Feature request: pluggable / multilingual embedding model for semantic search

[junk151516]

The bundled Nomic embeddings work well for English-named codebases, but many real-world codebases name their domain in another language. Ours is a Colombian accounting/POS platform: business identifiers are Spanish (procesar_retenciones_batch, comprobante, lote, proveedor), infrastructure is English.

Observed: semantic queries phrased in Spanish (e.g. "match comprobante bancario con lote de pagos") return unrelated Makefile targets with min-cosine scores around 0.12:

cli search_graph --project P --semantic-query '["match comprobante bancario con lote de pagos", "monto tolerancia fecha"]' --limit 5
# → semantic_results: deploy-everything (0.132), setup-iam (0.121), deploy-api (0.121), ... (all Makefile targets)

BM25 works fine on the same index (identifier tokens are language-agnostic), so this is specifically the embedding model.

Ask: a config option to point at an alternative local embedding model (e.g. a multilingual ONNX model such as multilingual-e5-small), or a multilingual variant of the bundled model as an install option. Even a documented "semantic search is English-only" caveat in the tool description would help agents route around it.

[DeusData]

Thanks for the concrete multilingual example and the BM25 control. I have classified this as a 0.9.2-rc enhancement rather than a parsing defect: the current bundled embedding model has a capability gap, while lexical search remains available. A pluggable model path would add model-format, dimension, tokenizer, compatibility, installation-size, and local-file trust contracts, so the first step should characterize one permissively licensed multilingual candidate against the existing English benchmark and document the current limitation. No hosted embedding API or external service is in scope.