Evaluation and Safety

TraceVerde includes 6 built-in evaluation detectors for monitoring LLM input/output quality and safety. All features are opt-in and can be enabled individually.

PII Detection

Detect personally identifiable information using Microsoft Presidio.

genai_otel.instrument(
    enable_pii_detection=True,
    pii_mode="redact",       # "detect", "redact", or "block"
    pii_threshold=0.5,
    pii_gdpr_mode=True,
    pii_hipaa_mode=True,
    pii_pci_dss_mode=True,
)

Detects 15+ entity types: email, phone, SSN, credit cards, IP addresses, and more.

Examples:

• Basic detect mode
• Redaction mode
• Blocking mode
• GDPR compliance
• HIPAA compliance
• PCI-DSS compliance
• Combined compliance
• Custom threshold
• Response detection
• Env var config
• PII with Anthropic
• PII with Ollama
• PII with HuggingFace

Toxicity Detection

Detect harmful content using Perspective API (cloud) or Detoxify (local).

genai_otel.instrument(
    enable_toxicity_detection=True,
    toxicity_threshold=0.7,
)

Categories: toxicity, severe_toxicity, identity_attack, insult, profanity, threat.

Examples:

• Basic Detoxify (local)
• Perspective API (cloud)
• Blocking mode
• Category detection
• Custom threshold
• Combined with PII
• Response detection
• Toxicity with Anthropic
• Toxicity with Ollama
• Toxicity with HuggingFace

Bias Detection

Identify demographic biases in prompts and responses.

genai_otel.instrument(
    enable_bias_detection=True,
    bias_threshold=0.4,
)

8 bias types: gender, race, ethnicity, religion, age, disability, sexual_orientation, political.

Examples:

• Basic detection
• Blocking mode
• Category-specific
• Custom threshold
• Hiring compliance
• Response detection
• Multiple evaluations
• Bias with Ollama
• Bias with HuggingFace

Prompt Injection Detection

Protect against prompt manipulation attacks.

genai_otel.instrument(
    enable_prompt_injection_detection=True,
    prompt_injection_threshold=0.5,
)

6 injection types: instruction_override, role_playing, jailbreak, context_switching, system_extraction, encoding_obfuscation.

Examples:

• Basic detection
• Blocking mode
• System override
• Jailbreak techniques
• Payload injection
• Prompt injection with HuggingFace

Restricted Topics Detection

Monitor and block sensitive topics.

genai_otel.instrument(
    enable_restricted_topics=True,
    restricted_topics=["medical_advice", "legal_advice", "financial_advice"],
    restricted_topics_threshold=0.5,
)

Examples:

• Basic detection
• Blocking mode
• Content policy
• Industry compliance

Hallucination Detection

Track factual accuracy and groundedness.

genai_otel.instrument(
    enable_hallucination_detection=True,
    hallucination_threshold=0.7,
)

Includes factual claim extraction, hedge word detection, citation tracking, and context contradiction detection.

Examples:

• Basic detection
• Citation verification
• RAG faithfulness
• Hallucination with Ollama
• Hallucination with Mistral
• Hallucination with HuggingFace

Enable All Evaluations

genai_otel.instrument(
    enable_pii_detection=True,
    enable_toxicity_detection=True,
    enable_bias_detection=True,
    enable_prompt_injection_detection=True,
    enable_restricted_topics=True,
    enable_hallucination_detection=True,
)

See the comprehensive evaluation example and Ollama multiple evaluations.

Metrics

Each detector records metrics for monitoring:

• genai.evaluation.<detector>.detections - Detection events
• genai.evaluation.<detector>.blocked - Blocked requests
• genai.evaluation.<detector>.score - Score distribution (histogram)

See the README for the complete list of span attributes.

All Examples

Browse all evaluation examples in the examples/ directory.