This repository contains the material for the "Knowledge Graphs Course" erogated by Cefriel. An example scenario (Competence Registry) is considered to implement and discuss modelling and querying over relational databases, graph databases and triplestores. The main topics are:
- Relational Databases using Postgres (CSV import and SQL queries)
- Graph Database using Neo4j (Graph modelling, CSV import and Cypher queries)
- RDF Database using rdflib (KG construction from CSV file, SPARQL queries)
- [Bonus] Natural language QA using LangChain (querying different databases using the OpenAI API)
A company would like to keep a record of the digital competences possessed by each employee. Each employee is described by their first name, last name, and the department they belong to. Each competence defined by the DigComp framework is described through four main fields: identifier, dimension, name, and description. Each employee can be associated with one or more competences. For each competence association, the following information can be provided: competence level (from 1 to 5) and the person's interest in gaining/increasing the skill.
The repository defines a docker-compose.yml file to run all the databases and the Jupyter notebooks as containers via Docker. The containers can be run all at once or separately.
The notebooks can be executed running the competence-kg-notebook container with the command:
docker-compose up --force-recreate competence-kg-notebookThe data are randomly generated via the script available in the notebooks/competence-kg-generate.ipynb notebook. A constant seed is currently defined to always generate the default data contained in the data folder.
The files employees.csv, assignment.csv and competences.csv contain the data. The *-iri.csv files are generated to support the discussion on the RDF graph.
The relational database is modelled according to the ER diagram defined in the model folder. The diagram can be edited using the draw.io webapp.
An init.sql file is defined to initialise a Postgres database from the CSV files in the data folder. The database can be executed using the command:
docker-compose up --force-recreate competence-kg-postgresNote that if the option --volumes is not used with the command docker-compose down the status of the database is automatically persisted by Docker.
The competence-kg-postgres.ipynb notebook contains the code to inspect and query the database using the SQL syntax.
The graph database is modelled according to the diagram defined in the model folder. The diagram can be edited using the arrows.app webapp.
The competence-kg-neo4j.ipynb notebook contains the code to import the data and to inspect and query the database. A Neo4j database is instantiated using the CSV files in the data folder and according to the graph model defined. The Cypher language is used to query the database.
The database can be executed using the command:
docker-compose up --force-recreate competence-kg-neo4jThe status of the database is automatically persisted in the neo4j folder as defined in the docker-compose.yml file. The folder can be deleted to erase all the data in the database.
The graph database is modelled according to the ontology ontology.ttl defined in the model folder and available online at https://knowledge.c-innovationhub.com/competence-kg/schema. The ontology reuses the foaf vocabulary and adds further classes and properties. The ontology can be visualised using the WebVOWL webapp. In the model folder the file content-negotiation.sh shows how different cURL requests return different representations of the ontology.
The competence-kg-rdf.ipynb notebook contains the code to construct the knowledge graph and to inspect and query the it. An in-memory database is instantiated using the rdflib library and the CSV files in the data folder. In this case, the *-iri.csv files are considered instead of the assignment.csv and competences.csv files. Indeed, we leverage the SKOS vocabulary describing the DigComp framework (already available online) to enrich the company's graph without having to instantiate from the CSV all the data about the competences.
The mappings from the CSV files to the ontology are defined using YARRRML, compiled to RML using the yatter tool and executed using the morph-kgc processor. The mappings are available in mappings folder. The SPARQL language is used to query the database.
The generated RDF graphs are serialised in the Turtle format and saved in the rdf folder together with a dump of the DigComp SKOS vocabulary.
The competence-kg-langchain.ipynb notebook provides the code to execute QA (i.e., natural language querying) over the three databases leveraging the langchain library and the OpenAI GPT models. The SQLDatabaseChain, GraphCypherQAChain, GraphSPARQLQAChain are used in the code.
A credentials.json file should be provided in the main folder with a valid key for the OpenAI API.
{
"OPENAI_API_KEY": "PUT_HERE_YOUR_KEY"
}