Architectural diagrams with LLMs (ChatGPT)
Over the weekend, I reviewed internal architectural documents and came across instructions on how to write RFC/ADR correctly and how to visualize architectural diagrams using chatbots. The instructions were good and looked like this.
Draw a diagram on plantuml.
There's a postgres base. It has a table called XXX. The data in this tablet XXX writes frontend YYY by referring to the backend ZZZ working with this database. There's an ETL process that puts this table into a clickhouse WWW, into a VVV table. With Jupyter, we connect to DWH and analyze this VVV table.
In principle, such a macaroma can be asked to generate LLM. Class diagram, component diagram, sequence or activity if we need UML diagrams C4 Model: Context, Containers, Components, and Code
- any other diagrams: ER, DFD, IDEF, BPMN, ...
However, in this instruction, the authors forgot that the output of the LLM sometimes gets thrash, which means you need to know these notations yourself and be able to read them, and it is also desirable to model. As a result, visualizing with ChatGPT is fun, but visualization should help + be accurate enough. And in order to check how good it is, I would recommend not just to command the chatbot, but also to study the documentation of the plate modeling, such as:
- Standard for UML
- Platuml documentation
- Documentation for the C4 Model
- And any other documentation on the modeling notation you need
Without such basic knowledge, it is difficult to validate what LLM gave out.
#Software #SoftwareArchitecture #Architecture #Engineering #SoftwareDevelopment