KGraphLang is a query language designed for LLMs to query Knowledge Graphs.
KGraphLang uses predicates to access knowledge graph nodes and edges, and provides a comprehensive syntax to compose predicates and generate query results.
KGraphLang is both human friendly and easy for LLMs to generate.
Features include:
- Defining predicates which can be implemented via code or queries to an underlying data source.
- Annotating predicates with extra-logical inputs like @topk(10) to control predicate output
- Aggregations for collections, count, sum, average, max, min
- Math functions for add, subtract, multiply, divide
- Base datatypes: string, number, boolean
- Data types for time, currency, geolocation, units, URIs
- Complex data types for List, Map
- Single and multi-line comments
KGraphLang is used with Ensemble Reasoning to implement the KGraphService Ensemble Member (EM).
Using Ensemble Reasoning a reasoning LLM generates KGraphLang queries and processes results during the inference. This avoids the latency of "tool calls" by integrating the knowledge graph directly with the reasoning LLM's model.
A KGraphLang expression always terminates with a period (.).
Example:
?x = 10, ?y = 20.
Logical expressions can use AND (,) and OR (;) to combine multiple statements:
- AND expressions:
?x = 10, ?y = 20.
- OR expressions:
?x = 'Alice'; ?y = 'Bob'.
Variables start with a question mark (?):
?variable_name
Use is to assign arithmetic expressions to variables:
?total is ?price * ?quantity
Use = for unification:
?status = 'active'
Supports arithmetic operations: addition (+), subtraction (-), multiplication (*), and division (/):
?result is ?a + ?b * 2
Use parentheses () to group expressions and control evaluation order:
(?a + ?b) * 2
Collections are enclosed in brackets ([ ]) and represent lists or maps:
- List:
[1, 2, 3, ?var]
- Map:
['key1' = 'value1', 'key2' = ?var]
- Nested Maps:
[ 'outer' = [ 'inner_key' = [ 1, 2, 3 ] ] ]
Booleans are represented as true or false:
?is_enabled = true
Supports comparisons: >, <, >=, <=, ==, !=:
?score >= 85
- Membership (
in):
?item in ['apple', 'banana']
['key' = 'value'] in ?map
- Subset (
subset):
['a', 'b'] subset ['a', 'b', 'c']
['key1' = 'value1'] subset ?larger_map
Use not to negate expressions:
not(?status = 'completed')
- Predicates:
hasRole(?user, 'Admin')
Nested predicate calls within arguments are invalid:
child(?x, father(?y)) // INVALID
Valid usage:
hasRole(?user, 'Admin')
- Annotations (
@):
@topk(10) entity_search(?name)
Explicitly typed literals include:
-- Date, DateTime, Time, and Duration:
'2023-02-18'^Date
'2023-02-18T14:30:00'^DateTime
'14:30:00'^Time
'P2Y4M'^Duration
- GeoLocation:
'40.7128,-74.0060'^GeoLocation
- Currency and Unit:
'10.00'^Currency(USD)
'100'^Unit('http://qudt.org/vocab/unit/kg')
- URI:
'http://example.org/resource'^URI
Aggregation functions summarize query results:
sum{ ?amount | ?amount in ?fees }
Supported aggregates: collection, set, average, sum, min, max, count.
Strings use single, double, or triple-double quotes:
'simple string'
"Another string"
"""Multiline
string"""
- Single-line (
//):
// This is a comment
- Multi-line (
/* */):
/* Multi-line
comment */
A comprehensive query demonstrating multiple language features:
// define property URIs we want to retrieve
?uri_prop = 'urn:uri_prop'^URI,
?name_prop = 'urn:name_prop'^URI,
?email_prop = 'urn:email_prop'^URI,
?prop_list = [ ?uri_prop, ?name_prop, ?email_prop ],
// get a list of person URIs from the KG
person_uri_list(?PersonList),
// create a collection of People, with each person
// having a map containing the properties we want
?PersonEmailMapList = collection {
?PersonMapRecord |
?Pid in ?PersonList,
// get the person info from the KG
get_person_map(?Pid, ?prop_list, ?PersonMapRecord)
}.
// now the LLM has the desired info from the KG which it can
// use for the next step in the reasoning process...
KGraphLang enables expressive and intuitive querying of Knowledge Graphs.