Record glossary provenance without transcript mutation

2026-09-12 11:40:57 +02:00
parent 9177a2660b
commit 92247ef43e
7 changed files with 160 additions and 0 deletions
@@ -9,6 +9,8 @@ from unittest.mock import patch

 from scripts import run_mvp_meeting as cli
 from src.meeting_lab.audio import PreparedAudio
+from src.meeting_lab.llm.ollama import OllamaGeneration
+from src.meeting_lab.protocol.generate_direct_protocol import generate_direct_protocol
 from src.meeting_lab.models.meeting_context import load_meeting_context
 from src.meeting_lab.orchestration import mvp as mvp_api
 from src.meeting_lab.orchestration.mvp import MvpMeetingConfig, MvpRunResult
@@ -202,6 +204,85 @@ class MvpApiTests(unittest.TestCase):
            persisted = load_meeting_context(run_dir / "context/meeting_context.yaml")
            self.assertEqual(persisted.speaker_mappings, {"SPEAKER_00": "person-1"})

+    def test_regeneration_keeps_glossary_out_of_transcript_and_in_context(self):
+        with tempfile.TemporaryDirectory() as directory:
+            run_dir = Path(directory)
+            diarization_dir = run_dir / "diarization"
+            diarization_dir.mkdir()
+            source = diarization_dir / "transcript_diarized.json"
+            source.write_text(
+                json.dumps(
+                    {
+                        "text": "SPEAKER_00: Lumini.",
+                        "segments": [
+                            {
+                                "start": 0.0,
+                                "end": 1.0,
+                                "speaker_id": "SPEAKER_00",
+                                "text": "Lumini.",
+                            }
+                        ],
+                        "speaker_labels_anonymous": True,
+                    }
+                ),
+                encoding="utf-8",
+            )
+            source_before = source.read_bytes()
+            transcript_dir = run_dir / "transcript"
+            transcript_dir.mkdir()
+            whisper_source = transcript_dir / "transcript.json"
+            whisper_source.write_text(
+                json.dumps({"text": "Lumini.", "segments": []}), encoding="utf-8"
+            )
+            whisper_source_before = whisper_source.read_bytes()
+            context = context_data()
+            context["known_entities"] = {
+                "Authoritative terminology": ["Luminy (aliases: Lumini)"]
+            }
+
+            def generate_without_network(transcript, context_path, **kwargs):
+                return generate_direct_protocol(
+                    transcript,
+                    context_path,
+                    model=kwargs["model"],
+                    num_ctx=kwargs["num_ctx"],
+                    num_thread=kwargs["num_thread"],
+                    safe_input_token_budget=kwargs["safe_input_token_budget"],
+                    glossary_aliases=kwargs["glossary_aliases"],
+                    model_check=lambda *_: {},
+                    generation_call=lambda *_args, **_kwargs: OllamaGeneration(
+                        raw_response={"response": "# Protocol", "done": True},
+                        text="# Protocol",
+                        client_wall_time_seconds=0.1,
+                    ),
+                )
+
+            with (
+                patch.object(
+                    mvp_api,
+                    "generate_direct_protocol",
+                    side_effect=generate_without_network,
+                ),
+            ):
+                mvp_api.regenerate_mvp_protocol(
+                    run_dir,
+                    meeting_context=context,
+                    glossary_aliases={"Lumini": "Luminy"},
+                )
+
+            generation = run_dir / "protocol"
+            transcript_input = (generation / "transcript_input.txt").read_text()
+            exact_prompt = (generation / "exact_prompt.txt").read_text()
+            metadata = json.loads((generation / "runtime_metadata.json").read_text())
+            self.assertIn("SPEAKER_00: Lumini.", transcript_input)
+            self.assertNotIn("SPEAKER_00: Luminy.", transcript_input)
+            self.assertIn("Luminy (aliases: Lumini)", exact_prompt)
+            self.assertIn("SPEAKER_00", exact_prompt)
+            self.assertIn("Test Person", exact_prompt)
+            self.assertEqual(metadata["glossary_replacements"], [])
+            self.assertEqual(source.read_bytes(), source_before)
+            self.assertEqual(whisper_source.read_bytes(), whisper_source_before)
+
    def test_failure_emits_terminal_failure_event(self):
        with tempfile.TemporaryDirectory() as directory:
            root = Path(directory)