Frequently Asked Questions (FAQ)

Common questions about elsai Guardrails.

General

What is elsai Guardrails?

elsai Guardrails is a Python library that adds programmable safety checks to LLM applications, including toxicity detection, sensitive data detection, and content classification.

What LLM providers are supported?

We support:

• OpenAI
• Azure OpenAI
• Anthropic Claude
• Google Gemini
• AWS Bedrock

Is it free to use?

The library is open source. You'll need API keys for the LLM providers you use.

Configuration

How do I configure guardrails?

You can configure guardrails using YAML files or YAML strings. See the Configuration Guide.

Can I use environment variables for API keys?

Yes, you can use environment variables for API keys. See the Configuration Guide.

How do I adjust toxicity threshold?

Set the toxicity_threshold in your configuration:

guardrails:
  toxicity_threshold: 0.7  # Adjust as needed

Usage

How do I check input only?

Use GuardrailSystem directly:

guardrail = GuardrailSystem(config=config)
result = guardrail.check_input("user input")

How do I get detailed results?

Use return_details=True:

result = rails.generate(
    messages=messages,
    return_details=True
)

Can I use it asynchronously?

Yes, use generate_async():

result = await rails.generate_async(messages=messages)

Troubleshooting

Why is my input being blocked?

Check the detailed results to see which check failed:

• Toxicity detection
• Sensitive data detection
• Content classification

How do I disable a specific check?

Set the check to false in configuration:

guardrails:
  check_toxicity: false

Why do I need an embedding encoder for content classification?

Content classification uses embeddings for semantic routing. The encoder type is configurable via the ENCODER_TYPE environment variable. Supported encoders include Azure OpenAI (default), OpenAI, Cohere, HuggingFace, FastEmbed, Ollama, Local, and more.

How do I persist guardrail runs to a database?

Enable ARMS Backend storage in your policy and set API_BASE_URL and ELSAI_ARMS_API_KEY. The Backend routes to MongoDB, DynamoDB, or ClickHouse automatically. See ARMS Storage.

guardrails:
  storage:
    enabled: true
    project: my-app
    arms_correlation: true

How do I link guardrails storage to an ARMS run?

Call link_arms(arms) after creating both clients, or link_run_context(run_id=..., project_id=...). You can also set GUARDRAILS_ARMS_RUN_ID and GUARDRAILS_ARMS_PROJECT_ID environment variables.

How do I detect data leaks in LLM output?

Enable data_exfiltration in your guardrails policy. It runs on output only and can warn (mask) or block based on risk score. See Data Exfiltration Detection.

guardrails:
  data_exfiltration:
    enabled: true
    action_thresholds:
      warn: 20
      block: 80

Why is my output blocked for exfiltration?

Check result.exfiltration on the GuardrailResult. High-risk responses (credentials, bulk PII, export-style payloads) exceed the block threshold. Lower thresholds or disable specific detectors to tune behavior.

Integration

Can I use it with Flask/FastAPI?

Yes, see Integration Examples.

Can I use it with LangChain?

Yes, you can wrap LangChain chains with guardrails. See Integration Examples.