@lancedb/lancedb / Exports / VectorQuery
Class: VectorQuery
A builder used to construct a vector search
This builder can be reused to execute the query many times.
Hierarchy
QueryBase\<NativeVectorQuery,VectorQuery>
↳ VectorQuery
Table of contents
Constructors
Properties
Methods
- [asyncIterator]
- bypassVectorIndex
- column
- distanceType
- execute
- limit
- nativeExecute
- nprobes
- postfilter
- refineFactor
- select
- toArray
- toArrow
- where
Constructors
constructor
• new VectorQuery(inner): VectorQuery
Parameters
| Name | Type |
|---|---|
inner |
VectorQuery |
Returns
Overrides
Defined in
Properties
inner
• Protected inner: VectorQuery
Inherited from
Defined in
Methods
[asyncIterator]
▸ [asyncIterator](): AsyncIterator\<RecordBatch\<any>, any, undefined>
Returns
AsyncIterator\<RecordBatch\<any>, any, undefined>
Inherited from
Defined in
bypassVectorIndex
▸ bypassVectorIndex(): VectorQuery
If this is called then any vector index is skipped
An exhaustive (flat) search will be performed. The query vector will be compared to every vector in the table. At high scales this can be expensive. However, this is often still useful. For example, skipping the vector index can give you ground truth results which you can use to calculate your recall to select an appropriate value for nprobes.
Returns
Defined in
column
▸ column(column): VectorQuery
Set the vector column to query
This controls which column is compared to the query vector supplied in the call to
Parameters
| Name | Type |
|---|---|
column |
string |
Returns
See
This parameter must be specified if the table has more than one column whose data type is a fixed-size-list of floats.
Defined in
distanceType
▸ distanceType(distanceType): VectorQuery
Set the distance metric to use
When performing a vector search we try and find the "nearest" vectors according to some kind of distance metric. This parameter controls which distance metric to use. See
Parameters
| Name | Type |
|---|---|
distanceType |
string |
Returns
See
IvfPqOptions.distanceType for more details on the different distance metrics available.
Note: if there is a vector index then the distance type used MUST match the distance type used to train the vector index. If this is not done then the results will be invalid.
By default "l2" is used.
Defined in
execute
▸ execute(): RecordBatchIterator
Execute the query and return the results as an
Returns
See
- AsyncIterator of
- RecordBatch.
By default, LanceDb will use many threads to calculate results and, when the result set is large, multiple batches will be processed at one time. This readahead is limited however and backpressure will be applied if this stream is consumed slowly (this constrains the maximum memory used by a single query)
Inherited from
Defined in
limit
▸ limit(limit): VectorQuery
Set the maximum number of results to return.
By default, a plain search has no limit. If this method is not called then every valid row from the table will be returned.
Parameters
| Name | Type |
|---|---|
limit |
number |
Returns
Inherited from
Defined in
nativeExecute
▸ nativeExecute(): Promise\<RecordBatchIterator>
Returns
Promise\<RecordBatchIterator>
Inherited from
Defined in
nprobes
▸ nprobes(nprobes): VectorQuery
Set the number of partitions to search (probe)
This argument is only used when the vector column has an IVF PQ index. If there is no index then this value is ignored.
The IVF stage of IVF PQ divides the input into partitions (clusters) of related values.
The partition whose centroids are closest to the query vector will be exhaustiely searched to find matches. This parameter controls how many partitions should be searched.
Increasing this value will increase the recall of your query but will also increase the latency of your query. The default value is 20. This default is good for many cases but the best value to use will depend on your data and the recall that you need to achieve.
For best results we recommend tuning this parameter with a benchmark against your actual data to find the smallest possible value that will still give you the desired recall.
Parameters
| Name | Type |
|---|---|
nprobes |
number |
Returns
Defined in
postfilter
▸ postfilter(): VectorQuery
If this is called then filtering will happen after the vector search instead of before.
By default filtering will be performed before the vector search. This is how filtering is typically understood to work. This prefilter step does add some additional latency. Creating a scalar index on the filter column(s) can often improve this latency. However, sometimes a filter is too complex or scalar indices cannot be applied to the column. In these cases postfiltering can be used instead of prefiltering to improve latency.
Post filtering applies the filter to the results of the vector search. This means
we only run the filter on a much smaller set of data. However, it can cause the
query to return fewer than limit results (or even no results) if none of the nearest
results match the filter.
Post filtering happens during the "refine stage" (described in more detail in
Returns
See
VectorQuery#refineFactor). This means that setting a higher refine factor can often help restore some of the results lost by post filtering.
Defined in
refineFactor
▸ refineFactor(refineFactor): VectorQuery
A multiplier to control how many additional rows are taken during the refine step
This argument is only used when the vector column has an IVF PQ index. If there is no index then this value is ignored.
An IVF PQ index stores compressed (quantized) values. They query vector is compared against these values and, since they are compressed, the comparison is inaccurate.
This parameter can be used to refine the results. It can improve both improve recall and correct the ordering of the nearest results.
To refine results LanceDb will first perform an ANN search to find the nearest
limit * refine_factor results. In other words, if refine_factor is 3 and
limit is the default (10) then the first 30 results will be selected. LanceDb
then fetches the full, uncompressed, values for these 30 results. The results are
then reordered by the true distance and only the nearest 10 are kept.
Note: there is a difference between calling this method with a value of 1 and never
calling this method at all. Calling this method with any value will have an impact
on your search latency. When you call this method with a refine_factor of 1 then
LanceDb still needs to fetch the full, uncompressed, values so that it can potentially
reorder the results.
Note: if this method is NOT called then the distances returned in the _distance column will be approximate distances based on the comparison of the quantized query vector and the quantized result vectors. This can be considerably different than the true distance between the query vector and the actual uncompressed vector.
Parameters
| Name | Type |
|---|---|
refineFactor |
number |
Returns
Defined in
select
▸ select(columns): VectorQuery
Return only the specified columns.
By default a query will return all columns from the table. However, this can have a very significant impact on latency. LanceDb stores data in a columnar fashion. This means we can finely tune our I/O to select exactly the columns we need.
As a best practice you should always limit queries to the columns that you need. If you pass in an array of column names then only those columns will be returned.
You can also use this method to create new "dynamic" columns based on your existing columns.
For example, you may not care about "a" or "b" but instead simply want "a + b". This is often
seen in the SELECT clause of an SQL query (e.g. SELECT a+b FROM my_table).
To create dynamic columns you can pass in a Map
For example, an SQL query might state SELECT a + b AS combined, c. The equivalent
input to this method would be:
Parameters
| Name | Type |
|---|---|
columns |
string[] | Record\<string, string> | Map\<string, string> |
Returns
Example
new Map([["combined", "a + b"], ["c", "c"]])
Columns will always be returned in the order given, even if that order is different than
the order used when adding the data.
Note that you can pass in a `Record<string, string>` (e.g. an object literal). This method
uses `Object.entries` which should preserve the insertion order of the object. However,
object insertion order is easy to get wrong and `Map` is more foolproof.
Inherited from
Defined in
toArray
▸ toArray(): Promise\<unknown[]>
Collect the results as an array of objects.
Returns
Promise\<unknown[]>
Inherited from
Defined in
toArrow
▸ toArrow(): Promise\<Table\<any>>
Collect the results as an Arrow
Returns
Promise\<Table\<any>>
See
ArrowTable.
Inherited from
Defined in
where
▸ where(predicate): VectorQuery
A filter statement to be applied to this query.
The filter should be supplied as an SQL query string. For example:
Parameters
| Name | Type |
|---|---|
predicate |
string |
Returns
Example
x > 10
y > 0 AND y < 100
x > 5 OR y = 'test'
Filtering performance can often be improved by creating a scalar index
on the filter column(s).