Master Thesis Project

Master Thesis project as part of a Double Master Degree at the University of Camerino and the University of Applied Sciences of Northwestern Switzerland.

Abstract

The success of research project proposals heavily depends on the formation of an effective consortium, ensuring the inclusion of experienced researchers aligned with the topics of research calls, such as those in the EU's Horizon Europe program.

One of the most challenging aspects of consortium formation is the identification of suitable research collaborators. Traditional approaches, which primarily rely on social networks and author citations, often fail to provide effective recommendations.

This thesis introduces an Agentic Graph Retrieval-Augmented Generation (RAG) method that enhances contextual and explainable collaborator recommendations. By leveraging Knowledge Graphs (KGs) and Large Language Models (LLMs), this method tailors recommendations based on researchers' areas of expertise and project relevance, surpassing traditional approaches in effectiveness.

The Design Science research methodology was followed to develop this approach, and the GPT-4o model was used for evaluation, demonstrating the method's capabilities in enhancing consortium formation.

---

Installation

Prerequisites

To run this project, ensure you have the following installed on your system:

• Docker (to containerize and manage services)
  • Please, ensure you have also installed Docker Compose V2
• GraphDB (for storing and querying the knowledge graph)

Clone the Repository

  git clone https://github.com/Piermuz7/MasterThesisProject.git

Streamlit Configuration

Create a .streamlit directory in the project root and add the following file:

  mkdir .streamlit
  touch .streamlit/secrets.toml

Add the following configuration to the secrets.toml file:

GRAPHDB_URL = "http://graphdb:7200/repositories/CORDIS-eurio-KG"
EURIO_ONTOLOGY_PATH = "ontology/EURIO.ttl"
GRAPHDB_USERNAME = "admin"
GRAPHDB_PASSWORD = "root"
AZURE_OPENAI_ENDPOINT = "<YOUR_AZURE_OPENAI_ENDPOINT>"
[api_key]
ANTHROPIC_KEY = "<YOUR_ANTHROPIC_API_KEY>"
AZURE_OPENAI_API_KEY = "<YOUR_AZURE_OPENAI_API_KEY>"

---

Importing the EURIO Knowledge Graph on GraphDB locally

1. Download the EURIO Knowledge Graph
2. Import it into GraphDB:

   • Open GraphDB Workbench (http://localhost:7200)

   • Create a new repository

   • Since the EURIO RDF file is large, you may need to increase the maxInMemorySize and maxUploadSize properties in GraphDB.

   • Go to META-INF GraphDB directory. On MacOS, it is located at '/Applications/GraphDB Desktop.app/Contents/app/lib/common/WEB-INF/classes/META-INF'.

     cd /Applications/GraphDB\ Desktop.app/Contents/app/lib/common/WEB-INF/classes/META-INF

   • Open properties.xml and set the following values:

     • <entry key="graphdb.workbench.maxInMemorySize">5368709120</entry>
     • <entry key="graphdb.workbench.maxUploadSize">5368709120</entry>

   • Load the EURIO RDF N-Quad file into the repository

• Due to the large size of the EURIO knowledge graph, it is recommended to import it locally first and then use Docker’s volume mapping to persist the data. This approach ensures better performance and avoids the need to repeatedly re-import the knowledge graph when restarting the container.

---

Quick Start

Under the project directory, you can use the following commands to build, run, and manage the project.

Build & Run the Project

  sudo docker compose up --build

This command builds the Docker images and starts all services. In particular, it starts the Streamlit app, GraphDB, Chroma vector database, and Ollama service.

Restart Services

  sudo docker compose restart

This restarts the Streamlit app, GraphDB, Chroma vector database, and Ollama service

Stop Services

  sudo docker compose stop

Stops all running containers without removing them.

Stop & Remove Containers

  sudo docker compose down

Stops and removes all containers and networks created by docker-compose.

---

Technology Stack

Component	Description
Docker	Manages containerized services
GraphDB	Stores and queries the knowledge graph
SPARQL	Query language for RDF-based knowledge graphs
Python	Core programming language
Streamlit	Provides an interactive user interface
GPT-4o	Large Language Model used for RAG-based recommendations
LangChain	Framework for integrating LLMs with external tools
LlamaIndex	Framework to handle agentic graph retrieval
Ollama	Service for generating embeddings from text
Chroma	Vector database for storing and querying embeddings
all-minilm:l6-v2	sentence-transformers model for generating embeddings

---

Evaluation

The evaluation of the Agentic Graph Retrieval-Augmented Generation (RAG) method was mainly conducted using the GPT-4o model. Some experiments were also conducted using the Claude 3.5 Sonnet model. The framework used for evaluation is RAGAS. Our evaluation aims to measure the relevance and consistency of the suggested contributors, ensuring that the retrieved knowledge is in line with user demands.

The results of the artifact evaluation are presented under the /evaluation directory.

---

License

This project is licensed under the MIT License. Feel free to modify and distribute it.

---

Contact

For any questions or contributions, feel free to contact:
Piermichele Rosati

Personal email: piermichele.rosati@gmail.com

Institutional emails:

• piermichele.rosati@studenti.unicam.it
• piermichele.rosati@students.fhnw.ch

GitHub: Piermuz7