Indexing Language Reference

This reference documents the full Vespa indexing language. If more complex processing of input data is required, implement a document processor.

The indexing language is analogous to UNIX pipes, in that statements consists of expressions separated by the pipe symbol where the output of each expression is the input of the next. Statements are terminated by semicolon and are independent of each other (except when using variables).

Indexing script

An indexing script is a sequence of indexing statements separated by a semicolon (;). A script is executed statement-by-statement, in order, one document at a time.

Vespa derives one indexing script per search cluster based on the search definitions assigned to that cluster. As a document is fed to a search cluster, it passes through the corresponding indexing cluster, which runs the document through its indexing script. Note that this also happens whenever the document is reindexed, so expressions such as now must be thought of as the time the document was (last) indexed, not when it was fed.

You can examine the indexing script generated for a specific search cluster by retrieving the configuration of the indexing document processor.

$ vespa-get-config -i search/cluster.<cluster-name> -n vespa.configdefinition.ilscripts

The current execution value is set to null prior to executing a statement.

Indexing statement

An indexing statement is a sequence of indexing expressions separated by a pipe (|). A statement is executed expression-by-expression, in order.

Within a statement, the execution value is passed from one expression to the next.

The simplest of statements passes the value of an input field into an attribute:

input year | attribute year;

The above statement consists of 2 expressions; input year and attribute year. The former sets the execution value to the value of the "year" field of the input document. The latter writes the current execution value into the attribute "year".

Indexing expression

Primitives

A string, numeric literal and true/false can be used as an expression to explicitly set the execution value. Examples: "foo", 69, true).

Outputs

An output expression is an expression that writes the current execution value to a document field. These expressions also double as the indicator for the type of field to construct (i.e. attribute, index or summary). It is important to note that you can not assign different values to the same field in a single document (e.g. attribute | lowercase | index is illegal and will not deploy).

Arithmetics

Indexing statements can contain any combination of arithmetic operations, as long as the operands are numeric values. In case you need to convert from string to numeric, or convert from one numeric type to another, use the applicable converter expression. The supported arithmetic operators are:

You may use parenthesis to declare precedence of execution (e.g. (1 + 2) * 3). This also works for more advanced array concatenation statements such as (input str_a | split ',') . (input str_b | split ',') | index arr.

Converters

There are several expressions that allow you to convert from one data type to another. These are often used within a for_each to convert e.g. an array of strings to an array of integers.

Other expressions

The following are the unclassified expressions available:

Execution value example

Accessing the execution value (the value passed into this expression) explicitly is useful when it is to be used as part of an expression such as concatenation. In this example we have a document with a title and an array of sentences, and we prepend each sentence by the document title (and a space), before converting it to a set of embedding vectors (represented by a 2d mixed tensor).

  input mySentenceArray | for_each { input title . " " . _ } | embed | attribute my2dTensor | index my2dTensor

Choice (||) example

The choice expression is used to provide alternatives if an expression may return null.

  (input myField1 || "") . " " . (input myField2 || "") | embed | attribute | index

In this example two fields are concatenated, but if one of the fields is empty, the empty string is used instead. If the empty string alternatives are not provided, no embedding will be produced if either input field is missing.

select_input example

The select_input expression is used to choose a statement to execute based on which fields are non-empty in the input document:

select_input {
    CX:   input CX | set_var CX;
    CA:   input CA . " " . input CB | set_var CX;
}

This statement executes input CX | set_var CX; unless CX is empty. If so, it will execute input CA . " " . input CB | set_var CX; unless CA is empty.

Switch example

The switch-expression behaves similarly to the switch-statement in other programming languages. Each case in the switch-expression consists of a string and a statement. The execution value is compared to each string, and if there is a match, the corresponding statement is executed. An optional default operation (designated by default:) can be added to the end of the switch:

input mt | switch {
    case "audio": input fa | index;
    case "video": input fv | index;
    default: 0 | index;
};

Indexing statements example

Using indexing statements, multiple document fields can be used to produce one index structure field. For example, the index statement:

input field1 . input field2 | attribute field2;

combines field1 and field2 into the attribute named field2. When partially updating documents which contains indexing statement which combines multiple fields the following rules apply:

• Only attributes where all the source values are available in the source document update will be updated
• The document update will fail when indexed (only) if no attributes end up being updated when applying the rule above

Example: If a schema has the indexing statements

input field1 | attribute field1;
input field1 . input field2 | attribute field2;

the following will happen for the different partial updates: