Implementing Safety Settings in Python: A Comprehensive Guide to Gemini API

📚 Recommended Python Learning Resources

Level up your Python skills with these hand-picked resources:

def explain_safety_system():
    """
    Explain how Gemini's safety system works.
    """
    print("\n" + "=" * 60)
    print("  UNDERSTANDING GEMINI'S SAFETY SYSTEM")
    print("=" * 60)
    
    print("\n  What is the Safety System?")
    print("-" * 60)
    print("""
Gemini has built-in safety filters that detect and block potentially
harmful content in both prompts and responses.

The system evaluates content across multiple harm categories and assigns
a probability rating for each category.
""")
    
    categories = [
        {
            "name": "HARM_CATEGORY_HARASSMENT",
            "description": "Negative or harmful comments targeting identity/protected attributes"
        },
        {
            "name": "HARM_CATEGORY_HATE_SPEECH",
            "description": "Content promoting hatred toward groups based on identity"
        },
        {
            "name": "HARM_CATEGORY_SEXUALLY_EXPLICIT",
            "description": "Contains explicit sexual content or references"
        },
        {
            "name": "HARM_CATEGORY_DANGEROUS_CONTENT",
            "description": "Promotes dangerous activities or could cause harm"
        }
    ]
    
    for cat in categories:
        print(f"\n  * {cat['name']}")
        print(f"    {cat['description']}")

def default_safety_behavior(client):
    """
    Demonstrate default safety behavior without custom settings.
    
    Args:
        client: The initialized Gemini client
    """
    print("\n" + "=" * 60)
    print("  EXAMPLE 1: Default Safety Behavior")
    print("=" * 60)
    
    print("\n[NOTE] Using default safety settings (BLOCK_MEDIUM_AND_ABOVE)")
    
    safe_prompt = "Write a fun story about a robot learning to bake cookies."
    print(f"Prompt: {safe_prompt}\n")
    
    response = client.models.generate_content(
        model='gemini-2.5-flash',
        contents=safe_prompt
    )
    
    print(f"Response: {response.text[:200]}...\n")
    
    if hasattr(response, 'candidates') and response.candidates:
        candidate = response.candidates[0]
        if hasattr(candidate, 'safety_ratings') and candidate.safety_ratings:
            print("Safety Ratings:")
            for rating in candidate.safety_ratings:
                print(f"  * {rating.category}: {rating.probability}")

def custom_safety_settings(client):
    """
    Demonstrate configuring custom safety settings.
    
    Args:
        client: The initialized Gemini client
    """
    print("\n" + "=" * 60)
    print("  EXAMPLE 2: Custom Safety Settings")
    print("=" * 60)
    
    prompt = "Tell me about online safety for children."
    
    settings_variants = [
        {
            "name": "Very Restrictive",
            "settings": [
                types.SafetySetting(
                    category='HARM_CATEGORY_HARASSMENT',
                    threshold='BLOCK_LOW_AND_ABOVE'
                ),
                types.SafetySetting(
                    category='HARM_CATEGORY_HATE_SPEECH',
                    threshold='BLOCK_LOW_AND_ABOVE'
                ),
                types.SafetySetting(
                    category='HARM_CATEGORY_SEXUALLY_EXPLICIT',
                    threshold='BLOCK_LOW_AND_ABOVE'
                ),
                types.SafetySetting(
                    category='HARM_CATEGORY_DANGEROUS_CONTENT',
                    threshold='BLOCK_LOW_AND_ABOVE'
                )
            ]
        },
        {
            "name": "Balanced (Default)",
            "settings": [
                types.SafetySetting(
                    category='HARM_CATEGORY_HARASSMENT',
                    threshold='BLOCK_MEDIUM_AND_ABOVE'
                ),
                types.SafetySetting(
                    category='HARM_CATEGORY_HATE_SPEECH',
                    threshold='BLOCK_MEDIUM_AND_ABOVE'
                ),
                types.SafetySetting(
                    category='HARM_CATEGORY_SEXUALLY_EXPLICIT',
                    threshold='BLOCK_MEDIUM_AND_ABOVE'
                ),
                types.SafetySetting(
                    category='HARM_CATEGORY_DANGEROUS_CONTENT',
                    threshold='BLOCK_MEDIUM_AND_ABOVE'
                )
            ]
        },
        {
            "name": "More Permissive",
            "settings": [
                types.SafetySetting(
                    category='HARM_CATEGORY_HARASSMENT',
                    threshold='BLOCK_ONLY_HIGH'
                ),
                types.SafetySetting(
                    category='HARM_CATEGORY_HATE_SPEECH',
                    threshold='BLOCK_ONLY_HIGH'
                ),
                types.SafetySetting(
                    category='HARM_CATEGORY_SEXUALLY_EXPLICIT',
                    threshold='BLOCK_ONLY_HIGH'
                ),
                types.SafetySetting(
                    category='HARM_CATEGORY_DANGEROUS_CONTENT',
                    threshold='BLOCK_ONLY_HIGH'
                )
            ]
        }
    ]
    
    for variant in settings_variants:
        print(f"\n  Safety Level: {variant['name']}")
        print("-" * 60)
        
        config = types.GenerateContentConfig(
            safety_settings=variant['settings']
        )
        
        try:
            response = client.models.generate_content(
                model='gemini-2.5-flash',
                contents=prompt,
                config=config
            )
            print(f"[OK] Response generated successfully")
            print(f"Response (first 150 chars): {response.text[:150]}...")
        except Exception as e:
            print(f"[X] Response blocked or error: {str(e)}")

def handling_blocked_responses(client):
    """
    Demonstrate how to handle blocked responses gracefully.
    
    Args:
        client: The initialized Gemini client
    """
    print("\n" + "=" * 60)
    print("  EXAMPLE 3: Handling Blocked Responses")
    print("=" * 60)
    
    prompt = "Write a story about cybersecurity."
    
    config = types.GenerateContentConfig(
        safety_settings=[
            types.SafetySetting(
                category='HARM_CATEGORY_DANGEROUS_CONTENT',
                threshold='BLOCK_MEDIUM_AND_ABOVE'
            )
        ]
    )
    
    try:
        response = client.models.generate_content(
            model='gemini-2.5-flash',
            contents=prompt,
            config=config
        )
        print(f"Response: {response.text[:200]}...\n")
    except Exception as e:
        print(f"[X] Response blocked or error: {str(e)}")