---
title: "Create a Function"
sidebarTitle: "Create Function"
description: "Register Python as a column: CPU, GPU, and remote-service Functions, called once per row or once per batch."
icon: code
keywords: ["function", "udf", "cpu", "gpu", "batch", "scalar", "blob", "remote api", "embedding", "rate limiter"]
availability:
embedded: unavailable
enterprise: available
---
Some curation steps cannot be written as a SQL expression: an embedding
model, an image decode, a call to an external service. A Function is a
Python callable registered with the database, versioned, and then applied as
a column or called in a query -- the code runs where the data is, not pulled
into your process first. This page covers how to shape one; later curation
tasks such as [Chunking](/curation/chunking) assume you have read it. The
full registration, listing, refresh and drop lifecycle is covered on the
[Function management reference](/management/function); the examples below
focus on shaping a Function for common curation tasks.
## Scalar: once per row {#scalar}
The plain form takes one row's values and returns one result. Types are
inferred from the annotations:
```python Python icon="python"
from lancedb import udf
@udf(pip=["numpy==2.2.0"])
def score(value: float) -> float:
return value * 2
version = db.create_function_async(score).wait()
table.add_columns({"doubled": version(value=col("value"))})
```
`pip` declares the worker's dependencies; they are part of the function's
immutable version, so changing one produces a new version rather than
altering rows a previous version already filled.
## Batch: once per call, many rows at a time {#batch}
A function is called once per batch instead of once per row when its
parameters are annotated `pyarrow.Array`, or when it takes a single
`pyarrow.RecordBatch` parameter to read more than one column together. The
same doubling logic, vectorized over a whole batch at once:
```python Python icon="python"
import pyarrow as pa
import pyarrow.compute as pc
from lancedb import udf
@udf(
pip=["numpy==2.2.0"],
batch_size=32,
row_timeout_seconds=90,
)
def score_batch(value: pa.Array) -> pa.Array:
return pc.multiply(value, 2)
```
Batching amortizes per-call overhead across many rows, which matters most
once a function also does something with real fixed cost per call -- loading
a model, opening a connection, making a network round trip.
`batch_size` caps the number of input rows in each execution batch; actual
batches may be smaller. `row_timeout_seconds` bounds each Python call: one
row for scalar Functions or the entire batch for vectorized Functions. It
does not include the instance's `__init__`. Both settings are fixed into the
registered version and apply to direct calls, computed columns, and
materialized-view refreshes alike. Omit them to use adaptive batching and the
service's default per-call deadline.
## Stateful Functions: load once, call many times {#stateful}
A `@udf`-decorated **class** is the same callable with state. Each remote
instance calls `__init__` once, calls `__call__` for every row or batch it
processes, and calls `close()`, if defined, once when it retires:
```python Python icon="python"
from lancedb import udf
@udf
class scale:
def __init__(self, factor: float, offset: float = 0.0):
self.factor = factor
self.offset = offset
def __call__(self, value: float) -> float:
return value * self.factor + self.offset
```
Constructor parameters are **initialization fields**, bound once per binding
rather than read per row -- mixed with per-row columns in the same call:
`version(value=col("value"), factor=2.0)`. Load models, open clients, and
build rate limiters in `__init__`; every batch the instance serves then
reuses them, since an instance's state is shared across its own batches but
never across other instances or workers. This is the pattern GPU inference
and remote-API Functions both build on below.
## GPU: batch inference, loaded once {#gpu}
`db.list_gpu_classes()` lists the GPU classes configured for the deployment.
Each record reports the operator's hardware description and limits; it does
not mean a GPU is currently free. Listing requires the System Usage
permission. Pass a returned class name to `gpu_class` to fix that placement
in the Function version. `gpu=True` remains required:
```python Python icon="python"
import pyarrow as pa
from lancedb import udf
gpu_class = next(
gpu.name for gpu in db.list_gpu_classes() if gpu.gpu_model == "NVIDIA A10G"
)
@udf(
gpu=True,
gpu_class=gpu_class,
pip=["torch==2.12.0", "sentence-transformers==5.7.0"],
)
class embed_text:
def __init__(self, model_name: str = "all-MiniLM-L6-v2"):
import torch
from sentence_transformers import SentenceTransformer
device = "cuda" if torch.cuda.is_available() else "cpu"
self.model = SentenceTransformer(model_name, device=device)
def __call__(self, text: pa.Array) -> pa.Array:
texts = [t or "" for t in text.to_pylist()]
embeddings = self.model.encode(texts)
return pa.array(embeddings.tolist(), type=pa.list_(pa.float32(), list_size=384))
version = db.create_function_async(embed_text).wait()
table.add_columns({"embedding": version(text=col("caption"))})
```
Older servers that do not implement `GET /v1/gpu/classes` raise
`NotImplementedError`. Without `gpu_class`, the deployment's default GPU
class is used. The class is immutable with the Function version; set a
per-refresh class or aggregate GPU budget through
[`refresh_column_async`](/feature-engineering/refresh-column#resources).
The model loads once per worker instance, in `__init__`; `__call__` only
encodes the batch it was handed. Registering the same logic as a plain
function instead would reload the model on every single row.
## Calling a remote service {#remote}
A function that calls an external API needs a credential without putting it
in the function's source. A Secret is a database-scoped named credential:
`db.create_secret` stores the value once, and binding it at registration time
puts it in the environment variable the function reads, without ever
returning it from any API or recording it in the function's own definition.
Combine it with the stateful form so the HTTP client -- and any rate limiter
around it -- is built once, not reopened on every row. There is no built-in
retry or per-row error-skip policy -- a function that calls a flaky service
should handle timeouts and retries itself, inside `__call__`.
See [Vend a Secret to a Function](/curation/vend-secret#specify) for the
full worked example, how the bound value actually reaches a running
evaluation, and how that differs from rotating or revoking it later.
## Read and produce blobs {#blobs}
A Function can bind a [Blob v2 column](/management/table#managed-blobs) and
return another blob, including a blob nested inside a struct field, not only
at the top level. In row mode, a top-level blob parameter arrives as Python
`bytes`, or `None` for a nullable input, after the worker has read the
payload. Return bytes and let the destination table store the Managed
payload and descriptor.
This section extracts a frame from a stored video and captions it,
continuing as one example across chaining, persisting and incremental reads
below. `videos` holds raw clips in a `video` Blob v2 column; `frames` names
one timestamp per clip to extract, with a `selected` flag to skip some:
```python Python icon="python"
import pyarrow as pa
videos = db.create_table("videos", schema=pa.schema([
pa.field("id", pa.utf8(), nullable=False),
lancedb.blob("video", nullable=False),
]))
frames = db.create_table("frames", [
{"frame_id": 1, "video_id": "clip-1", "ts": 0.5, "selected": True},
{"frame_id": 2, "video_id": "clip-1", "ts": 1.5, "selected": True},
])
```
### Produce a blob {#produce-blobs}
Declare blob inputs and outputs explicitly. `extract_frame` reads one frame
out of a video at a timestamp and returns it as a JPEG, without loading the
whole clip first -- see [Read a blob incrementally](#incremental) below for
how the `video: BlobReader` parameter does that:
```python Python icon="python"
from datetime import timedelta
from lancedb import BlobReader, col
from lancedb.functions import udf
@udf(
name="extract_frame",
gpu=True,
pip=["av==13.1.0", "pillow==11.0.0"],
input_schema=pa.schema([
lancedb.blob("video"),
pa.field("ts", pa.float64(), nullable=False),
]),
output_schema=lancedb.blob("jpeg", nullable=False),
)
def extract_frame(video: BlobReader, ts: float) -> bytes:
... # decode the frame at ts, encode as JPEG, return its bytes
extract_frame_version = db.create_function_async(extract_frame).wait(
timeout=timedelta(minutes=15)
)
```
An empty byte string is a filled result. A scalar output cannot be nullable,
because a physical NULL marks an unfilled computed row. To return a nullable
blob as a value, use a named-struct output schema with a nullable blob field;
the struct itself records whether the result has been filled. Read a nested
blob with `table.fetch_blobs("payload.image", hits)`.
A list or large-list output may also contain Blob v2 elements; inputs and
fixed-size lists do not support this layout. List-nested values are queryable
as descriptors, but `fetch_blobs` and `blob_mode="bytes"` cannot fetch their
payloads.
`output_schema=lancedb.blob(...)` is what makes a persisted result Blob v2
rather than plain Arrow `binary`. Returning `bytes` from a function whose
output type is inferred from a bare annotation, with no explicit
`lancedb.blob(...)` output schema, persists as ordinary `binary`: the column
still holds the right bytes, but it skips the Managed storage path, descriptor
reads, and `fetch_blobs`. Declare it explicitly on every function whose output
should be a blob, not just the first one in a chain.
### Chain a Function on its output {#chain}
A second Function can bind a blob another Function produced. Register
`caption` the same way, consuming a blob instead of producing one:
```python Python icon="python"
@udf(gpu=True, pip=["transformers==4.47.0"])
def caption(jpeg: bytes) -> str:
... # run a vision-language model over the JPEG, return one sentence
caption_version = db.create_function_async(caption).wait(timeout=timedelta(minutes=15))
```
Compose both in one SQL statement over a join, instead of applying each as
its own column. The join supplies `video` and `ts` from two different
tables, and `caption` reads `extract_frame`'s result as it is computed,
before anything is written:
```sql SQL
SELECT f.frame_id, extract_frame(v.video, f.ts) AS jpeg,
caption(extract_frame(v.video, f.ts)) AS caption
FROM videos v JOIN frames f ON v.id = f.video_id
WHERE f.selected = true
```
[`EXPLAIN`](/sql/dql/explain) on a query like this shows `RemoteTake`: the
blob is read once, where the worker runs, and never pulled into the query
node first.
### Persist it as a column, or a materialized view {#persist}
Apply the same registered version with `add_columns` to persist its output as
an ordinary table column, on a table that already has both of its inputs --
here, a single-clip table with its own `video` and `ts`, rather than the two
tables the join above reads from:
```python Python icon="python"
clips = db.create_table("clips", schema=pa.schema([
pa.field("frame_id", pa.int64(), nullable=False),
lancedb.blob("video", nullable=False),
pa.field("ts", pa.float64(), nullable=False),
]))
clips.add_columns({"jpeg": extract_frame_version(video=col("video"), ts=col("ts"))})
result = clips.refresh_column_async("jpeg").wait(timeout=timedelta(minutes=15))
assert result.rows_failed == 0
hits = clips.search().select(["frame_id", "jpeg"]).with_row_id(True).to_arrow()
values = clips.fetch_blobs("jpeg", hits).to_pylist()
```
Or wrap the join and both Functions in a
[materialized view](/sql/ddl/materialized-view) instead, so the extracted
frame and its caption stay current as new videos and frames arrive. See
[Auto Refresh](/enterprise/auto-refresh) for
keeping either one fresh without calling refresh yourself.
On deployments that support computed-column inputs, another Function can
bind to `col("jpeg")` directly, the same as the SQL join above does through
its own expression. Refresh the producer first, then the consumer. Once
outputs are current, a later refresh fills appended rows. A filtered
refresh preserves payload references for unselected Managed outputs. An
unfiltered refresh can recompute a whole fragment when it cannot establish
that the fragment's outputs are current.
Function output blobs follow the same compaction and cleanup lifecycle as
ingested Managed blobs. Retained versions continue to own their payloads;
cleanup reclaims an object only after its retained references are gone.
### Read a blob incrementally {#incremental}
Annotate a row Function parameter with `BlobReader` to read a Blob v2 input
incrementally, the way `extract_frame` above does, instead of materializing
the whole payload first. The signature selects the access mode without
another registration option. A minimal example, reading only the first MP4
box type without loading the rest of the video:
```python Python icon="python"
from lancedb import BlobReader, udf
@udf
def first_box_type(video: BlobReader) -> bytes:
video.seek(4)
return video.read(4)
```
The worker passes a seekable reader backed by the stored blob --
`video.seekable()` is `True`. `read(size)` reads from the current cursor;
`read_range(offset, length)` reads a byte range without moving it, the same
bytes a `seek(offset)` followed by `read(length)` would read. `size()`
returns the payload size without reading it. Each row call receives its own
reader, which closes when the call returns or raises. Do not retain it for
another invocation.
Use `video: bytes` when the Function needs the complete payload. Bytes inputs
materialize before invocation and remain subject to eager input limits.
Reader inputs keep the unread payload out of the input batch, but calling
`video.read()` without a size loads all remaining bytes in Python. Reader
inputs are supported in row mode. Return values remain ordinary values such
as `bytes`; this feature does not add streamed outputs.
## What to read next {#next}
List, describe and drop registered Function versions.
Bind a credential by reference instead of putting it in the function's source.
A Function whose output is a list, exploded into one row per element.
Filling the column a Function is bound to, and re-filling it as inputs change.