Beyond Vector Search
Hybrid Vector/Knowledge Graph Query Engine for Enhanced Data Retrieval

Installation

We used Python==3.9.18, and we recommend using a virtual environment to install the required packages.

pip install -r requirements.txt

Creating filtered data

• notebooks/parsing_json.ipynb: filter data for arxiv
• notebooks/parsing_cnn_news.ipynb: filter data for CNN news
• notebooks/parsing_wiki_movies.ipynb: filter data for wiki movies

Dataset

We used a subset of the ArXiv dataset from Kaggle which contains 12,926 samples, and the the pickle file is provided at "beyond_vector_search/data/filtered_data.pickle".

Source Code

make_vectordb.py: a script to build a vector database from a "data/filtered_data.pickle"

utils/
    - build_graph.py: a script containing helper functions for building the knowledge graph
    - parse_arxiv.py: a script containing helper functions for parsing the arxiv dataset
vector_graph/
    - bipartite_graph_dict.py: A custom implementation of the bipartite graph
    - bipartite_graph_networkx.py: An experimental implementation of the bipartite graph using networkx
    - embedding_models.py: A custom implementation of the embedding models for generating the text embeddings
workloads
    - keyword_extractor.py
    - query_gen.py: A script for generating the text queries given paper data points
    - workload_gen.sh: This is the script for generating the workloads we described in the report
testing
    - inference.py: A script for executing our various search query engines on the generated workloads
zy_testing
    - compute_metrics_cos.py: A script for computing the accuracy of our results utilizing various performance compute_metrics_cos