): keys_seen: list[str] = [] self._stub_openai(monkeypatch, keys_seen) cfg = {"api_key": "PRIMARY", "primary_api_key": "PRIMARY", "fallback_api_key": "", "base_url": "http://x", "model": "m"} with pytest.raises(Exception, match="402"): await nlu._call_llm("hi", cfg) assert keys_seen == ["PRIMARY"] def test_prompt_lists_every_business_and_required_field(knowledge): prompt = nlu._build_prompt("帮我填查课日志", knowledge["businesses"], {}, None) assert "counselor_worklog" in prompt and "generic_explore" in prompt assert "class_name" in prompt and "observe_date" in prompt assert json.dumps # 提示词要求模型输出 JSON assert "JSON" in prompt.upper()