Manticore Search 29.9.0 introduces chunked and multi-vector auto-embeddings, allowing documents to be split into fragments. Users can now choose from five chunking strategies, including truncate, mean, fixed, recursive, and sentence-based approaches.
The update sets mmap as the default access mode for columnar attributes, replacing buffered file reads. Users can revert to the previous behavior by explicitly setting access_columnar_attrs to file.
New features include support for UTF-8 identifiers for table and field names, and a MAX_INPUT_TOKENS option to limit the input size for local embedding models. Additionally, Manticore Backup now supports authentication via username, password, or bearer tokens.
The release addresses several operational issues, including fixes for hybrid search, KNN queries, and bulk ingestion. It also improves compatibility with S3 backups and adds benchmark capabilities to manticore-load via HTTP JSON API.
No mandatory data migration is required for this release. However, users utilizing German AOT morphology or authenticated backups should review the specific update notes to ensure proper configuration.
