ard search is intentionally simple for the first public release. The implementation
prioritizes deterministic, inspectable behavior over opaque ranking.
POST /search,ard search, andpkg/client.Searchuse the same registry search path.- Query text is split on whitespace and matched case-insensitively against normalized catalog entry text.
- The indexed text includes identifier, publisher, display name, media type, description, tags, capabilities, and representative queries.
- Query terms are recall-oriented: an entry can match when any term is present.
- Request filters still apply after recall, and inactive lifecycle states are excluded from public search.
score is semantic relevance only. It is not a trust, safety, governance, policy, or
compliance signal.
The first-release score is a deterministic approximation:
- Empty query text scores matching entries as
50. - A query with no matched terms scores
0. - Any matched term starts at
50. - Additional matched query terms increase the score linearly up to
100.
Local search results are ordered by:
scoredescending.identifierascending.displayNameascending.sourceascending.
This keeps page-token behavior stable and makes local search compatible with
score-ranked auto federation. Future ranking improvements should preserve the contract
that score remains relevance-only and should keep trust and policy signals separate.