Semantic Project Search

Embeddings help when a visitor describes a need rather than naming a technology.

Keyword search fails when the visitor and the author use different vocabulary.

• Index case studies with both keyword and embedding search
• Run the same natural-language queries through both
• Compare the top results by relevance judgement

Text embeddings with cosine similarity ranking.

No human approval step is required, because the output is only a ranking or suggestion and is never published as fact.

Pending — the comparison harness is defined but not yet run at scale.

• The archive is too small to draw statistically meaningful conclusions.
• Embedding cost and latency matter more than accuracy at this size.

Only public portfolio content is indexed.

Search ranking is presented as a suggestion, not an authoritative answer.

Collect a set of realistic visitor queries to evaluate against.