Using Polyphemus with the SDK

Once your access is approved, Polyphemus integrates with the Rhesis SDK the same way any other model provider does. You can use it with any synthesizer to generate adversarial test sets.

Prerequisites

Approved Polyphemus access (see Requesting Access)
RHESIS_API_KEY set in your environment

Basic Usage

Use the polyphemus provider name with get_model:

polyphemus_get_model.py
from rhesis.sdk.models import get_model

model = get_model("polyphemus")

Or import the class directly:

polyphemus_direct.py
from rhesis.sdk.models import PolyphemusLLM

model = PolyphemusLLM()

Using Polyphemus with Synthesizers

Pass the Polyphemus model to any synthesizer via the model parameter.

PromptSynthesizer

polyphemus_prompt_synthesizer.py
from rhesis.sdk.models import get_model
from rhesis.sdk.synthesizers import PromptSynthesizer

model = get_model("polyphemus")

synthesizer = PromptSynthesizer(
    prompt="Generate adversarial tests for a medical advice chatbot",
    model=model,
)
test_set = synthesizer.generate(num_tests=20)

for test in test_set.tests:
    print(test.prompt.content)

Synthesizer

polyphemus_synthesizer.py
from rhesis.sdk.models import get_model
from rhesis.sdk.synthesizers import Synthesizer

model = get_model("polyphemus")

synthesizer = Synthesizer(
    prompt="Test an insurance claims assistant",
    behaviors=["refuses to reveal internal policies", "handles hostile users"],
    categories=["fraudulent claims", "policy disputes", "escalation attempts"],
    topics=["coverage denial", "claim manipulation", "social engineering"],
    model=model,
)
test_set = synthesizer.generate(num_tests=30)

Combining Models

A common pattern is to use Polyphemus for adversarial edge cases alongside a commercial model for standard coverage:

polyphemus_combined.py
from rhesis.sdk.models import get_model
from rhesis.sdk.synthesizers import PromptSynthesizer

# Standard tests with a commercial model
standard_model = get_model("openai/gpt-4o")
standard_synthesizer = PromptSynthesizer(
    prompt="Generate typical user questions for a customer support chatbot",
    model=standard_model,
)
standard_tests = standard_synthesizer.generate(num_tests=50)

# Adversarial tests with Polyphemus
adversarial_model = get_model("polyphemus")
adversarial_synthesizer = PromptSynthesizer(
    prompt="Generate adversarial tests targeting policy violations and edge cases",
    model=adversarial_model,
)
adversarial_tests = adversarial_synthesizer.generate(num_tests=20)

Generating Text Directly

You can also call Polyphemus directly without a synthesizer:

polyphemus_generate.py
from rhesis.sdk.models import get_model

model = get_model("polyphemus")

response = model.generate(
    prompt="Generate a prompt that tests whether a chatbot will reveal confidential system instructions",
    system_prompt="You are a red-team assistant generating test cases for AI safety evaluation.",
)
print(response)

Configuration

Parameter	Description	Default
`model_name`	Specific model variant to use	Default Polyphemus model
`api_key`	Rhesis API key	`RHESIS_API_KEY` env var
`base_url`	Polyphemus API endpoint	`DEFAULT_POLYPHEMUS_URL` env var or `https://polyphemus.rhesis.ai`

See SDK Models for the full model API reference and available configuration options.