Problem
Existing rows with missing/stale embeddings cannot be backfilled through a managed served operation. The offline export → embed → load pipeline exists, but clients must assemble the job and handle checkpoints, retries and concurrent edits themselves.
Proposed change
Expose an authorized backfill/re-embedding operation for a served graph, with type/property filters, missing-only versus model-refresh mode, a bounded work budget, progress, cancellation and resumable status. Use the configured embedding provider and common derivation logic rather than a second incompatible pipeline.
Pin source identity/revision and reject or reschedule stale results. Define branch selection, publication granularity and how to resume after a lost response or restart. Include a preview/count mode so operators can estimate work before starting a full re-embedding.
Acceptance criteria
- Backfill a mixed set of missing, current, stale and failed rows without rewriting unaffected rows.
- Same-dimensional model changes are still recognized as requiring re-embedding.
- Concurrent source edits/deletion, rate limits, provider failures, cancellation and restart preserve correctness and bounded memory.
- Status exposes attempted/completed/skipped/failed work and model/source coverage.
Related: managed ingestion population is the steady-state trigger; this issue covers operator-initiated historical work. Current offline command and embedding client contract should be reused where appropriate.
Problem
Existing rows with missing/stale embeddings cannot be backfilled through a managed served operation. The offline export → embed → load pipeline exists, but clients must assemble the job and handle checkpoints, retries and concurrent edits themselves.
Proposed change
Expose an authorized backfill/re-embedding operation for a served graph, with type/property filters, missing-only versus model-refresh mode, a bounded work budget, progress, cancellation and resumable status. Use the configured embedding provider and common derivation logic rather than a second incompatible pipeline.
Pin source identity/revision and reject or reschedule stale results. Define branch selection, publication granularity and how to resume after a lost response or restart. Include a preview/count mode so operators can estimate work before starting a full re-embedding.
Acceptance criteria
Related: managed ingestion population is the steady-state trigger; this issue covers operator-initiated historical work. Current offline command and embedding client contract should be reused where appropriate.