52f4ba99f8
A model that loops until the output cap returns non-empty text, so the existing empty-content check lets it through and it overwrites the last good summary. Discard a summary only when it both hit the output cap and scores below a distinct-4-gram ratio of 0.35: set is_fallback so the previous summary is retained, zero the token counts on that path, and count the rejection via summary_rejections_counter. |
||
|---|---|---|
| .. | ||
| test_enqueue.py | ||
| test_message_embeddings.py | ||
| test_representation.py | ||
| test_telemetry.py | ||
| test_token_metrics.py | ||