Yes, please open an issue 😄

Added a couple comments, although since this is a draft PR you are probably addressing them already.

Working on this PR made me wonder: why do we integrate Jinja2 (or any specific templating engine) into Outlines? I mean, we could have just provided a Prompt interface that takes an already rendered str (and eventually keep a function decorator if that sounds fancy and helps better infer the template arguments), leaving the templating up to the user. That way, users could choose a different templating engine if they wanted, and e.g., how to use Outlines in combination with Jinja2 could be rather provided as an example in the documentation. WDYT?

I am not sure what you mean by "taking an already rendered str", could you provide an example?

I am not sure what you mean by "taking an already rendered str", could you provide an example?

TL;DR: I mean passing the output/result of the templating process to the Prompt constructor, thereby externalizing the templating from Outlines. IMHO, this would lead to better separation of concerns and make Outlines lighter and simpler. One idea could be to maintain tighter integration with templating only as an optional feature, if we believe it simplifies the workflow and improves the developer experience.

import outlines

# Using Jinja2 ...
import jinja2

env = jinja2.Environment(loader=jinja2.FileSystemLoader("."))
template = env.get_template("prompt.jinja2")

# ... or, Using Mako ...
import Mako

template = mako.template.Template(filename="prompt.mako")

@outlines.prompt
def fancy(examples, question):
    return template.render(examples=examples, question=question)

examples = [
    {"question": "What is the capital of France?", "answer": "Paris"},
    {"question": "What is 2 + 2?", "answer": "4"},
]
question = "What is the Earth's diameter?"

prompt = fancy(examples, question)

# ... but why not just?
rendered = template.render(examples=examples, question=question)
prompt = outline.Prompt(rendered)
# EDIT: Nevermind, while writing this down, I realized that this would
# essentially mean that arguments are evaluated too early, leaving the prompt
# without any arguments and simply storing a `str`...

You can already use any templating engine you want with Outlines, by rendering the template and passing the returned string to a generator:

import outlines

model = outlines.models.openai("gpt-4o")

# Input

examples = [
    {"question": "What is the capital of France?", "answer": "Paris"},
    {"question": "What is 2 + 2?", "answer": "4"},
]
question = "What is the Earth's diameter?"


# Using Jinja2 ...
import jinja2

env = jinja2.Environment(loader=jinja2.FileSystemLoader("."))
template = env.get_template("prompt.jinja2")
rendered = template.render(examples=examples, question=question)

generator = outlines.generate.text(model)
results = generator(rendered)


# ... or, Using Mako ...
import Mako

template = mako.template.Template(filename="prompt.mako")
rendered = template.render(examples=examples, question=question)

generator = outlines.generate.text(model)
results = generator(rendered)

The role of the Prompt object is to wrap a Jinja2 template so that it returns a string when called with the template's variable values. I chose Jinja2 to make the common path easier; users of outlines are not necessarily familiar with templating engines and Jinja2 seemed like a reasonable default. As I demonstrated above you can always use another templating engine, at the cost of extra complexity.

The confusion may come from poor naming on my part; this object should probably be called Template instead.

Thank you very much @rlouf for the clarification. I'm not sure about the naming either, but if it's extensible, that already sounds like a great design! Apologies for sharing ideas maybe a bit too quickly as they pop from my mind, I probably should have checked the documentation first!

(Do not merge this PR yet, I will rewrite the commit history a bit once we've settled on deprecating or removing outlines.prompts.render!)

Help! I don't understand the failed test coverage. It says that only line 90 of prompt.py is not covered, which doesn't make sense to me…

It seems that tests don't cover that part of the flow, where if is True and cleaned_template += "\n" is actually being executed:

        ends_with_linebreak = content.replace(" ", "").endswith("\n\n")
        if ends_with_linebreak:
            cleaned_template += "\n"

This is one of the cases, I suppose, where one of the old render tests would exactly cover (I would integrate them all into template_from_str):

    tpl = """
        A test line
            An indented line

    """
    assert render(tpl) == "A test line\n    An indented line\n"

(Do not merge this PR yet, I will rewrite the commit history a bit once we've settled on deprecating or removing outlines.prompts.render!)

I've cleaned up the git history, so it's ready to be merged now :)

Great job! 🎉