升级 llm-compat 到 v0.10.0：根治总结/校对被拒绝检测误判

pyproject.toml

@@ -79,4 +79,4 @@ exclude_lines = [
 ]
 
 [tool.uv.sources]
-llm-compat = { git = "https://github.com/zlxlabs/llm-compat.git", tag = "v0.8.0" }
+llm-compat = { git = "https://github.com/zlxlabs/llm-compat.git", tag = "v0.10.0" }

src/video_transcript_api/llm/llm.py

@@ -19,7 +19,6 @@
 from llm_compat.providers import (
     build_request_payload,
     describe_from_payload,
-    detect_provider,
 )
 
 # 注意：不能在模块顶层 `from .core.usage_context import ...`——
@@ -163,6 +162,7 @@ def set_default_config(config: Optional[Dict[str, Any]]) -> None:
         collector_api_key=llm_cfg.get("collector_api_key", ""),
         refusal_keywords_url=llm_cfg.get("refusal_keywords_url"),
         sensitive_detector=sensitive_detector,
+        on_all_refused="raise",
     )
     if sensitive_detector:
         logger.info(f"[LLM] SyncLLMClient initialized with SensitiveDetector ({len(sensitive_detector._words)} words)")

tests/unit/test_refusal_contract.py

@@ -0,0 +1,181 @@
+"""Cross-library refusal detection contract tests (llm-compat v0.10.0).
+
+Uses real llm_compat.refusal.detect_refusal — the library itself is never mocked.
+Wire-format response dicts mirror what SyncLLMClient receives from providers.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+from unittest.mock import patch
+
+import pytest
+
+from llm_compat import ContentPolicyError
+from llm_compat.refusal import detect_refusal
+
+from video_transcript_api.llm.core.errors import RetryableError, map_llm_compat_error
+
+
+def _is_refusal(data: dict[str, Any]) -> bool:
+    """Normalize detect_refusal return across llm-compat API versions."""
+    result = detect_refusal(data)
+    if isinstance(result, bool):
+        return result
+    return result.is_refusal
+
+
+def _make_response(content: str | None, finish_reason: str = "stop") -> dict[str, Any]:
+    return {
+        "choices": [
+            {
+                "message": {"content": content},
+                "finish_reason": finish_reason,
+            }
+        ]
+    }
+
+
+# >300 chars, legal/medical tone with repeated 违反 (production false-positive pattern)
+LONG_LEGAL_SUMMARY = (
+    "这是一段关于法律题材的长文总结，涵盖医疗合规与侵权责任。"
+    + "该行为可能违反相关法律法规，涉及敏感话题的讨论需要谨慎对待。"
+    * 15
+)
+
+# >300 chars, embed every legacy substring keyword from pre-v0.10.0 table
+_LEGACY_KEYWORDS = ["无法提供", "涉及敏感", "不适合讨论", "违反"]
+LONG_ALL_LEGACY_KEYWORDS = (
+    "法律医疗题材总结。"
+    + "。".join(
+        f"段落{i}讨论{k}相关问题"
+        for i in range(30)
+        for k in _LEGACY_KEYWORDS
+    )
+)
+
+
+REFUSAL_CONTRACT_CASES: list[tuple[str, str | None, str, bool, str]] = [
+    (
+        "long_with_wei_fan",
+        LONG_LEGAL_SUMMARY,
+        "stop",
+        False,
+        "long legal/medical summary containing wei_fan",
+    ),
+    (
+        "long_all_legacy_keywords",
+        LONG_ALL_LEGACY_KEYWORDS,
+        "stop",
+        False,
+        "long text with all legacy keyword substrings",
+    ),
+    (
+        "short_cn_refusal",
+        "抱歉，我无法提供这方面的帮助。",
+        "stop",
+        True,
+        "short CN first-person refusal at sentence start",
+    ),
+    (
+        "short_en_refusal",
+        "I'm sorry, but I cannot assist with that request.",
+        "stop",
+        True,
+        "short EN first-person refusal",
+    ),
+    (
+        "pseudo_refusal_with_turn",
+        "我不能协助这个请求，但可以换个角度说明。",
+        "stop",
+        False,
+        "pseudo refusal with concessive turn (but/however)",
+    ),
+    (
+        "content_filter",
+        "ok",
+        "content_filter",
+        True,
+        "provider finish_reason content_filter",
+    ),
+    (
+        "malformed_none_content",
+        None,
+        "stop",
+        True,
+        "None content with finish_reason stop",
+    ),
+]
+
+
+@pytest.mark.parametrize(
+    "case_id,content,finish_reason,expected,detail",
+    REFUSAL_CONTRACT_CASES,
+    ids=[c[0] for c in REFUSAL_CONTRACT_CASES],
+)
+def test_detect_refusal_contract(
+    case_id: str,
+    content: str | None,
+    finish_reason: str,
+    expected: bool,
+    detail: str,
+) -> None:
+    """Table-driven contract: real detect_refusal on wire-format response dicts."""
+    if case_id.startswith("long"):
+        text = content if isinstance(content, str) else ""
+        assert len(text) > 300, f"case {case_id} requires >300 chars, got {len(text)}"
+
+    data = _make_response(content, finish_reason)
+    actual = _is_refusal(data)
+    assert actual == expected, f"case {case_id}: {detail}"
+
+
+class TestSyncLLMClientRefusalPolicy:
+    """Construction-time on_all_refused must stay raise (not v0.10.0 default)."""
+
+    @pytest.fixture(autouse=True)
+    def _restore_llm_module_globals(self):
+        """Restore module-level globals mutated by set_default_config().
+
+        set_default_config() writes _default_config and _sync_client. Without
+        this fixture the mock client built here leaks into any later test that
+        reads the module default instead of setting its own, making results
+        depend on collection order.
+        """
+        from video_transcript_api.llm import llm as llm_mod
+
+        saved_config = llm_mod._default_config
+        saved_client = llm_mod._sync_client
+        try:
+            yield
+        finally:
+            llm_mod._default_config = saved_config
+            llm_mod._sync_client = saved_client
+
+    @patch("video_transcript_api.llm.llm.SyncLLMClient")
+    def test_on_all_refused_raise_passed_at_construction(self, mock_client_cls) -> None:
+        from video_transcript_api.llm.llm import set_default_config
+
+        config = {
+            "llm": {
+                "api_key": "test-key",
+                "base_url": "https://api.test.com/v1",
+            },
+        }
+        set_default_config(config)
+        call_kwargs = mock_client_cls.call_args[1]
+        assert call_kwargs.get("on_all_refused") == "raise"
+
+
+class TestContentPolicyErrorChain:
+    """All-refused terminal path: ContentPolicyError -> map_llm_compat_error -> RetryableError."""
+
+    def test_content_policy_error_maps_to_retryable(self) -> None:
+        err = ContentPolicyError(
+            "All models refused",
+            attempted_models=["deepseek-v4", "gemini-3-flash"],
+            raw_content="I cannot assist",
+            original_model="deepseek-v4",
+        )
+        result = map_llm_compat_error(err)
+        assert isinstance(result, RetryableError)