The descriptions of all the objects should be walked to form a nested markdown file that mirrors the filesystem. Only use the covenants, not every single object. We would represent items in a collection as their parent plus an example of one child.
This manual would be regenerated by AI to cover the uses described in all the user stories that assailed the object. The manual would get updated with a description that would have been better for the AI to use, with the assessor deciding if an improvement is good enough to warrant a change.
The change assesor should look at everything that came before, so that the chosen output isn't just better than only some - it must be better than all.
This file is transformed to be a fact list style document that performs well under embedding.
This would then be used by HAL as a way to figure out what path needs to changed to. This file is updloaded with the HAL assistant.
We might get better results if the manual was generated, and then a fact sheet was made out of which pointed to sections in the manual. Possibly parsing it with a set of rules to make it best for embedding, like correcting the language to use the smallest number of definitions, generating a dictionary at the end, and a TOC at the front.
The descriptions of all the objects should be walked to form a nested markdown file that mirrors the filesystem. Only use the covenants, not every single object. We would represent items in a collection as their parent plus an example of one child.
This manual would be regenerated by AI to cover the uses described in all the user stories that assailed the object. The manual would get updated with a description that would have been better for the AI to use, with the assessor deciding if an improvement is good enough to warrant a change.
The change assesor should look at everything that came before, so that the chosen output isn't just better than only some - it must be better than all.
This file is transformed to be a fact list style document that performs well under embedding.
This would then be used by HAL as a way to figure out what path needs to changed to. This file is updloaded with the HAL assistant.
We might get better results if the manual was generated, and then a fact sheet was made out of which pointed to sections in the manual. Possibly parsing it with a set of rules to make it best for embedding, like correcting the language to use the smallest number of definitions, generating a dictionary at the end, and a TOC at the front.