0.1.0

ai_assert

Runtime constraint verification for AI outputs. 278 lines. Zero dependencies.

from ai_assert import ai_assert, valid_json, max_length, contains

result = ai_assert(
    prompt="Return a JSON object with a 'greeting' key",
    constraints=[valid_json(), max_length(200), contains("hello")],
    generate_fn=my_llm,    # any function: str → str
    max_retries=3,
)

print(result.output)       # guaranteed valid JSON, ≤200 chars, contains "hello"
print(result.passed)       # True
print(result.attempts)     # 1–4 (retried with feedback until constraints pass)

The Problem

LLMs don't reliably follow instructions. You ask for JSON, you get markdown. You ask for 100 words, you get 500. You ask for a list of 5 items, you get 7.

Every AI application handles this with ad-hoc validation scattered across the codebase. Or worse — hopes for the best.

The Solution

ai_assert is a universal check → score → retry loop:

1. Generate: Call your LLM (any model, any provider)
2. Check: Run every constraint (multiplicative gate — ALL must pass)
3. Retry: If any constraint fails, feed back failure details and regenerate
4. Return: Verified output with full audit trail

Prompt → LLM → Check all constraints → All pass? → Return ✓
                      ↓ (any fail)
              Build feedback prompt → Retry (up to max_retries)

Key Properties

• Zero dependencies — stdlib only (dataclasses, typing, json, functools)
• Model-agnostic — works with OpenAI, Anthropic, local models, anything with str → str
• 278 lines — small enough to read in one sitting, audit in an hour
• Multiplicative gate — a zero in ANY constraint = failure (not averaged away)
• Continuous scoring — each check returns a score in [0, 1), never 1.0
• Full audit trail — every attempt, every check result, preserved in result.history

Install

pip install ai-assert

Or just copy ai_assert.py into your project. It's one file with zero dependencies.

Quick Start

1. Define your generator

Any function that takes a str and returns a str:

from openai import OpenAI

client = OpenAI()

def my_llm(prompt: str) -> str:
    response = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user", "content": prompt}],
    )
    return response.choices[0].message.content

2. Declare constraints

from ai_assert import valid_json, max_length, min_length, contains, not_contains, matches_schema, custom

constraints = [
    valid_json(),                                          # must be parseable JSON
    max_length(500),                                       # at most 500 chars
    matches_schema({"required": ["name", "age"]}),         # must have these keys
    not_contains("```"),                                   # no markdown code fences
    custom("positive_age", lambda s: json.loads(s).get("age", 0) > 0,
           fail_msg="Age must be positive"),               # any boolean function
]

3. Call ai_assert

from ai_assert import ai_assert

result = ai_assert(
    prompt="Generate a user profile as JSON with name and age.",
    constraints=constraints,
    generate_fn=my_llm,
    max_retries=3,
)

if result.passed:
    user = json.loads(result.output)  # safe — guaranteed valid
    print(f"Got {user['name']}, age {user['age']}")
else:
    print(f"Failed after {result.attempts} attempts")
    for check in result.checks:
        if not check.passed:
            print(f"  ✗ {check.name}: {check.message}")

4. Or use the decorator

from ai_assert import reliable, valid_json, max_length

@reliable(constraints=[valid_json(), max_length(500)], max_retries=3)
def generate_profile(prompt: str) -> str:
    return my_llm(prompt)

result = generate_profile("Generate a user profile as JSON.")
# result is an AiAssertResult — same interface as ai_assert()

Built-in Constraints

Custom constraints

from ai_assert import Constraint

def word_count_between(min_w, max_w):
    def check(output):
        count = len(output.split())
        if min_w <= count <= max_w:
            return True, 0.99, f"{count} words"
        return False, 0.0, f"{count} words, need {min_w}-{max_w}"
    return Constraint(name=f"word_count({min_w}-{max_w})", check_fn=check)

The AiAssertResult

@dataclass
class AiAssertResult:
    output: str                  # the final output
    passed: bool                 # did all constraints pass?
    checks: list[CheckResult]    # per-constraint results
    attempts: int                # how many tries it took
    history: list[dict]          # full audit trail of all attempts

    @property
    def composite_score(self) -> float:   # average score, always < 1.0

Each CheckResult:

@dataclass
class CheckResult:
    name: str       # constraint name
    passed: bool    # pass/fail
    score: float    # continuous in [0, 1), never 1.0
    message: str    # human-readable explanation

Benchmark Results

Tested on IFEval (541 prompts, 25 constraint types):

• 30 prompts rescued from failure by the retry mechanism
• All improvements from runtime verification — no model changes

Design Decisions

Why multiplicative, not averaged? If your JSON is invalid, it doesn't matter that the length was perfect. A zero in any dimension is system failure. This is the multiplicative gate — all(c.passed for c in checks).

Why scores in [0, 1) and not [0, 1]? No verification achieves perfect fidelity. The score participates in correctness without claiming identity with it. This prevents false confidence in downstream aggregation.

Why retry with feedback, not just Best-of-N? Feedback-directed retry is more sample-efficient. The failure messages tell the model exactly what to fix. Best-of-N generates blindly N times and picks the best. (Both approaches compose — you can use ai_assert inside a Best-of-N loop.)

Why zero dependencies? So you can drop it into any project. No framework lock-in. No transitive dependency hell. Copy the file and go.

Works With

ai_assert works with any str → str function. Tested with:

• OpenAI (GPT-4o, GPT-4o-mini)
• Anthropic (Claude 3.5 Sonnet)
• Local models (Ollama, vLLM, llama.cpp)
• Any HTTP API (wrap the call in a function)

See examples/basic_usage.py for integration code.

Compared To

Contributing

Issues and PRs welcome. The codebase is 278 lines — reading the entire thing takes less time than reading this README.

License

MIT — see LICENSE.