Define the embedding function registry in LanceDB?

The embedding function registry is LanceDB's mechanism for attaching an embedding model directly to a table's schema, so that inserting raw data (like text) automatically generates and stores the corresponding vector, without the application needing to call an embedding model explicitly before every insert.

from lancedb.embeddings import get_registry
from lancedb.pydantic import LanceModel, Vector

registry = get_registry()
func = registry.get("sentence-transformers").create()

class Document(LanceModel):
    text: str = func.SourceField()
    vector: Vector(func.ndims()) = func.VectorField()

table = db.create_table("documents", schema=Document)
table.add([{"text": "hello world"}])  # vector generated automatically

The registry ships with built-in support for many popular embedding providers — OpenAI, Cohere, sentence-transformers, and multimodal models like CLIP — and a table's schema marks which field is the SourceField (the raw input, like text) and which is the VectorField (the generated embedding), so both query-time and insert-time embedding generation are handled consistently.

The trade-off is that if the underlying embedding function later changes, the table generally needs to be reconfigured and its embeddings regenerated to stay consistent, since old rows were embedded with a different model or set of weights than new ones would be.