Operator Runbooks

Investigate and retry failed ingestion

Diagnose a stuck or failed document and recover it.

When a document doesn't become searchable, use the Ingestion Health panel (/operator/health) to diagnose and recover it.

Diagnose

Open Command Center → Knowledge → Ingestion Health and review the four counters.
Find the document in the failed jobs table; note its organization, title, and the recorded error.
Determine whether the job is failed (retries exhausted) or dead_letter (unrecoverable).

Common causes

SymptomLikely causeAction
Fails at parsingUnsupported file type or a corrupt fileRe-upload a supported format (PDF/DOCX/XLSX).
Fails at embeddingEmbedding provider unavailable, or an API key issueResolve the provider/config, then retry.
Embedding-dimension mismatchEMBEDDING_DIMENSIONS doesn't match the indexed corpusA configuration/migration fix; a plain retry won't help.
dead_letterDocument deleted mid-flight or invalid metadataRe-upload the document fresh.

Retry

In the failed-jobs table, use Retry on the affected job and confirm.
The job re-enters the pipeline; the health counters and list refresh.
Confirm the document reaches indexed.

Retries use exponential back-off, so a transient provider issue often clears on its own. A retry only helps once the underlying cause is resolved. For a dimension mismatch or an unsupported file, fix the root cause first. See failure handling.

GP-scoped view

GP users can see and retry their own organization's failed jobs from Settings → Knowledge. Operators see failures across all tenants.

On this page