Problem

EmbeddingFunction methods (compute_source_embeddings, compute_query_embeddings) are synchronous-only. When used with AsyncTable, this blocks the event loop in production async frameworks like FastAPI.

Inconsistency between add() and search()

AsyncTable.search() correctly wraps embedding calls in run_in_executor:

# table.py – AsyncTable.search()
async def make_embedding(embedding, query):
    loop = asyncio.get_running_loop()
    return (await loop.run_in_executor(
        None, embedding.function.compute_query_embeddings_with_retry, query,
    ))[0]

But AsyncTable.add() calls embeddings synchronously on the event loop via _append_vector_columns():

# table.py – _append_vector_columns()
col_data = func.compute_source_embeddings_with_retry(batch[conf.source_column])

This blocks all concurrent requests when ingesting data with registered embedding functions.

Broader issue

The EmbeddingFunction ABC has no async interface at all. The TypeScript SDK already supports async computeSourceEmbeddings() / async computeQueryEmbeddings(), but the Python SDK has no equivalent. Most embedding providers (OpenAI, Cohere, Bedrock, etc.) offer async clients that could be leveraged.

Suggested improvements

1. Short-term: Wrap compute_source_embeddings_with_retry in run_in_executor inside AsyncTable.add(), matching the pattern already used in AsyncTable.search()
2. Long-term: Add async variants to EmbeddingFunction (async_compute_source_embeddings, async_compute_query_embeddings) so implementations can use native async HTTP clients instead of thread pool workarounds

Environment

• lancedb 0.30.2
• Python 3.13
• FastAPI