Annotation improvements - part 2

[prabhu]

Adds ml-tiny profile to trim the bom size further. Gemini, unfortunately, still complains, but is sorta working.

[prabhu]

few-shot prompting

bom.json

Gemini failed to detect the BOM type despite offering a system prompt about annotations. Unfortunately, annotations is an array of object, where the text attribute could contain interesting information. This nested structure is breaking some models.

By simply copying the annotations text from the file and using it as a prompt (few-shot prompt), we get high quality result.

[prabhu]

Crypto properties are getting understood properly.

[prabhu]

meta-prompting

By introducing the definition of OBOM along with the file, the model seems to be focusing on the correct attributes.

obom-windows.json

I was hoping the model would infer windows_programs from the word programs. After an initial hiccup, it has still managed to answer which is cool.

With REPL, we can do .search windows_programs, although the columns cannot be specified dynamically.

The solution is not magical though. When asked for "windows drivers", it couldn't figured out the tag windows_drivers.

Our REPL does work correctly here.

Despite my best attempts, I couldn't get the model to return all results correctly. This must be due to tags being an array.

• Need to figure out a prompting technique for OBOM tags.

[prabhu]

obom-linux.json

This Operations Bill-of-Materials (OBOM) document was created on Wednesday, November 13, 2024 with cdxgen. The lifecycles phases represented are: pre-build, post-build, and operations. The document describes an operating system named 'AlmaLinux' with the build name 'AlmaLinux release 9.4 (Seafoam Ocelot)'. The system appears to be set up for remote development, with 21 Visual Studio Code extensions installed. In addition, there are 1057 applications installed on the system.

Including the question along with the meta-prompt appears to be less effective.

The below answer, for example, is not accurate since those packages can be used by IT for deploying applications.

Simply separate out the initial prompt and the question improves quality.

Presence of VS Code extensions was correctly used to answer the question.