epilys / knowledge-graph-language

A query language for exploring knowledge graphs.

Knowledge Graph Language (KGL)

Knowledge Graph Language is a query language for interacting with graphs. It accepts semantic triples (i.e. ("James", "Enjoys", "Coffee")), indexes them, and makes them available for querying.

Here is a demo of the language in action:

kgl-interpreter-demo.mp4

You can use this language to:

• Find all attributes associated with a node in a graph.
• Return all nodes that are connected to a node.
• Return all nodes that are connected to a node and meet a specified condition.
• Find how two nodes connect in a graph.

This README serves as the official language reference.

The following blog posts discuss the design of the KGL language and the interpreter, respectively:

• Designing a knowledge graph query language
• Designing an interpreter for Knowledge Graph Language (KGL)

You can try the language on a knowledge graph calculated from James' Coffee Blog.

Syntax

Query a Single Item

You can query a single item:

{ James }

This will return all items associated with the James entry:

{'Birthday': ['March 20th, 2024'], 'WorksFor': ['Roboflow', 'PersonalWeb', 'IndieWeb'], 'Enjoys': ['Coffee'], 'Hobbies': ['Making coffee']}

Sequential Queries

The Knowledge Graph Language flows from left to right. You can make a statement, then use an arrow (->) to query an attribute related to the result:

Consider the following query:

{ James -> WorksFor -> Makes }

This query gets the James item, retrieves for whom James works, then reports the Makes attribute for the employer.

The query returns:

['Computer vision software.']

Filter Queries

Important

This feature is not fully implemented.

You can filter queries so that the flow of data is constrained to only work with results that match a condition.

Consider this query:

{ Roboflow -> WorksFor ("Enjoys" = "Coffee") -> Hobbies }

This query gets the Roboflow node. Then, the query gets everyone who works at Roboflow who enjoys coffee. The query then finds who everyone works for, and returns their hobbies.

This returns:

['Making coffee']

You can filter by the number of items connected to a node in the graph, too.

Consider these triples:

("CLIP", "isA", "Paper")
("CLIP", "Authors", "Person 1")
("CLIP", "Authors", "Person 2")
("Person 1", "Citations", "Paper 42")
("Person 2", "Citations", "Paper 1")
("Person 2", "Citations", "Paper 2")
("Person 2", "Citations", "Paper 3")
("Person 2", "Citations", "Paper 4")

Suppose you want to find all authors of the CLIP paper in a research graph, but you only want to retrieve authors whose work has been cited at least three times. You can do this with the following query:

{ CLIP -> Authors ("Citations" > "3") }

This query returns:

['Person 2']

This is because only Person 2 has greater than three citations to their works.

Describe Relationships

Suppose you want to know how James and Roboflow relate. For this, you can use the interrelation query operator (<->).

Consider this query:

{ Roboflow <-> James }

This returns:

['Roboflow', ('James', 'WorksFor')]

Serialized into Knowledge Graph Language, this response is represented as:

Roboflow -> WorksFor

If we execute that query in introspection mode, we can see all information about James:

{ Roboflow -> WorksFor }!

This returns:

[{'James': {'Birthday': ['March 20th, 2024'], 'WorksFor': ['Roboflow', 'PersonalWeb', 'IndieWeb'], 'Enjoys': ['Coffee'], 'Hobbies': ['Coffee']}}, {'Lenny': {'WorksFor': ['MetaAI', 'Roboflow']}}]

Introspection

By default, all Sequential Queries return single values. For example, this query returns the names of everyone who works at Roboflow:

{ Roboflow -> WorksFor }

The response is:

['James', 'Lenny']

We can enable introspection mode to learn more about each of these responses. To enable introspection mode, append a ! to the end of your query:

{ Roboflow -> WorksFor }!

This returns all attributes related, within one degree, to James and Lenny, who both work at Roboflow:

[{'James': {'Birthday': ['March 20th, 2024'], 'WorksFor': ['Roboflow', 'PersonalWeb', 'IndieWeb'], 'Enjoys': ['Coffee'], 'Hobbies': ['Coffee']}}, {'Lenny': {'WorksFor': ['MetaAI', 'Roboflow']}}]

Description Operators

By default, Knowledge Graph Language returns the value associated with your query. You can add operators to the end of your query to change the output.

You can use:

• ? to return True if your query returns a response and False if your query returns no response.
• # to count the number of responses
• ! to return an introspection response.

Data Format

This project allows you to index triples of data like:

("James", "Enjoys", "Coffee")
("James", "Hobbies", "Making coffee")
("James", "WorksFor", "Roboflow")
("Roboflow", "Makes", "Computer Vision")
("Roboflow", "EntityType", "Company")

A graph is then constructed from the triples that you can then query.

Python API

First, install KGL:

pip install kgl

Create a Knowledge Graph

from kgl import KnowledgeGraph

kg = KnowledgeGraph()

Ingest Items

You can ingest triples of strings:

kg.add_node(("Lenny", "Owns", "Roboflow"))

You can also ingest triples whose third item is a list:

kg.add_node(("Alex", "Citations", ["MetaAI", "GoogleAI", "Coffee", "Teacup", "Roboflow"]))

Add Data in Native KGL

You can add a relation in native KGL using the following syntax:

{ subject, predicate, object }

Here is an example:

{ Lenny, Owns, Roboflow }

Ingest from CSV

To load data from a CSV file, use:

kg = KnowledgeGraph().load_from_csv("./file.csv")

Each line in the CSV file should use the structure:

subject, predicate, object

There should be no header row in your file.

Ingest from TSV

To load data from a TSV file, use:

kg = KnowledgeGraph().load_from_tsv("./file.tsv")

Each line in the TSV file should use the structure:

subject  predicate  object

There should be no header row in your file.

Ingest from JSON

To load data from a JSON file, use:

kg = KnowledgeGraph().load_from_json_file("./file.json")

The JSON file should have the structure:

[
  {
    "Entity": "Tessa Violet",
    "is": "a singer",
    "wrote": ["Games", "Haze"]
  }
]

Entity is a reserved key (case sensitive) that states the entity to which all pairs in a specific JSON entry.

This will be converted into the triples:

("Tessa Violet", "is", "a singer")
("Tessa Violet", "wrote", ["Games", "Haze"]

These triples will then be ingested into a KGL data store.

Evaluate a Query

result, query_time = kg.evaluate("{ James }")
print(result)
print(query_time)

Responses are valid Python objects, whose type varies depending on your query.

By default, KGL returns a list.

But:

• ! queries return dictionaries.
• # queries return integers.
• ? queries return booleans.

Run the Web Interpreter

To run the KGL web interpreter, first generate a knowledge graph in a CSV file. Each line should use the format subject, predicate, object. Then, open app/app.py and replace the all1.csv reference with the name of your CSV file. Then, install the required dependencies:

cd app/
pip3 install -r requirements.txt

To run the web interpreter, execute the following command:

python3 app.py

The interpreter will open at http://localhost:5000.

Tests

To run the project test suite, run:

pytest test/test.py

License

This project is licensed under an MIT license.