LLM Providers

Patterns - Interface: LLMProviderBase (abstract class) or LLMProvider protocol — implement complete_json(prompt, schema, temperature) -> dict. - Decorators: BudgetedProvider (prompt/output limits), RetryingProvider (exponential backoff + jitter + logging). Stackable. - Circuit-breaker: CircuitBreakerProvider adds trip/half-open/close with optional concurrency guard. - Judge: LLMJudge consumes any provider.

Built-ins - OpenAIProvider: Chat Completions with JSON mode, prompt/output budgets, timeout. Pass to LLMJudge directly or wrap with BudgetedProvider + RetryingProvider.

Implement your own

from contextguard import LLMProviderBase, LLMJudge

class MyProvider(LLMProviderBase):
    def complete_json(self, prompt, schema, temperature=0.0):
        # Call your model and return parsed JSON
        return {"support": 0.6, "contradict": 0.1, "rationale": "...", "reasons": [], "confidence": 0.7}

judge = LLMJudge(MyProvider())

Budget + retry example

from contextguard import OpenAIProvider, BudgetedProvider, RetryingProvider, LLMJudge
base = OpenAIProvider(model="gpt-4o-mini", max_prompt_chars=8000, max_output_tokens=300, timeout=30)
llm = RetryingProvider(BudgetedProvider(base, max_prompt_chars=8000, max_output_tokens=300), max_attempts=3)
judge = LLMJudge(llm)

Notes - Enforce budgets to avoid runaway cost/latency. - Add your own logging/metrics by injecting a logger or subclassing the wrappers.