This solution does not have a pre-defined data structure. The JSON documents in Cosmos DB can be structured as per general NoSQL design best practices.
This is because the graph, in the CosmosAIGraph solution, only exists in-memory within the Apache Jena process, while all of the source data for it comes from Cosmos DB during initialization.
For example, one way to model this source data is to use an edges attribute in JSON documents, and populate it with the list of outgoing edges (relationships) from the given a node (entity) document.
In the case of the CosmosAIGraph reference dataset of python libraries, however, more specifically-named developers and dependency_ids JSON fields are used to construct the in-memory RDF graph:
...
"developers": [
"[email protected]"
],
...
"dependency_ids": [
"pypi_asgiref",
"pypi_blinker",
"pypi_click",
"pypi_importlib_metadata",
"pypi_itsdangerous",
"pypi_jinja2",
"pypi_python_dotenv",
"pypi_werkzeug"
],
...
The impl\data directory in this repo contains a curated set of PyPi (Python) library JSON documents.
This domain of software libraries was chosen because it should be relatable to most customers, and it also suitable for Bill-of-Materials graphs.
The PyPi JSON files were obtained with HTTP requests to public URLs such as https://pypi.org/pypi/{libname}/json, and their HTML contents were transformed into JSON.
Subsequent data wrangling fetched referenced HTML documentation, produced text summarization with Azure OpenAI and semantic-kernel and produced a vectorized embedding value from several concatenated text attributes within each library JSON document. A full description of this data wrangling process is beyond the scope of this documentation, but the process itself is in file 'impl/app/wrangle.py'.