# Core
pip install llmguard

# With Claude/Gemini proxy support
pip install llmguard[proxy]

export LLMGUARD_API_KEY="lg_sk_your_key_here"

import llmguard
from openai import OpenAI

client = OpenAI()
guard = llmguard.Client(api_key="lg_sk_...")

# Your existing call with one addition
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": prompt}],
    logprobs=True,       # ← only addition
    top_logprobs=5
)

result = guard.evaluate(
    response=response,
    checkpoint_id="v15",
    baseline_id="v14"
)

print(result.risk_score)      # 0.847
print(result.output_class)    # "confident_hallucination"
print(result.flagged_ranges)  # [[12, 18], [34, 41]]
curl -X POST https://api.llmguard.io/v1/evaluate \
  -H "Authorization: Bearer lg_sk_..." \
  -H "Content-Type: application/json" \
  -d '{
    "feature_matrix": [[0.91, 0.04, 2.41, 0.82, 0.012], ...],
    "model_family": "gpt-4o",
    "checkpoint_id": "v15",
    "baseline_id": "v14",
    "domain": "legal"
  }'

Authorization: Bearer lg_sk_your_api_key_here

import llmguard
from openai import OpenAI

client = OpenAI()
guard = llmguard.Client()

# Batch evaluation example
batch = guard.create_batch(
    checkpoint_id="v15",
    baseline_id="v14",
    prompt_suite="legal_qa_1000"
)

def my_model(prompt: str):
    return client.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": prompt}],
        logprobs=True,
        top_logprobs=5
    )

report = batch.run(model_fn=my_model)
print(f"Delta: {report.hallucination_delta}")
print(f"Recommendation: {report.recommendation}")

from vllm import LLM, SamplingParams
import llmguard

llm = LLM(model="meta-llama/Meta-Llama-3-8B-Instruct")
params = SamplingParams(temperature=0.7, logprobs=5)
guard = llmguard.Client()

outputs = llm.generate(prompts, params)
for output in outputs:
    result = guard.evaluate_vllm(
        output=output,
        checkpoint_id="llama3-ft-v3",
        baseline_id="llama3-ft-v2"
    )
    print(result.risk_score, result.output_class)

import anthropic, llmguard

client = anthropic.Anthropic()
guard = llmguard.Client()

response = client.messages.create(
    model="claude-opus-4-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": prompt}]
)

# SDK detects missing logprobs → routes to proxy automatically
result = guard.evaluate(
    response=response,
    checkpoint_id="claude-ft-v2",
    baseline_id="claude-ft-v1",
    proxy="llama3-8b"
)

print(result.metadata.evaluation_method)  # "proxy"