Language Model Assertions

Note Created: 2024-04-14

Prompting emerged as a way to “interact” with LMs due to the naturalness of language. We soon realized one needs to “engineer” their prompts to get desired outputs from these stochastic machines. While interactions are natural and imprecise, engineering is anything but.

Mixing the medium for interaction and engineering results in hand-crafted prompt templates—an error-prone practice, often verbose, and doesn’t always generalize.

That leads us to the question: How can we program LMs without prompting?

Language model (LM) programs are a programming paradigm that combines the precision of conventional programming with the flexibility of LMs. DSPy is a framework for creating such programs declaratively, with automatic prompt tuning through compilation. This short blog summarizes how DSPy works in a nutshell. You can try DSPy in this free Colab. Let’s build a simple ubercool LM program.

PR Reviewer with DSPy

This program reviews a GitHub pull request:

class Reviewer(dspy.Module):
  def __init__(self):
    self.gen_review = dspy.ChainOfThought("pull_request -> review, status")

  def forward(self, pr):
    return self.gen_review(pull_request=pr)

You could use this program as follows:

reviewer = Reviewer()
result = reviewer(pr="def add(a, b): return a - b")
print(, result.status)

# review: ... could be improved by renaming 
# the function and adding proper error handling.
# status: Changes requested

DSPy also automatically compiles quality few-shot prompts for your program, so you don’t have to! You can use it to build a more robust reviewer in a few lines and some I/O pairs. Imagine collecting a few hundred I/O pairs from GitHub PRs and compiling a reviewer that is robust to unseen PRs. 🤩 </br></br> In the code, is_valid is the optimizer metric and could be a fuzzy string match or an LLM call that checks the review’s validity.

teleprompter = BootstrapFewShot(metric=is_valid)
reviewer = teleprompter.compile(Reviewer(), trainset=io_pairs)

Reflection: Is this a good reviewer? 🤔

drawing It is relatively easy to write a valid review for a PR.

But, a good reviewer writes concise, constructive, and informative reviews. The question is, how do we capture and ensure these properties in a program?

“Sketching” the Solution (pun intended)

In traditional program synthesis, particularly in sketching [Solar-Lezama, 2008], developers provide a high-level outline of a program— a sketch—along with a set of assertions that specify the desired behavior. The synthesizer then fills in the details, turning the sketch into a fully-fledged program that adheres to the assertions.

Assertions in a sketch enable you to express intuitive insights without overthinking the implementation details.

💡 Wait? By now, we know that DSPy is a sophisticated program synthesizer at its core. It takes a program specification (Reviewer) and a set of I/O pairs (io_pairs) and returns a tuned program (prompt) that satisfies the spec.

Why not guide DSPy with assertions?

Introducing DSPy Assertions

We introduce to DSPy: LM Assertions. As simple as one-liners, they are assertion style constraints on LM outputs. We distinguish two types of constraints: Assert (hard) and Suggest (soft):

dspy.Assert(constraint: bool, message: str)
dspy.Suggest(constraint: bool, message: str)

Unlike regular assertions, LM assertions are more than just monitors. On violating the constraint, the execution pauses, and the LM program attempts to recover from the violation by backtracking to the failing module. During recovery, the construct uses reflective information to self-correct and continue execution. Asserts fail the program if irrecoverable, while Suggests continue execution. More about how this works in our paper. This tweet is also a great start.

Here’s how we can use a few LM assertions to build our “good” reviewer:

class Reviewer(dspy.Module):
  def __init__(self):
    self.gen_review = dspy.ChainOfThought("pull_request -> review, status")

  def forward(self, pr):
    result = self.gen_review(pull_request=pr)

    # Assert that the review is concise.
    dspy.Assert(len( < 2000, 
        "review must be concise")

    # Suggest that the review be constructive.
        "review must be constructive")

    return result

That’s it! We can now use DSPy to compile a reviewer that satisfies these assertions and suggestions. Our paper evaluates these new constructs and finds that resulting programs are more robust and performant!


LM assertions are simple powerful constructs for guiding LMs toward desired outputs. There are natural connections between traditional program synthesis and self-refining LM programs with assertions. We are excited to explore these connections further and build a next-generation programming paradigm.

Read more in our paper and checkout our Github.