YmerNode.References.Search (Ymer Node v0.2.1)

Copy Markdown View Source

The find surface — two composable layers on one call.

  1. Filters — tag-AND and derived source, fetched through YmerNode.References.list_classified/1 rather than from the repo directly, so those rules stay implemented once. The resulting cycle between the two modules is runtime-only: a defdelegate out and a plain call back.
  2. Keyword floor — a score of distinct query terms matched against YmerNode.References.Reference.search_text/1, which is every text field, so a term carried only by a uri still matches.

mode names what ran: :keyword when a query was given, :filter when none was — filters only, newest first. There are two modes and not three. Ranking by embedding is a knowing omission rather than an unimplemented branch: the node ships no embedding provider for the registry, and a mode that silently fell back would tell a caller its query had been ranked one way when it had been ranked the other.

Ties break newest-updated first, then by id descending, so an equal-scoring pair has a stable order rather than the repo's.

Summary

Functions

run(opts \\ [])

Runs a find. Options: :query (a string, or nil for filter mode), :tags (tag-AND list), :source (a source token), :limit (a positive integer, or nil for all).

tokenize(query)

Splits a query into the distinct terms a keyword score counts.

Functions

run(opts \\ [])

Runs a find. Options: :query (a string, or nil for filter mode), :tags (tag-AND list), :source (a source token), :limit (a positive integer, or nil for all).

Answers %{mode: :keyword | :filter, results: results}, where each result is %{reference: %YmerNode.References.Reference{}, classification: %{source: source, recipe: recipe}, score: number | nil} — the score is nil in filter mode, because nothing ranked.

The classification rides along rather than being derived by whoever renders the hit: YmerNode.References.list_classified/1 had to compute it to apply the source filter, and passing :declarations through means one operation reads the seam once. :declarations is optional here for the same reason it is there — a caller with none in hand gets one read rather than a crash.

tokenize(query)

Splits a query into the distinct terms a keyword score counts.

Downcased, split on anything that is not a letter, a number or a hyphen — so an identifier like ABC-1234 survives as one term — then deduplicated and capped.

Examples

iex> YmerNode.References.Search.tokenize("Release notes, ABC-1234")
["release", "notes", "abc-1234"]

iex> YmerNode.References.Search.tokenize("a I x")
[]

iex> YmerNode.References.Search.tokenize("dup DUP dup")
["dup"]