"""nodes/llm.py 的模块级测试(数据 → 测试过程 → 验证结果)。 被测模块:`nodes/llm.py`(LLM 翻译节点:分批请求 + 按 ID 回填),可独立调用。 网络属于允许 mock 的 I/O 边界:单元用例注入假 HTTP 响应验证请求体、ID 校验、 重试与回填;集成用例调用真实 LLM 验证真实字幕翻译质量。 """ from __future__ import annotations import json import urllib.error import urllib.request from pathlib import Path import pytest from nodes.llm import ( CHUNK_SIZE, MAX_BATCH_RETRIES, _parse_translations, _system_prompt, invoke, translate_lines, ) from wov_sdk.models import InvokeRequest # 模块专用数据目录(真实字幕产物缺失时相关用例跳过)。 DATA_DIR = Path(__file__).resolve().parent / "data" class _FakeHTTPResponse: """假的 HTTP 响应:返回预置 JSON 体(供 urlopen mock 使用)。""" def __init__(self, payload: dict) -> None: self._data = json.dumps(payload, ensure_ascii=False).encode("utf-8") def read(self) -> bytes: return self._data def __enter__(self): return self def __exit__(self, *exc) -> None: return None def _llm_reply(translations: list[tuple[int, str]], total_tokens: int = 42) -> _FakeHTTPResponse: """构造 OpenAI 兼容接口的响应体(content 为 {id,text} JSON 数组)。""" content = json.dumps( [{"id": i, "text": t} for i, t in translations], ensure_ascii=False ) return _FakeHTTPResponse({ "choices": [{"message": {"content": content}}], "usage": {"total_tokens": total_tokens}, }) def _capture_urlopen(calls: list[dict], responses: list[_FakeHTTPResponse]): """返回一个假 urlopen:记录请求体,按顺序返回预置响应。""" def fake_urlopen(http_request, timeout=None): calls.append({ "url": http_request.full_url, "body": json.loads(http_request.data.decode("utf-8")), "headers": dict(http_request.headers), "timeout": timeout, }) return responses.pop(0) if responses else _llm_reply([]) return fake_urlopen # --------------------------------------------------------------------------- # 提示词与响应解析(纯函数) # --------------------------------------------------------------------------- def test_system_prompt_mentions_target_language_and_json_contract() -> None: """系统提示词说明目标语言与 JSON 条目契约(时间戳不进入模型)。""" # 数据:目标语言 zh-CN。 # 测试过程 prompt = _system_prompt("zh-CN") # 验证结果 assert "zh-CN" in prompt assert "id" in prompt and "text" in prompt def test_parse_translations_accepts_out_of_order_ids() -> None: """乱序返回的条目按 ID 回填(不依赖数组顺序)。""" # 数据:ID 为 3、1、2 的乱序结果。 content = json.dumps([ {"id": 3, "text": "三"}, {"id": 1, "text": "一"}, {"id": 2, "text": "二"}, ]) # 测试过程 parsed = _parse_translations(content, {1, 2, 3}) # 验证结果 assert parsed == {1: "一", 2: "二", 3: "三"} def test_parse_translations_rejects_missing_id() -> None: """缺少任一 ID 时明确报错(不允许静默漏译导致时间轴错位)。""" # 数据:缺少 ID 2。 content = json.dumps([{"id": 1, "text": "一"}, {"id": 3, "text": "三"}]) # 测试过程与验证结果 with pytest.raises(ValueError, match="missing"): _parse_translations(content, {1, 2, 3}) def test_parse_translations_rejects_duplicate_id() -> None: """重复 ID 报错。""" # 数据:ID 1 出现两次。 content = json.dumps([{"id": 1, "text": "一"}, {"id": 1, "text": "壹"}]) # 测试过程与验证结果 with pytest.raises(ValueError, match="duplicate"): _parse_translations(content, {1}) def test_parse_translations_rejects_empty_text() -> None: """空正文或非字符串正文报错。""" # 数据:空字符串与数字正文。 # 测试过程与验证结果 with pytest.raises(ValueError, match="empty or invalid"): _parse_translations(json.dumps([{"id": 1, "text": " "}]), {1}) with pytest.raises(ValueError, match="empty or invalid"): _parse_translations(json.dumps([{"id": 1, "text": 3}]), {1}) def test_parse_translations_rejects_unexpected_id() -> None: """返回了未请求的 ID 时报错。""" # 数据:包含 ID 9(未请求)。 content = json.dumps([{"id": 9, "text": "九"}]) # 测试过程与验证结果 with pytest.raises(ValueError, match="invalid or duplicate"): _parse_translations(content, {1}) def test_parse_translations_rejects_non_array_payload() -> None: """顶层不是数组时报错。""" # 数据:对象形式的返回。 # 测试过程与验证结果 with pytest.raises(ValueError, match="JSON array"): _parse_translations(json.dumps({"id": 1, "text": "一"}), {1}) def test_parse_translations_keeps_multiline_text_structure() -> None: """译文多行结构保留(去除纯空行,不截断 cue)。""" # 数据:含空行的多行译文。 content = json.dumps([{"id": 1, "text": "第一行\n\n第二行"}]) # 测试过程 parsed = _parse_translations(content, {1}) # 验证结果:空行被去掉但两行都保留。 assert parsed[1] == "第一行\n第二行" # --------------------------------------------------------------------------- # translate_lines:请求体、ID 与重试 # --------------------------------------------------------------------------- def test_translate_lines_sends_global_ids_and_maps_back(monkeypatch) -> None: """按全局位置 ID 请求翻译,并把结果按位置回填(空 cue 不请求但占位)。""" # 数据:5 行,其中第 3 行为空(占位)。 lines = ["一", "二", "", "四", "五"] calls: list[dict] = [] monkeypatch.setenv("LLM_API_KEY", "sk-test") monkeypatch.setattr( urllib.request, "urlopen", _capture_urlopen(calls, [_llm_reply([(1, "one"), (2, "two"), (4, "four"), (5, "five")])]), ) # 测试过程 result = translate_lines(lines, {"target_language": "en"}) # 验证结果:请求体只含非空行的全局 ID(1,2,4,5),结果按位置回填且空行保留。 sent = json.loads(calls[0]["body"]["messages"][1]["content"]) assert [item["id"] for item in sent] == [1, 2, 4, 5] assert result == ["one", "two", "", "four", "five"] def test_translate_lines_retries_on_structure_error(monkeypatch) -> None: """结构校验失败时重试,最终成功(最多 MAX_BATCH_RETRIES 次)。""" # 数据:第一次返回缺 ID,第二次正确。 lines = ["一", "二"] calls: list[dict] = [] monkeypatch.setenv("LLM_API_KEY", "sk-test") monkeypatch.setattr( urllib.request, "urlopen", _capture_urlopen(calls, [ _llm_reply([(1, "one")]), # 缺 ID 2 → 触发重试 _llm_reply([(1, "one"), (2, "two")]), # 正确 ]), ) # 测试过程 result = translate_lines(lines, {}) # 验证结果:重试一次后成功,共发起 2 次请求。 assert result == ["one", "two"] assert len(calls) == 2 def test_translate_lines_raises_after_retries_exhausted(monkeypatch) -> None: """结构错误耗尽重试后抛错(不返回错位译文)。""" # 数据:每次都返回错误结构。 calls: list[dict] = [] monkeypatch.setenv("LLM_API_KEY", "sk-test") monkeypatch.setattr( urllib.request, "urlopen", _capture_urlopen(calls, [_llm_reply([]) for _ in range(MAX_BATCH_RETRIES)]), ) # 测试过程与验证结果 with pytest.raises(ValueError, match="alignment failed"): translate_lines(["一"], {}) assert len(calls) == MAX_BATCH_RETRIES def test_translate_lines_empty_input_makes_no_request(monkeypatch) -> None: """空输入直接返回空列表,不发请求。""" # 数据:空列表。 calls: list[dict] = [] monkeypatch.setattr(urllib.request, "urlopen", _capture_urlopen(calls, [])) # 测试过程与验证结果 assert translate_lines([], {}) == [] assert calls == [] def test_translate_lines_all_empty_cues_make_no_request(monkeypatch) -> None: """全部为空 cue 时不请求模型,返回等长空列表。""" # 数据:3 个空行。 calls: list[dict] = [] monkeypatch.setattr(urllib.request, "urlopen", _capture_urlopen(calls, [])) # 测试过程与验证结果 assert translate_lines(["", " ", ""], {}) == ["", "", ""] assert calls == [] def test_translate_lines_batches_by_chunk_size(monkeypatch) -> None: """超过 CHUNK_SIZE 行时分批请求(每批最多 CHUNK_SIZE 条)。""" # 数据:CHUNK_SIZE + 1 行。 total = CHUNK_SIZE + 1 lines = [f"行{i}" for i in range(1, total + 1)] calls: list[dict] = [] def fake_urlopen(http_request, timeout=None): body = json.loads(http_request.data.decode("utf-8")) calls.append(body) items = json.loads(body["messages"][1]["content"]) return _llm_reply([(item["id"], f"t{item['id']}") for item in items]) monkeypatch.setenv("LLM_API_KEY", "sk-test") monkeypatch.setattr(urllib.request, "urlopen", fake_urlopen) # 测试过程 result = translate_lines(lines, {}) # 验证结果:两批(CHUNK_SIZE + 1),结果等长且顺序正确。 assert len(calls) == 2 assert len(result) == total assert result[0] == "t1" and result[-1] == f"t{total}" def test_translate_lines_injects_proper_noun_rule(monkeypatch) -> None: """本批原文命中专名时,系统提示词追加规则(不被硬译)。""" # 数据:含专名 ジンゴ 的一行。 calls: list[dict] = [] monkeypatch.setenv("LLM_API_KEY", "sk-test") monkeypatch.setattr( urllib.request, "urlopen", _capture_urlopen(calls, [_llm_reply([(1, "译文")])]), ) # 测试过程 translate_lines(["ジンゴがここで味わいませんか"], {}) # 验证结果:系统提示词包含该专名与禁止硬译的说明。 system_prompt = calls[0]["body"]["messages"][0]["content"] assert "ジンゴ" in system_prompt assert "芒果" in system_prompt def test_translate_lines_default_model_and_env_override(monkeypatch) -> None: """默认模型来自 LLM_MODEL 环境变量(数据驱动,不改代码切换模型)。""" # 数据:设置环境变量为自定义模型。 calls: list[dict] = [] monkeypatch.setenv("LLM_API_KEY", "sk-test") monkeypatch.setenv("LLM_MODEL", "自定义/模型") monkeypatch.setattr( urllib.request, "urlopen", _capture_urlopen(calls, [_llm_reply([(1, "x")])]), ) # 测试过程 translate_lines(["一"], {}) # 验证结果:请求体使用环境变量指定的模型。 assert calls[0]["body"]["model"] == "自定义/模型" def test_translate_lines_param_model_wins_over_env(monkeypatch) -> None: """节点参数 model 优先于环境变量。""" # 数据:环境变量与参数都设置。 calls: list[dict] = [] monkeypatch.setenv("LLM_API_KEY", "sk-test") monkeypatch.setenv("LLM_MODEL", "env/模型") monkeypatch.setattr( urllib.request, "urlopen", _capture_urlopen(calls, [_llm_reply([(1, "x")])]), ) # 测试过程 translate_lines(["一"], {"model": "param/模型"}) # 验证结果 assert calls[0]["body"]["model"] == "param/模型" def test_translate_lines_uses_timeout_env(monkeypatch) -> None: """LLM_TIMEOUT_SECONDS 决定请求超时(默认 600)。""" # 数据:设置 45 秒。 calls: list[dict] = [] monkeypatch.setenv("LLM_API_KEY", "sk-test") monkeypatch.setenv("LLM_TIMEOUT_SECONDS", "45") monkeypatch.setattr( urllib.request, "urlopen", _capture_urlopen(calls, [_llm_reply([(1, "x")])]), ) # 测试过程 translate_lines(["一"], {}) # 验证结果 assert calls[0]["timeout"] == 45.0 def test_translate_lines_sends_bearer_key(monkeypatch) -> None: """请求头带 Bearer Key(来自 LLM_API_KEY)。""" # 数据:设置 key。 calls: list[dict] = [] monkeypatch.setenv("LLM_API_KEY", "sk-abc") monkeypatch.setattr( urllib.request, "urlopen", _capture_urlopen(calls, [_llm_reply([(1, "x")])]), ) # 测试过程 translate_lines(["一"], {}) # 验证结果 headers = {k.lower(): v for k, v in calls[0]["headers"].items()} assert headers.get("authorization") == "Bearer sk-abc" # --------------------------------------------------------------------------- # invoke 全流程 # --------------------------------------------------------------------------- def test_invoke_translates_srt_and_writes_artifact(monkeypatch, tmp_path: Path) -> None: """invoke 解析 SRT → 翻译 → 写出 cn_srt,时间轴保持原样。""" # 数据:两条真实 SRT。 srt = "1\n00:00:01,000 --> 00:00:02,000\nこんにちは\n\n2\n00:00:03,000 --> 00:00:04,000\nさようなら\n" srt_path = tmp_path / "in.srt" srt_path.write_text(srt, encoding="utf-8") calls: list[dict] = [] monkeypatch.setenv("LLM_API_KEY", "sk-test") monkeypatch.setattr( urllib.request, "urlopen", _capture_urlopen(calls, [_llm_reply([(1, "你好"), (2, "再见")])]), ) request = InvokeRequest( run_id="r", node_instance_id="n", params={}, inputs={"srt_uri": str(srt_path)}, output_dir=str(tmp_path / "out"), ) # 测试过程 response = invoke(request) # 验证结果:状态、产物时间轴与译文顺序。 assert response.status == "completed", response.error content = Path(response.outputs["cn_srt_uri"]).read_text(encoding="utf-8") assert "00:00:01,000 --> 00:00:02,000" in content assert "你好" in content and "再见" in content assert content.index("你好") < content.index("再见") def test_invoke_fails_without_input(tmp_path: Path) -> None: """缺少 srt_uri 时失败。""" # 数据:空输入。 request = InvokeRequest( run_id="r", node_instance_id="n", params={}, inputs={}, output_dir=str(tmp_path) ) # 测试过程 response = invoke(request) # 验证结果 assert response.status == "failed" assert "srt_uri" in (response.error or "") def test_invoke_fails_when_input_missing(tmp_path: Path) -> None: """输入文件不存在时失败。""" # 数据:不存在的路径。 request = InvokeRequest( run_id="r", node_instance_id="n", params={}, inputs={"srt_uri": str(tmp_path / "nope.srt")}, output_dir=str(tmp_path), ) # 测试过程 response = invoke(request) # 验证结果 assert response.status == "failed" assert "not found" in (response.error or "") def test_invoke_reports_failure_on_llm_error(monkeypatch, tmp_path: Path) -> None: """LLM 报错时节点返回 failed(不产出半成品产物)。""" # 数据:urlopen 抛 HTTPError。 srt_path = tmp_path / "in.srt" srt_path.write_text("1\n00:00:01,000 --> 00:00:02,000\nこんにちは\n", encoding="utf-8") monkeypatch.setenv("LLM_API_KEY", "sk-test") def fake_urlopen(http_request, timeout=None): raise urllib.error.HTTPError(http_request.full_url, 500, "server error", {}, None) monkeypatch.setattr(urllib.request, "urlopen", fake_urlopen) request = InvokeRequest( run_id="r", node_instance_id="n", params={}, inputs={"srt_uri": str(srt_path)}, output_dir=str(tmp_path / "out"), ) # 测试过程 response = invoke(request) # 验证结果 assert response.status == "failed" # --------------------------------------------------------------------------- # 真实 LLM 集成(需要 LLM_API_KEY 与网络) # --------------------------------------------------------------------------- @pytest.mark.integration def test_real_llm_translates_short_lines_with_domain_rules() -> None: """真实 LLM 校准:短句专名/隐语不按字面直译(真实 API,外部状态缺失则跳过)。""" # 数据:含专名 ジンゴ 的短句;Key 与账号可用性由共享判定处理。 from nodes.llm import translate_lines as real_translate from tests.shared.llm_service import require_llm_credentials, skip_on_service_unavailable require_llm_credentials() # 测试过程 with skip_on_service_unavailable(): out = real_translate(["ジンゴがここで味わいませんか"], {"target_language": "zh-CN"}) # 验证结果:有译文且未把专名硬译成"芒果"。 assert out and out[0].strip() assert "芒果" not in out[0]