Metadata filters should narrow company search using meaningful trusted attributes. Separate relevance filters from security controls so a user-selected label cannot decide which restricted documents the assistant may retrieve.
Filters define different boundaries
A relevance filter selects useful content, such as product, document type, language, or effective period. An access filter enforces who may see a resource. Both can appear in a query, but their authority and failure consequences differ.
For a proposed index, define metadata ownership, allowed values, missing-value behaviour, and update process. Keep tenant and access scope derived from trusted application context. Do not accept arbitrary model-generated security filters as permission evidence.
Version and freshness metadata can assist policy selection, but a date field alone does not establish which document is authoritative. The business needs an explicit source-precedence rule.
Test a fictional product-policy query
An employee asks about product Nova 2 in the current region. The relevance filter should exclude policies for other products and locations where the metadata is reliable. The access boundary separately excludes private operational documents the employee cannot view.
A document missing product metadata should not silently disappear from all evaluation without notice. Decide whether it is quarantined, searchable under a broader scope, or sent for content repair according to policy.
| Metadata condition | Expected treatment |
|---|---|
| Correct product and region | Eligible under relevance rules |
| Restricted document | Excluded by trusted access policy |
| Missing required metadata | Defined visible disposition |
| Stale permission tag | Revocation path prevents unauthorised retrieval |
| Conflicting effective dates | Authority rule or review |
Inspect eligible candidate sets as well as final rankings. A result list can look relevant while a missing filter exposes private content to the answer stage.
Evaluate overfiltering and underfiltering
Strict filters can remove the only supporting passage when metadata is incomplete. Broad filters can introduce distracting or inappropriate evidence. Track false exclusions, irrelevant additions, and access violations separately.
More metadata improves control only if it is maintained. Start with attributes that change important decisions and assign owners for their accuracy. A large unused taxonomy can create ingestion work without making answers more reliable.
This filter design is proposed. The voice integration case study describes related retrieval work without claiming this index configuration. Pair it with permissions before retrieval.
Search metadata FAQ
Can the model choose the tenant filter?
The trusted application should enforce tenant scope independently of the model’s proposed query.
Should missing metadata be treated as public?
Not by default. Apply the explicit approved policy for that attribute and consequence.
What proves filters are useful?
Test relevance gains, missed supporting evidence, freshness, and access boundaries under labelled fixtures.