Deployment, Versioning, and Resilience
Version parsers, chunkers, embedding models, index configuration, prompts, and answer models. Changes may require reindexing and evaluation. Use deadlines, cancellation, retries, circuit breakers, and bounded concurrency. Keep known-good configurations and rollback plans. Graceful degradation may return safe search results without generation if the model is unavailable.
Implementation example
{
"pipelineVersion":"rag-v5",
"parserVersion":"parser-v3",
"chunkerVersion":"semantic-v2",
"embeddingModel":"embedding-v2",
"indexVersion":"kb-green",
"promptVersion":"grounded-v4"
}💡 Version manifests make behavior reproducible and help diagnose regressions.
Key Takeaways
- Inspect the relevant pipeline stage instead of guessing.
- Preserve source provenance, versions, and access scope.
- Measure quality on representative queries.
Knowledge Check
Why version the whole pipeline?