Starting with this commit and in the remaining two folders, I translate the documents in the order they appear in the documentation. Should have done that from the beginning, as it helps the LLM learning things (I guess).
I created this before with GPT-4o, twice as much changes, plus it did not understand some of my rule definitons (Though the prompt was different). So I reverted everything and tried the same with GPT-5. This diff is a joy, only the relevant changes.
This commit started as an attempt to handle translation of abbr elements, which I was not able to get to work in a clean way. Shortly before I resigned I had the idea to try `openai:gpt-5` as agent in the translate.py, instead of `openai:gpt-4o` and BAM it worked. A lot of other things also worked much better. Modifications had to be made though because GPT-5 interpreted some rules stricter, so some reformulation and some more examples were needed. Will translate with that agent in the remaining translations to test if these prompts work well.
As it did so once (like in the example). This is actually super smart, the anchor exists (is the LLM caching the Anchors of headings in other documents?!), but I guess this possibly creates confusion, so avoid.
The LLM wrote `Pfadparameter`, which is not bad, but let's use or not use a dash unison for all FastAPI specific path operation parameters. As the translation so far used one, I now defined it to be so.
There are some instances of using `<code>...</code>` instead of backticks in the English document, which I fixed on the way, as it will not require a retranslation of other languages.
Removes one level of Use–mention distinction [1] by writing the examples in separate paragraphs, instead of wrapping them in backticks. This because the LLM still seems to ignore the rules sometimes, here it ignored the rule to translate the content of title attribute of abbr elements.
[1]: https://en.wikipedia.org/wiki/Use%E2%80%93mention_distinction
Before it seemingly randomly applied them, indicating that it has not understood them the way they were formulated.
Change two related things in two translations.
Remove the part to not translate link targets, as this may interfere with the more precise rulings in the general prompt.
..., which tells the LLM to fix wrongly translated anchor parts in links.
Back in those days I translated these parts, to link to the correct German heading, which then had a different anchor than the English heading. But now, as we use the same anchors
documentation-wide, this needs to be undone. The LLM missed some of these translated anchors, so I added this rule.
I will check via script, when gone through the whole documentation, if all those anchor parts are in sync and valid.
In this document it added language codes to urls starting with "https://sqlmodel.tiangolo.com/", resulting in broken links.
So I changed the instructions in the translate.py, so that it only adds the language code when the URL starts exactly with "https://fastapi.tiangolo.com".
Nils Lindemann