Run Protect from the SDK

This guide screens a single input by calling protect() directly from your own code: no dashboard involved. You can call it on incoming input before it reaches your model, or on the model’s output before it reaches whoever receives it; protect() takes a string either way. The guide walks through initializing the client, building a rules list, running the check, and reading the result back, for both text and non-text input. For every parameter and return field, see the Protect SDK reference.

Initialize Protect

Install the SDK first:

pip install ai-evaluation

from fi.evals import Protect

protector = Protect()

Protect reads FI_API_KEY and FI_SECRET_KEY from your environment. Generate these on the API Keys page if you don’t have them yet.

Build the rules list

protect_rules is a list of dicts, each naming one check to run with a metric key:

rules = [
    {"metric": "toxicity"},
    {"metric": "bias_detection"},
    {"metric": "prompt_injection"},
    {"metric": "data_privacy_compliance"},
]

For how these checks fit into the guardrail model, see Understanding Protect.

The SDK accepts these metric values: toxicity, bias, bias_detection, sexist, prompt_injection, data_privacy_compliance, and pii. A metric name outside this list raises an error, so confirm the name before you ship it.

Run the check

Call protector.protect() with:

• inputs: your input, as the first positional argument
• protect_rules: the rules list
• action: the message returned when a rule fails
• reason: set True to include an explanation with the result
• timeout: how long the check can run, in milliseconds

text_to_check = "the text you want to screen"

result = protector.protect(
    text_to_check,
    protect_rules=rules,
    action="I'm sorry, I can't help you with that.",
    reason=True,
    timeout=25000,
)

Read the result

protect() returns a dictionary shaped like this:

{
    "status": "failed",
    "completed_rules": ["toxicity"],
    "uncompleted_rules": ["bias_detection", "prompt_injection", "data_privacy_compliance"],
    "failed_rule": ["toxicity"],
    "messages": "I'm sorry, I can't help you with that.",
    "reasons": ["Message contains content flagged as toxic."],
    "time_taken": 0.42,
}

• status: "passed" or "failed"
• messages: on a failure, carries the action string instead of the original input
• completed_rules / uncompleted_rules: which checks ran to completion, and which didn’t; on a failure, the remaining checks can come back in uncompleted_rules
• failed_rule: the check(s) that tripped a failure, as a list, for example ["toxicity"]
• reasons: holds an explanation for a failure when reason=True is set
• time_taken: how long the check took, in seconds

If a rule doesn’t finish before timeout elapses, it lands in uncompleted_rules instead.

Use status to decide what to send onward: on a failure, forward messages instead of the original input; on a pass, forward the input unchanged.

if result["status"] == "failed":
    response = result["messages"]
else:
    response = text_to_check

Screen an image or audio input instead of text

inputs isn’t limited to text. Pass an image or audio path or URL in place of the text string, and call protect() the same way. The rules list carries over unchanged. For accepted URL and file formats, see the Protect SDK reference.

result = protector.protect(
    "/path/to/local/audio.wav",
    protect_rules=rules,
    action="Audio content cannot be processed",
    reason=True,
    timeout=25000,
)

result = protector.protect(
    "/path/to/local/image.png",
    protect_rules=rules,
    action="Image content cannot be processed",
    reason=True,
    timeout=25000,
)

Text, image, and audio are the only accepted input types for this call. Image sets, PDFs, and knowledge bases are not accepted.

Dive deeper