Module 2 · Running ingest in production · Lesson 6
Re-ingestion, versioning and backfills
How to change parsers, prompts, schemas and models without re-running everything or losing history - content hashing, version keys per stage, early cutoff, document lineage, diff-before-promote, and bulk backfills through batch APIs.
By the end of this lesson you can
- Model ingest as a build graph whose artefacts are keyed by input hash and stage version, and compute what a change invalidates
- Apply early cutoff so unchanged intermediate outputs stop a rebuild from propagating
- Version documents in the business sense (supersession, amendment) separately from pipeline versions, and keep what each agent saw
- Plan a re-ingestion with an impact estimate, a shadow run and a field-level diff before promotion
- Run a backfill through a batch API with stable request IDs, resumability and isolation from live traffic