Diagrams AI Can, and Cannot, Generate (Category Architecture)

I recently made it to autumn. articles from 2024 A year about drawing architectural diagrams using AI. The author reviewed three use cases and ended with three problems that prevent AI from drawing diagrams like a god. Here are three scenarios. 1) Generate generic diagrams on a given topic - I do not quite understand why such diagrams are needed by the author except as illustration examples, but AI copes with their drawing, however, Google image search copes even better. 2) Whiteboarding with AI When we draw a planned system for which we want to come up with a solution and consider possible problems. In principle, adapting the prompts with God's help and a strong word you can get something digestible, but sometimes it is easier to take a conditional DSL language like Plantuml and write with pens what exactly you need to correct:) But here AI in principle can help well. You can read a book about different approaches to drawing.Creating Software with Modern Diagramming Techniques"which I'm talking about." told 3) System diagramming with AI The point here is to reconstruct the architectural diagram from the project code. It is often difficult to get something meaningful here, as there may be a lot of code that is not clear in it, and AI can hallucinate something that is not. In general, if the project is not super-simple, it is difficult to get something sane. And the author of the article describes the key problems, among which three stand out. There is little training data in the form of projects with a codebase + diagrams that they are described not in generic form, but in detail. Problems with code analysis, where there are different languages and technologies, there is code separation and configurations, as well as a common porridge of dependencies between components on static diagrams (deployment diagrams, component diagrams), as well as porridge in dynamic charts (sequence, activity) due to the number of brunches with alternative logic The problem with the essence and focus on the main thing is to create good diagrams in order to highlight the main thing, and for this you need to understand the purpose of the system and its main parts. This is not always reflected in the code. Perhaps this can be fed as an additional AI context to an agent – something close to this is discussed in the whitepaper.From Requirements to Architecture: An AI-Based Journey to Semi-Automatically Generate Software Architecture"which I mean." told earlier

In general, the miracle did not happen, the silver bullet has not yet appeared, we continue to work and prepare robots that will clothe life with leather bags:)

#AI #Engineering #ML #Software #Architecture