Suggesting A Pathway

Search normally returns a list. A pathway needs a walk, so retrieval is only the first of four moves.

Retrieve — BM25 over a compact per-site index, blended with semantic similarity when the query can be embedded. Document frequencies are merged across every site being searched, or a term common in a big index site would score incomparably against a small one.

Expand — one hop along the link graph the wiki already carries, so a walk does not dead-end on a reference it needs. A link is an invitation, not a licence: a linked page comes along only if it is strongly relevant in its own right, at most two per seed. Otherwise a hub page’s link list quietly doubles what the grant opens.

Order — relevance sets the spine; links fill it in. Each seed in score order, then the pages it linked to. That reads like a curriculum rather than a random walk that strands a strong page behind a weak one.

Name — the package is named, by the searcher or from the query.

The floor and the upgrade

The lexical index is the floor: zero-dependency, deterministic, offline, and small enough to bake into a static web page so a browser can suggest pathways with no server at all. Embeddings are an upgrade layered on top, never a dependency. A system whose basic function requires a model download is a system that does not work.