Enterprise only
LanceDB Enterprise provides scalar, aggregate, window, and table functions. Use
this index to choose a function for your workload. Availability depends on the
server version deployed to your cluster. For standard functions and their dialect
version, see the SQL functions overview and
scalar function reference. This page adds Enterprise
discovery and workload-specific functions. For help choosing text or vector
matching and running a deduplication workflow, start with the
deduplication guide.
Discover built-in functions
On server versions that supportSHOW BUILTIN FUNCTIONS, query the running
engine for its scalar, aggregate, window, and table function catalog:
name, kind, syntax, and description columns. Names include
aliases and are sorted alphabetically. LIKE matches names without regard to
case. Syntax and description can be NULL when a function has no documentation.
Internal names beginning with __ are omitted.
The kind column identifies each function as scalar, aggregate, window,
or table. Older server versions may omit table functions from this catalog.
SHOW FUNCTIONS lists published Python Functions instead of the built-in SQL
catalog. If your server does not recognize SHOW BUILTIN FUNCTIONS, ask your
administrator about upgrading.
Vector functions
The distance functions accept vectors of the same dimension. Their optional
options string selects
metric=l2 (the default), metric=cosine, metric=dot,
or metric=hamming. Dot distance is the negated dot product, so smaller values
mean more similar vectors. Hamming requires UInt8 vectors; the other metrics
accept Float16, Float32, or Float64 vectors. euclidean is an alias for l2.
For ivf_partition, random selection requires a nonnegative seed.
Text similarity functions
Start with MinHash text deduplication for the workflow.
See the MinHash SQL guide for configuration limits,
candidate recall, and complete batch examples.
Blob functions
The
length, octet_length, and bit_length functions also support Blob V2
values.
Table functions
Use table functions in aFROM clause. They return rows rather than a scalar
value per input row.
For deduplication calls, table and column names are string literals. The SQL
name is
dedupe, not dedup. The selected column at the requested snapshot
determines whether dedupe and duplicate_pairs use text or vector matching.
Neither path modifies the source table. Both require a native table without WAL.
The vector path additionally requires a fully indexed snapshot; thresholds use
the vector metric’s units. The text path requires a top-level string column or a
FixedSizeList<UInt16> MinHash signature column with configuration metadata,
and its threshold is estimated Jaccard distance from zero through one. Pass
distance_threshold by name for text calls. Text matching considers MinHash LSH
candidates and can miss pairs that meet the distance threshold; it does not
compare every pair of rows. No MinHash or vector index is required for text calls.
Use dataset_version to pin a snapshot. Deduplication supports a resolution
policy; pair enumeration does not. Text options include num_hashes,
shingle_size, and num_bands. See the linked references for vector options
and the MinHash guide for text limits and resolution
policies.
Published Python Functions
UseSHOW FUNCTIONS to discover published Functions in your current catalog
context. See Function management for publishing AI
workloads such as embedding or captioning as Functions. These are separate from the built-in functions above: