Abstract provider class

Concrete providers

Tenacity retry decorator

Retry parameters

Fallback implementation

tiktoken token counting

Cost calculation

Correct pricing values

Pydantic response model

Response validation used

Response caching

Cost summary output

python:3.11-slim base image

Health check endpoint

Uvicorn CMD

Port 8080 exposed

Model export step

Canary at 5%

1 hour canary window

p95 latency threshold

Error rate threshold

K8s memory limits

K8s CPU limits

K8s readiness probe

RecursiveCharacterTextSplitter

Chunking separators

Embedding cache using hash

Batch embedding support

BM25 sparse retrieval

Hybrid score combination

Alpha parameter

Reranking step

Reranker sorts results

Query function integration

Demo runs without errors

KS test for drift

KS output fields

KS drift flag threshold

p95 latency warning

p95 latency critical

Error rate warning

Error rate critical

PSI warning threshold

PSI critical threshold

Accuracy drop thresholds

Retraining at PSI>0.2

Demo output

Feast Entity definition

Feast FeatureView

FileSource in FeatureView

TTL set on FeatureView

mlflow.start_run used

mlflow.log_metric called

Model logged to MLflow

mlflow.register_model called

PSI drift trigger

Scheduled retrain trigger

Performance drop trigger

Pipeline runs without error

Hash-based assignment

Experiment key in hash input

Bucket conversion

Sticky assignment

Configurable control_pct

Returns control/treatment

Primary metric collection

Guardrail metric tracking

p-value < 0.05 threshold

Minimum sample size check

Demo runs without errors