Multimodal API
pypaimon.multimodal provides a compact interface for scalar, text, vector, and BLOB data on Paimon data-evolution tables without primary keys. A connection wraps a catalog and a default database.
Start with Multimodal Tables to connect, define a schema, and add rows. Then follow the guide for your workload:
| Task | Guide |
|---|---|
| Create or change rows | Multimodal tables |
| Read payloads and coordinate jobs | Multimodal reads and row IDs |
| Retrieve text and vector candidates | Search |
| Store media | BLOB storage, object API, video frames |
| Import robotics datasets | HDF5 and ROSBag, LeRobot |
| Update selected columns | Data evolution |
| Train and process data | PyTorch, Ray Data, Daft |
Create and modify tables
Connect to a catalog, define scalar and payload columns, add rows, and apply overwrite, update, delete, or merge operations.
Read rows and payloads
Filter and project rows, read BLOBs in batches, build training windows, and pass row IDs between retrieval and processing stages.
Search text and vectors
Create indexes and run single-vector, full-text, hybrid, or batch-vector queries. Includes pre-filter and result-filter semantics.
Store video frames
Write complete encoded videos with logical frame rows, replace a video payload, and decode frames in a PyTorch DataLoader.
Import HDF5 and ROSBag
Transform local or remote HDF5 and ROSBag sources into table rows, with source-specific options and distributed ingestion paths.
Work with LeRobot
Import Dataset v3, capture frame data directly, retain component tags, and train from a complete Paimon table group.
Use the BLOB object API
Put, read, delete, and update objects using table columns as keys and payloads. For the lower-level storage API, see BLOB Storage.