New: - vllm_tuning_pipeline.py: A/B benchmark different vLLM configs, logs latency/TPS/TTFT to MLflow (vllm-tuning experiment) - vllm_tuning_pipeline.yaml: compiled KFP YAML Updated: - voice_pipeline.py: per-step NamedTuple outputs with latency tracking, new log_pipeline_metrics MLflow component - voice_pipeline.yaml, tts_pipeline.yaml, rag_pipeline.yaml: recompiled
- Add create_mlflow_run and log_evaluation_to_mlflow KFP components - Log accuracy, correct/total counts, pass/fail to MLflow experiment - Upload evaluation_results.json as artifact - Wire MLflow run into pipeline DAG before NATS publish
- voice_pipeline: STT → RAG → LLM → TTS - document_ingestion_pipeline: Extract → Chunk → Embed → Milvus - document_ingestion_mlflow_pipeline: With MLflow tracking - evaluation_pipeline: Model benchmarking - kfp-sync-job: K8s job to sync pipelines
