COREAICMSSAO

ES Multilingual Index & Multi-Model Embeddings

Refactor Elasticsearch multilingual indexing (per-language analyzers, locale-keyed text objects, index all content) and make embeddings per-translation, model-aware, and searched by a single cross-lingual kNN.

State
Done
Progress
100%
Steps
86 / 86
Started
2026-09-15
Finished
2026-09-15
Modules
LaraCore, LaraAI, LaraCMS, LaraSAO

Tasks.

  • Task 1: Schema translator — locale object sub-fields with analyzers + nested vector10/10
  • Task 2: resolveVectorField returns the nested dotted path5/5
  • Task 3: createIndex pushes the full mapping5/5
  • Task 4: Content document — locale-keyed text objects, locales, agnostic embeddings6/6
  • Task 5: Shared locale→analyzer map4/4
  • Task 6: Index all content — strip LocaleScope from the import query5/5
  • Task 7: Keyword search over locale sub-fields5/5
  • Task 8: Embedding-model registry (config + resolver)6/6
  • Task 9: Query/passage prefixes6/6
  • Task 10: ModelEmbedding schema — locale + model_key6/6
  • Task 11: Per-locale generation + agnostic ES embeddings8/8
  • Task 12: Incremental re-embed triggers on ContentTranslation5/5
  • Task 13: Vector search over the nested field + document locales filter5/5
  • Task 14: Ticket embeddable (mono-lingual exemplar)5/5
  • Task 15: ai:embeddings:repair per-locale + /health cross-check5/5

Read the full plan on GitHub