L'étape 'Building inverted index...' au démarrage bloquait l'event loop
asyncio, empêchant le serveur de répondre aux requêtes HTTP (navigation,
refresh) jusqu'à la fin de l'indexation.
- init_inverted_index() exécuté dans ThreadPoolExecutor via run_in_executor
- get_inverted_index() ne fait plus de rebuild auto bloquant
- Le serveur répond immédiatement, même pendant l'indexation
Replace double-nested stem loop (stems × tokens) with single-pass
stem frequency map. For 100 unique tokens per file: 10,000 iterations
→ 100 iterations per file. Critical for large vaults.
Add French snowball stemmer to tokenization pipeline:
- Index both original tokens AND their stems in InvertedIndex
- Query terms are also stemmed before lookup
- Stemmed forms accumulate TF from all original forms
- Lazy-init singleton pattern for stemmer
Register a hook with the indexer so that file add/remove events
incrementally maintain the inverted index, removing the need for
periodic staleness checks and cooldowns. Rebuild the index once on
startup via init_inverted_index().
- Add ext: operator support to query parser in backend and frontend
- Update search documentation in README and help modal with ext: examples
- Parse ext: operator to extract file extension filter (strips leading dot, converts to lowercase)
- Filter search candidates by file extension in advanced_search function
- Add ext chip display in search UI alongside existing tag/vault/title/path chips
- Update API documentation and function
bruno