A document import should not be considered healthy merely because the connector run completed. Track source discovery through extraction, indexing, and query eligibility so missing or broken content is visible before users depend on it.
Import health needs stage evidence
A failed import can occur during enumeration, download, parsing, metadata mapping, permission capture, embedding, indexing, or activation. Each failure can leave a different partial state.
For a proposed import record, store stable source identity, version, stage, safe error category, expected derived artifacts, and current owner. Separate successful fetch from successful searchable representation.
Do not infer content quality from a nonempty text field. A scanned document may yield only a header, or a table parser may drop the column labels that make values meaningful. Quality checks need representative fixtures.
Reconstruct a fictional silent import gap
A connector discovers 100 approved documents. Ninety-eight reach the active index, one fails parsing, and one lacks permission metadata. A green top-level job obscures two resources that the assistant should not treat as fully available.
The dashboard should show the stage counts and unresolved resources. The permission-metadata failure follows the safe access policy rather than becoming public by default. These counts illustrate monitoring, not actual product results.
| Import stage | Detection evidence |
|---|---|
| Discovery | Expected resource and source version |
| Fetch | Download result and supported completeness |
| Extraction | Required text or structural checks |
| Metadata | Audience, authority, and access fields |
| Index activation | Resource queryable under expected scope |
Use synthetic documents with known markers, tables, and permissions. Add a targeted probe query for important content, but do not treat one matching phrase as proof that the entire document imported faithfully.
Alert according to consequence
A failed low-priority draft may belong in routine review, while a missing current emergency procedure may need faster response under the business policy. Route the item to the owner who can repair the actual stage.
More health checks add cost and false positives. Keep them tied to concrete risks and inspect both true failures and ordinary successful imports. A document count alone can remain stable while one source silently replaces another or loses important pages.
This monitoring design is proposed. The voice integration case study describes related retrieval work. Pair it with connector acceptance checks.
Import failure FAQ
Does a successful fetch prove search readiness?
No. Extraction, metadata, permissions, indexing, and activation still need evidence.
Should every import error page someone?
Use impact, urgency, ownership, and the supported response policy.
What should the repair item include?
Provide safe resource and version references, failed stage, known partial state, and the next permitted action.