Changelog

2.0.1 — 2026-08-21

Non-English text no longer crashes the embedder (#10). Inputs were cut to 1,500 characters, but TaylorAI/bge-micro-v2 ships a placeholder model_max_length, so 1,500 chars of German (~560 tokens) overflowed its 512 position embeddings, onnxruntime threw, and search_notes silently dropped to keyword-fallback. The tokenizer is now capped at the model's real max_position_embeddings, so truncation happens at the token level for any language — including the parity check, which embeds a full stored note.
Stale model_key in smart_env.json no longer empties the index (#9). Switching models in Smart Connections can re-embed every .ajson under the new key while the config keeps the old one; every vector lookup then missed and search returned [] in "mode": "semantic". The server now trusts the data when the declared model has no vectors, logs the override, and reports the config value as declaredModelKey in get_stats / list_vaults.
A vault with no vectors indexed now falls back to keyword matching with an explicit warning instead of returning an empty semantic result.

True semantic search: search_notes now embeds your query locally with the same model your vault's Smart Connections index used (via transformers.js). Conceptual queries work; nothing leaves your machine.
Multi-vault: SMART_VAULT_PATH accepts comma-separated paths; tools take an optional vault parameter; new list_vaults tool.
Block-level retrieval: search matches individual sections (blocks) and returns content snippets inline.
Freshness: .smart-env changes are picked up automatically (throttled incremental reload) — no server restart after editing notes.
get_note_content include_blocks now actually extracts the named blocks.
Explicit mode: "keyword-fallback" + warning when the embedding model cannot load, instead of silently degraded results.

get_embedding_neighbors tool.
search_notes response shape changed (adds vault, scope, block, snippet, mode).