You can further limit the vector similarity search by providing a filter based on a specific metadata criteria.

Queries with metadata filters only return vectors which have metadata matching with the filter.

Upstash Vector allows you to filter keys which have the following value types:

  • string
  • number
  • boolean
  • object
  • array

Filtering is implemented as a combination of in and post-filtering. Every query is assigned a filtering budget, determining the number of candidate vectors that can be compared against the filter during query execution. If this budget is exceeded, the system fallbacks into post-filtering. Therefore, with highly selective filters, fewer than topK vectors may be returned.

Filter Syntax

A filter has a syntax that resembles SQL, which consists of operators on object keys and boolean operators to combine them.

Assuming you have a metadata like below:

{
    "city": "Istanbul",
    "country": "Turkey",
    "is_capital": false,
    "population": 15460000,
    "geography": {
        "continent": "Asia",
        "coordinates": {
            "latitude": 41.0082,
            "longitude": 28.9784
        }
    },
    "economy": {
        "currency": "TRY",
        "major_industries": [
            "Tourism",
            "Textiles",
            "Finance"
        ]
    }
}

Then, you can query similar vectors with a filter like below:

curl https://powerful-kodiak-60521-us1-vector.upstash.io/query \
  -H "Authorization: Bearer UPSTASH_VECTOR_TOKEN" \
  -d '{
   "vector":[0.9215,0.3897,...],
   "topK" : 5,
   "filter": "population >= 1000000 AND geography.continent = \"Asia\"",
   "includeMetadata": true
}'

Operators

Equals (=)

The equals operator filters keys whose value is equal to given literal.

It is applicable to string, number, and boolean values.

country = 'Turkey' AND population = 15460000 AND is_capital = false

Not Equals (!=)

The equals operator filters keys whose value is not equal to given literal.

It is applicable to string, number, and boolean values.

country != 'Germany' AND population != 12500000 AND is_capital != true

Less Than (<)

The less than operator filters keys whose value is less than the given literal.

It is applicable to number values.

population < 20000000 OR geography.coordinates.longitude < 30.0

Less Than or Equals (<=)

The less than or equals operator filters keys whose value is less than or equal to the given literal.

It is applicable to number values.

population <= 20000000 OR geography.coordinates.longitude <= 30.0

Greater Than (>)

The greater than operator filters keys whose value is greater than the given literal.

It is applicable to number values.

population > 10000000 OR geography.coordinates.latitude > 39.5

Greater Than or Equals (>=)

The greater than or equals operator filters keys whose value is greater than or equal to the given literal.

It is applicable to number values.

population >= 10000000 OR geography.coordinates.latitude >= 39.5

Glob

The glob operator filters keys whose value matches the given UNIX glob pattern.

It is applicable to string values.

It is a case sensitive operator.

The glob operator supports the following wildcards:

  • * matches zero or more characters.
  • ? matches exactly one character.
  • [] matches one character from the list
    • [abc] matches either a, b, or c.
    • [a-z] matches one of the range of characters from a to z.
    • [^abc] matches any one character other than a, b, or c.
    • [^a-z] matches any one character other than a to z.

For example, the filter below would only match with city names whose second character is s or z, and ends with anything other than m to z.

city GLOB '?[sz]*[^m-z]'

Not Glob

The not glob operator filters keys whose value does not match the given UNIX glob pattern.

It is applicable to string values.

It has the same properties with the glob operator.

For example, the filter below would only match with city names whose first character is anything other than A.

city NOT GLOB 'A*'

In

The in operator filters keys whose value is equal to any of the given literals.

It is applicable to string, number, and boolean values.

country IN ('Germany', 'Turkey', 'France')

Semantically, it is equivalent to equals operator applied to all of the given literals with OR boolean operator in between:

country = 'Germany' OR country = 'Turkey' OR country = 'France'

Not In

The not in operator filters keys whose value is not equal to any of the given literals.

It is applicable to string, number, and boolean values.

economy.currency NOT IN ('USD', 'EUR')

Semantically, it is equivalent to not equals operator applied to all of the given literals with AND boolean operator in between:

economy.currency != 'USD' AND economy.currency != 'EUR'

Contains

The contains operator filter keys whose value contains the given literal.

It is applicable to array values.

economy.major_industries CONTAINS 'Tourism'

Not Contains

The not contains operator filter keys whose value does not contain the given literal.

It is applicable to array values.

economy.major_industries NOT CONTAINS 'Steel Production'

Boolean Operators

Operators above can be combined with AND and OR boolean operators to form compound filters.

country = 'Turkey' AND population > 10000000

Boolean operators can be grouped with parentheses to have higher precendence.

country = 'Turkey' AND (population > 10000000 OR is_capital = false)

When no parentheses are provided in ambigous filters, AND will have higher precendence than OR. So, the filter

country = 'Turkey' AND population > 10000000 OR is_capital = false

would be equivalent to

(country = 'Turkey' AND population > 10000000) OR is_capital = false

Filtering Nested Objects

It is possible to filter nested object keys by referencing them with the . accessor.

Nested objects can be at arbitrary depths, so more than one . accessor can be used in the same identifier.

economy.currency != 'USD' AND geography.coordinates.latitude >= 35.0

Filtering Array Elements

Apart from the CONTAINS and NOT CONTAINS operators, individual array elements can also be filtered by referencing them with the [] accessor by their indexes.

Indexing is zero based.

economy.major_industries[0] = 'Tourism'

Also, it is possible to index from the back using the # character with negative values. # can be thought as the number of elements in the array, so [#-1] would reference the last character.

economy.major_industries[#-1] = 'Finance'

Miscellaneous

  • Identifiers (the left side of the operators) should be of the form [a-zA-Z_][a-zA-Z_0-9.[\]#-]*. In simpler terms, they should start with characters from the English alphabet or _, and can continue with same characters plus numbers and other accessors like ., [0], or [#-1].
  • The string literals (strings in the right side of the operators) can be either single or double quoted.
  • Boolean literals are represented as 1 or 0.
  • The operators, boolean operators, and boolean literals are case insensitive.