This is a seriously beautiful guide. I really appreciate you putting this together! I especially love the tab-through animations on the various pages, and this is one of the best explanations that I've seen. I generally feel I understand grammar-constrained generation pretty well (I've merged a handful of contributions to the llama.cpp grammar implementation), and yet I still learned some insights from your illustrations -- thank you!
I'm also really glad that you're helping more people understand this feature, how it works, and how to use it effectively. I strongly believe that structured outputs are one of the most underrated features in LLM engines, and people should be using this feature more.
Constrained non-determinism means that we can reliably use LLMs as part of a larger pipeline or process (such as an agent with tool-calling) and we won't have failures due to syntax errors or erroneous "Sure! Here's your output formatted as JSON with no other text or preamble" messages thrown in.
Your LLM output might not be correct. But grammars ensure that your LLM output is at least _syntactically_ correct. It's not everything, but it's not nothing.
And especially if we want to get away from cloud deployments and run effective local models, grammars are an incredibly valuable piece of this. For practical examples, I often think of Jart's example in her simple LLM-based spam-filter running on a Raspberry Pi [0]:
> llamafile -m TinyLlama-1.1B-Chat-v1.0.f16.gguf \
> --grammar 'root ::= "yes" | "no"' --temp 0 -c 0 \
> --no-display-prompt --log-disable -p "<|user|>
> Can you say for certain that the following email is spam? ...
Even though it's a super-tiny piece of hardware, by including a grammar that constrains the output to only ever be "yes" or "no" (it's impossible for the system to produce a different result), then she can use a super-small model on super-limited hardware, and it is still useful. It might not correctly identify spam, but it's never going to break for syntactic reasons, which gives a great boost to the usefulness of small, local models.
What does it do when the model wants to return something else, and what's better/worse about doing it in llamafile vs whatever wrapper that's calling it? How do I set retries? What if I want JSON and a range instead?
This is a fantastic guide! I did a lot of work on structured generation for my PhD. Here are a few other pointers for people who might be interested:
Some libraries:
- Outlines, a nice library for structured generation
- https://github.com/dottxt-ai/outlines
- Guidance (already covered by FlyingLawnmower in this thread), another nice library
- https://github.com/guidance-ai/guidance
- XGrammar, a less-featureful but really well optimized constrained generation library
- https://github.com/mlc-ai/xgrammar
- This one has a lot of cool technical aspects that make it an interesting project
Some papers:
- Efficient Guided Generation for Large Language Models
- By the outlines authors, probably the first real LLM constrained generation paper
- https://arxiv.org/abs/2307.09702
- Automata-based constraints for language model decoding
- A much more technical paper about constrained generation and implementation
- https://arxiv.org/abs/2407.08103
- Pitfalls, Subtleties, and Techniques in Automata-Based Subword-Level Constrained Generation
- A bit of self-promotion. We show where constrained generation can go wrong and discuss some techniques for the practitioner
- https://openreview.net/pdf?id=DFybOGeGDS
Some blog posts:
- Fast, High-Fidelity LLM Decoding with Regex Constraints
- Discusses adhering to the canonical tokenization (i.e., not just the constraint, but also what would be produced by the tokenizer)
- https://vivien000.github.io/blog/journal/llm-decoding-with-regex-constraints.html
- Coalescence: making LLM inference 5x faster
- Also from the outlines team
- This is about skipping inference during constrained generation if you know there is only one valid token (common in the canonical tokenization setting)
- https://blog.dottxt.ai/coalescence.html
Question for the well-informed people reading this thread: do SoTA models like Opus, Gemini and friends actually need output schema enforcement still, or has all the the RLVR training they do on generating code and json etc. made schema errors vanishingly unlikely? Because as a user of those models, they almost never make syntax mistakes in generating json and code; perhaps they still do output schema enforcement for "internal" things like tool call schemas though? I would just be surprised if it was actually catching that many errors. Maybe once in a while; LLMs are probabilistic after all.
(I get why you need structured generation for smaller LLMs, that makes sense.)
Schemas can get pretty complex (and LLMs might not be the best at counting). Also schemas are sometimes the first way to guard against the stochasticity of LLMs.
Yes. Most common failure mode for sota models is to put ```json\n first, but they often do just fail often enough to be worth calling api with json response schema.
This is good. It covers the two easiest dominant methods people use. It even touches on my main complaint for the one they seem to recommend.
That said:
- Constrained generation yields a different distribution from what a raw LLM would provide. This can be pathologically bad. My go-to example is LLMs having a preference for including ellipses in long, structured objects. Constrained generation forces closing quotes or whatever it takes to recover from that error according to a schema, nevertheless yielding an invalid result. Resampling tends to repeat till the LLM fully generates the data in question, always yielding a valid result which also adheres to the schema. It can get much worse than that.
- The unconstrained "method" has a few possible implementations. Increasing context length by complaining about schema errors is almost always worse from an end quality perspective than just retrying till the schema passes. Effective context windows are precious, and current models bias heavily toward earlier data which has been fed into them. In a low-error regime you might get away with a "try it again" response in a single chat, but in a high-error regime you'll get better results at a lower cost by literally re-sending the same prompt till the model doesn't cause errors.
If the authors or readers are interested in some of the more technical details of how we optimized guidance & llguidance, we wrote up a little paper about it here: https://guidance-ai.github.io/llguidance/llg-go-brrr
Are there output formats that are more reliable (better adherence to the schema, easier to get parse-able output) or cheaper (fewer tokens) than JSON? YAML has its own problems and TOML isn't widely adopted, but they both seem like they would be easier to generate.
Just brainstorming. Human beings have trouble writing json, cause it is too annoying. Too strict. In my experience, for humans writing typescript is a lot better than writing json directly, even when the file is just a json object. It allows comments, it allows things like trailing commas which are better for readability.
So maybe an interesting file to have the LLM generate is instead of the final file, a program that creates the final file?
Now there is the problem of security of course, the program the LLM generates would need to be sandboxed properly, and time constrained to prevent DOS attacks or explosive output sizes, not to mention the cpu usage of the final result, but quality wise, would it be better?
You should do your own evals specific to your case. In my evals XML outperforms JSON on every model for out of distribution tasks (i.e. not for JSON that was in the data).
I agree that building agents is basically impossible if you cannot trust the model to output valid json every time. This seems like a decent collection of the current techniques we have to force deterministic structure for production systems.
These are cool tricks but this seems like an impedence mismatch: why would you use an LLM (a probabilistic source of plausible text) in a situation where you want a deterministic source of text where plausibility is not enough?
You... don't. That's exactly what structured outputs are for! You're offloading any formally defined generation to a tool that better serves the case, leaving the ambiguous part of the task to the model.
Code is an example of a mixed case. Getting any mechanistically parsable output from a model is another. Sure, you can format it after the generation, but you still need the generation to be parsable for that. In many cases, using the required format right away will also provide the context for better replies.
> We use a lenient parser like ast.literal_eval instead of the standard json.loads(). It will handle outputs that deviate from strict JSON format. (single quotes, trailing commas, etc.)
A nitpick: that's probably a good idea and I've used it before, but that's not really a lenient json parser, it's a Python literal parser and they happen to be close enough that it's useful.
I'm also really glad that you're helping more people understand this feature, how it works, and how to use it effectively. I strongly believe that structured outputs are one of the most underrated features in LLM engines, and people should be using this feature more.
Constrained non-determinism means that we can reliably use LLMs as part of a larger pipeline or process (such as an agent with tool-calling) and we won't have failures due to syntax errors or erroneous "Sure! Here's your output formatted as JSON with no other text or preamble" messages thrown in.
Your LLM output might not be correct. But grammars ensure that your LLM output is at least _syntactically_ correct. It's not everything, but it's not nothing.
And especially if we want to get away from cloud deployments and run effective local models, grammars are an incredibly valuable piece of this. For practical examples, I often think of Jart's example in her simple LLM-based spam-filter running on a Raspberry Pi [0]:
> llamafile -m TinyLlama-1.1B-Chat-v1.0.f16.gguf \ > --grammar 'root ::= "yes" | "no"' --temp 0 -c 0 \ > --no-display-prompt --log-disable -p "<|user|> > Can you say for certain that the following email is spam? ...
Even though it's a super-tiny piece of hardware, by including a grammar that constrains the output to only ever be "yes" or "no" (it's impossible for the system to produce a different result), then she can use a super-small model on super-limited hardware, and it is still useful. It might not correctly identify spam, but it's never going to break for syntactic reasons, which gives a great boost to the usefulness of small, local models.
* [0]: https://justine.lol/matmul/
Some libraries:
- Outlines, a nice library for structured generation
- Guidance (already covered by FlyingLawnmower in this thread), another nice library - XGrammar, a less-featureful but really well optimized constrained generation library Some papers:- Efficient Guided Generation for Large Language Models
- Automata-based constraints for language model decoding - Pitfalls, Subtleties, and Techniques in Automata-Based Subword-Level Constrained Generation Some blog posts:- Fast, High-Fidelity LLM Decoding with Regex Constraints
- Coalescence: making LLM inference 5x fasterAutomata-based constraints is fun.
(I get why you need structured generation for smaller LLMs, that makes sense.)
With that said, the model is pretty good at it.
That said:
- Constrained generation yields a different distribution from what a raw LLM would provide. This can be pathologically bad. My go-to example is LLMs having a preference for including ellipses in long, structured objects. Constrained generation forces closing quotes or whatever it takes to recover from that error according to a schema, nevertheless yielding an invalid result. Resampling tends to repeat till the LLM fully generates the data in question, always yielding a valid result which also adheres to the schema. It can get much worse than that.
- The unconstrained "method" has a few possible implementations. Increasing context length by complaining about schema errors is almost always worse from an end quality perspective than just retrying till the schema passes. Effective context windows are precious, and current models bias heavily toward earlier data which has been fed into them. In a low-error regime you might get away with a "try it again" response in a single chat, but in a high-error regime you'll get better results at a lower cost by literally re-sending the same prompt till the model doesn't cause errors.
If the authors or readers are interested in some of the more technical details of how we optimized guidance & llguidance, we wrote up a little paper about it here: https://guidance-ai.github.io/llguidance/llg-go-brrr
edit: Somehow that link doesn't work... It's the diagram on the "constrained method" page
Every commercial model provider is adding structured outputs so will keep updating the guide.
What have folks tried?
https://github.com/toon-format/toon
So maybe an interesting file to have the LLM generate is instead of the final file, a program that creates the final file? Now there is the problem of security of course, the program the LLM generates would need to be sandboxed properly, and time constrained to prevent DOS attacks or explosive output sizes, not to mention the cpu usage of the final result, but quality wise, would it be better?
XML is better for code, and for code parts in particular I enforce a cdata[[ part so there LLM is pretty free to do anything without escaping.
OpenAI API lets you do regex structured output and it's much better than JSON for code.
Code is an example of a mixed case. Getting any mechanistically parsable output from a model is another. Sure, you can format it after the generation, but you still need the generation to be parsable for that. In many cases, using the required format right away will also provide the context for better replies.
A nitpick: that's probably a good idea and I've used it before, but that's not really a lenient json parser, it's a Python literal parser and they happen to be close enough that it's useful.