Closed rlouf closed 1 year ago
As a non-expert practitioner, I started testing outlines today, and for the first time (after many months of postponing it) gave it a try with HF transformers (once I realized that the regex method did not work with OpenAI).
On the other hand, outlines also offers very helpful features (prompt Jinja templating, "is_in"), albeit probably those are less unique (see LMQL).
However myself, I would be happy to use the above features when querying the OpenAI API. Is it your view that, since there are other repositories that also provide elegant prompting, we should use other libraries for that use-case and let outlines focus on the neuro-symbolic
problem?
We might keep the base OpenAI calling feature (it can be useful for comparison, after all), but remove everything that resembles an attempt to do guided generation with it. Does that make sense?
That sounds fair to me. And I think maintaining basic compatibility with openAI would give this project more visibility while getting people to try HF transformers to get all the benefits of this library.
What behavior of the library made you think about the improvement?
OpenAI is not compatible with the main offering of the library: guided generation.
How would you like it to behave?
Not having any OpenAI integration at all.