Would it be a good idea to show examples integrating dspy with outlines?

[torchss]

outlines does a phenomenal job of structured outputs and feels like time travelling into the future especially when using self hosted models. However, the prompting experience still reminds me of 2023.

I'm wondering:

• whether showing examples integrating dspy with outlines would be a value add for structured output
• if so, are there samples that show how to integrate dspy with outlines?
• if not, I would really appreciate some insights and recommendations for superior approaches

I'm unsure if #1495 , #502 are related.

[RobinPicard]

Hi @torchss! Those issues are not related to DSPy, but improving the prompting experience is something on our minds. As far as I can tell, there has not been anything done related to DSPy in Outlines. If you feel like working on an example on how to integrate the use of the 2 libraries that would be very valuable! I think that it could help us think about new features to add to Outlines.

[torchss]

@RobinPicard Thank You for looking at this. I would love to hear more thoughts about adding Outlines to DSPy - Omar has had thoughts about that too as well and I believe he has spoken to you about it?

Here's one implementation that's obviously outdated: https://github.com/Scale3-Labs/dspy-examples/blob/main/src/structured_output/outlines_lm.py

Before I/We/anyone began work on adding Outlines v1 to DSPy 3.0, I wanted to share these observations:

• These models generate pretty solid JSON responses already in DSPy 3 without adding Outlines: HuggingFaceTB/SmolLM3-3B, THUDM/SWE-Dev-7B, Qwen/Qwen3-4B (and the larger models)
• Sometimes these models output large strings and run out of tokens (when adding Outlines, I have empirically noticed that the probability this will happen is much lower)
• Smaller models struggle to generate valid JSON responses out of box. For example: Qwen/Qwen3-1.7B (or lower), tinyllama-1.1b-chat-v1.0, qwen2.5-coder-0.5b-instruct (when adding Outlines, I have empirically noticed that this is no longer an issue, i.e., valid JSON is always produced. Quality of response and resilience to prompt injection are separate matters).

Would love to hear your thoughts and recommendations on how to make DSPy 3 based applications more effective on smaller models.