Searching several company systems adds identity, freshness, and authority boundaries. A unified result list should preserve where each passage came from and whether the current user may access it.
Cross-system retrieval is a coordinated pipeline
Multi-system retrieval gathers eligible evidence from sources such as document stores, knowledge articles, and approved operational records. It can use a central index, federated queries, or a combination, each with different freshness and availability tradeoffs.
For a proposed connector contract, record source identity, stable resource reference, version, access metadata, ownership, update status, and deletion behaviour. Keep query and result context scoped to the user’s trusted permissions.
Do not assume two systems use the same customer, document, or employee identifiers. Mapping must preserve source namespaces and authorised relationships. A coincidental identifier match should not join unrelated records.
Test a fictional policy-and-ticket search
An employee asks how to handle a recurring export issue. The assistant searches approved product guidance and an internal incident library. The user may access only part of the incident collection, so candidate retrieval must respect those boundaries.
The public guide contains the current supported remedy, while an old ticket contains a temporary workaround. Ranking similarity should not make the old workaround authoritative. Source type, status, and approved precedence still matter.
| Connector condition | Expected behaviour |
|---|---|
| Source available and permitted | Eligible current evidence returned |
| One source unavailable | Accurate degraded-scope result |
| Permission revoked upstream | Supported propagation prevents stale access |
| Duplicate document copies | Preserve authority and avoid repeated context |
| Conflicting source facts | Apply policy or surface uncertainty |
Replay fixtures with delayed updates, deleted resources, and one failing connector. Inspect both the returned evidence and what the answer claims about search completeness.
Report coverage limits to users
A partial search should not be described as having checked every company system. Explain material unavailable sources at an appropriate level and offer the supported fallback. Keep source-specific diagnostics accessible to the responsible operator.
Central indexing can simplify ranking but creates synchronisation and permission propagation work. Federation can preserve live source checks while adding latency and partial availability. Evaluate the actual corpus, security guarantees, and operational ownership rather than choosing from architecture labels alone.
This retrieval design is proposed. The voice integration case study establishes related knowledge and backend experience. Pair it with document ownership design.
Cross-system search FAQ
Does one search box imply complete coverage?
No. State the connected sources, supported scope, and current availability limits.
Should source failures silently return fewer results?
Use a defined degraded-service policy and avoid implying a complete search when a material source was unavailable.
What should connectors preserve?
Keep source identity, permissions, version, ownership, deletion state, and the relationships needed for trustworthy evidence.