ADR-00019: Semantic Scholar's similarity capability — placement, not adoption¶
Context¶
Semantic Scholar's /paper/search (and bulk search) is keyword search against an Elasticsearch index with a custom relevance re-ranker — not free-text embedding search, so it isn't a search.semantic peer (see ADR-00015). It does compute SPECTER2 embeddings per paper, and the Recommendations API built on them takes a seed paper ID and returns similar papers — a citation/embedding-graph-flavored retrieval mechanism, which raises a corpus-scope placement question: discovery, or a locally-scoped citation/concept-linking mechanism instead.
Decision¶
If Semantic Scholar's Recommendations API is ever adopted, it belongs alongside research_discovery (a candidate future extension, in the same tier as CORE), not folded into a locally-scoped citation/concept-linking mechanism.
Alternatives considered¶
- Treating it as part of a locally-scoped citation/concept-linking mechanism, since its retrieval mechanism is citation/embedding-graph-flavored — rejected: the boundary is drawn by which corpus a mechanism operates over — already-fetched local corpus vs. not-yet-fetched external catalogue (discovery) — not by what kind of algorithm produced a result. Recommendations can return papers the local corpus has never seen, seeded by a paper ID instead of free text, which makes it a discovery mechanism regardless of its citation-graph-flavored implementation.
Consequences¶
None beyond the Decision above: Semantic Scholar's Recommendations API remains unadopted either way; this ADR only settles where it would land if and when adoption is revisited.