* Shows the autocomplete suggestion in a popup below the cursor.

Truncate text to approximately fit within max_tokens. Uses a simple character-to-token ratio estimation (roughly 4 characters per token).

Return absolute amplitudes and a threshold scaled to the input dtype.

Evict a specific model.

Best-effort extraction of segment timing data from NeMo ASR output.

Generate audio from pre-tokenized input with Kokoro chunking.

Get voicepack, falling back to default.

Bridge from legacy generate API to Qwen3.5 chat inference.

Check if we need to evict models based on memory pressure.

Simple fallback quality scoring based on text similarity.

Test if output matches baseline.

Warmup the model with a dummy inference to optimize performance.

Perform full analysis and return results.