Open source video conferencing app powered by LiveKit. Built with Django and React.
🔒️(summary) sanitize LLM failures and summary progress logs
Florent Chehab3cd2cb5591bf
diff --git a/src/summary/summary/core/celery_worker.py b/src/summary/summary/core/celery_worker.pyindex c1ee5ce2..c40562c3 100644--- a/src/summary/summary/core/celery_worker.py+++ b/src/summary/summary/core/celery_worker.py@@ -311,6 +311,17 @@ def format_actions(llm_output: dict) -> str:return ""+def parse_llm_json(value: str) -> dict:+ """Parse an LLM response without exposing it in decoding errors."""+ try:+ result = json.loads(value)+ if not isinstance(result, dict):+ raise ValueError+ return result+ except (TypeError, ValueError):+ raise LLMException("Invalid LLM response") from None++def summarize_transcription_internals(*, distinct_id: str, transcript: str, session_id: str) -> str:@@ -351,13 +362,13 @@ def summarize_transcription_internals()logger.info("Plan generated")- res = json.loads(parts)+ res = parse_llm_json(parts)parts = res.get("titles", [])- logger.info("Parts to summarize: %s", parts)+ logger.info("Parts to summarize: %s", len(parts))parts_summarized = []- for part in parts:+ for i, part in enumerate(parts):prompt_user_part = PROMPT_USER_PART.format(part=part, transcript=transcript)- logger.info("Summarizing part: %s", part)+ logger.info("Summarizing part %s", i + 1)parts_summarized.append(llm_service.call(PROMPT_SYSTEM_PART, prompt_user_part, name="part"))@@ -373,7 +384,7 @@ def summarize_transcription_internals(response_format=FORMAT_NEXT_STEPS,)- next_steps = format_actions(json.loads(next_steps))+ next_steps = format_actions(parse_llm_json(next_steps))logger.info("Next steps generated")diff --git a/src/summary/summary/core/llm_service.py b/src/summary/summary/core/llm_service.pyindex e8b52b0c..08c3822b 100644--- a/src/summary/summary/core/llm_service.py+++ b/src/summary/summary/core/llm_service.py@@ -140,6 +140,6 @@ class LLMService:response = self._client.chat.completions.create(**params)return response.choices[0].message.content- except Exception as e:- logger.exception("LLM call failed: %s", e)- raise LLMException(f"LLM call failed: {e}") from e+ except Exception:+ logger.error("LLM call failed")+ raise LLMException("LLM call failed") from Nonediff --git a/src/summary/tests/unit/test_privacy.py b/src/summary/tests/unit/test_privacy.pyindex 044f4278..8f5a0a0b 100644--- a/src/summary/tests/unit/test_privacy.py+++ b/src/summary/tests/unit/test_privacy.py@@ -6,7 +6,8 @@ from unittest.mock import Mockimport pytestimport requests-from summary.core import analytics, webhook_service+from summary.core import analytics, celery_worker, webhook_service+from summary.core.llm_service import LLMException, LLMServicefrom summary.core.privacy import redact_urlfrom summary.core.shared_models import SummarizeWebhookSuccessPayload@@ -19,6 +20,19 @@ def test_url_redaction():assert redact_url(URL) == "https://example.com/file?[redacted]"+def test_llm_exception_does_not_include_upstream_message(caplog):+ """LLM failures retain their category without logging upstream text."""+ observability = Mock(is_enabled=False)+ observability.get_openai_client.return_value.chat.completions.create.side_effect = (+ ValueError(CANARY)+ )+ with caplog.at_level(logging.DEBUG), pytest.raises(LLMException) as exc:+ LLMService(observability).call("system", CANARY, "test")+ assert CANARY not in str(exc.value)+ assert CANARY not in caplog.text+ assert exc.value.__suppress_context__++@pytest.mark.parametrize("log_level", [logging.DEBUG, logging.INFO])def test_webhook_content_is_only_logged_at_debug(monkeypatch, caplog, settings_override, log_level@@ -88,6 +102,14 @@ def test_analytics_source_url_removes_signature():assert CANARY not in str(metadata)+def test_invalid_llm_json_is_sanitized():+ """Invalid structured LLM output retains a retryable LLM category."""+ with pytest.raises(LLMException) as exc:+ celery_worker.parse_llm_json(CANARY)+ assert CANARY not in str(exc.value)+ assert exc.value.__suppress_context__++@pytest.mark.parametrize(("url", "expected"),[