Multilingual embedding for catalog.search (Norwegian catalog comments)

Carried out of the R118 MCP-server programme (discarded 2026-07-13 with slices 1-10 shipped) as its one live unresolved open question. The MCP semantic tools (catalog.search, docs.search) embed with bge-small-en-v1.5-q (graphitron-mcp/.../rag/BgeEmbedder.java, chosen in R372 D1), which is English-only. Consumer databases, notably Sikt’s own, may carry Norwegian table and column comments; against such a catalog the semantic half of catalog.search degrades to whatever the English model happens to extract from Norwegian text, and retrieval quality on the comment signal is unquantified. The docs corpus is English, so docs.search is unaffected.

Direction sketched in R118: swap to a multilingual model such as multilingual-e5-small behind the existing langchain4j EmbeddingModel seam (rag/Embedder.java keeps the model swappable by design). That may mean bundling the ONNX model ourselves rather than using a prebuilt langchain4j module, and e5 has its own asymmetric query-prefix convention to encode where BgeEmbedder encodes bge’s today. Gate the work on evidence: before swapping, measure retrieval quality of the current model against a real Norwegian-commented catalog (an opptak-style schema) to confirm the degradation is material. The hybrid BM25 half of the Lucene index already matches Norwegian tokens verbatim, which may be good enough for name-plus-comment discovery in practice.