bumpers 0.1.3

Safety/alignment guardrails for AI agents

🛡️ Bumpers

Add safety guardrails to your AI agents with just a few lines of code. Bumpers is a Python library that helps you build safer AI applications by validating agent actions and outputs in real-time.

✨ Features

• 🔒 Validation Engine: Pre-built validators for common safety checks
• 🤖 Framework Support: Works with both custom agents and popular frameworks like LangChain
• 🚫 Action Control: Whitelist allowed actions and block dangerous operations
• 🔍 Content Filtering: Detect and block sensitive or inappropriate content
• 📊 Semantic Monitoring: Track semantic drift and goal fulfillment
• ⏱️ Rate Limiting: Prevent infinite loops and resource abuse
• 📝 Logging & Analytics: Track and analyze agent behavior

🚀 Quick Start

Installation

1. Install:

pip install bumpers

Basic Usage with LangChain

from langchain.agents import AgentExecutor
from bumpers.core.engine import CoreValidationEngine
from bumpers.validators.action import ActionWhitelistValidator
from bumpers.validators.content import ContentFilterValidator
from bumpers.integrations.langchain_callback import BumpersLangChainCallback

# Set up validation engine
validation_engine = CoreValidationEngine()

# Add validators
validation_engine.register_validator(
    ActionWhitelistValidator(allowed_actions=["search", "calculate"]),
    ValidationPoint.PRE_ACTION
)
validation_engine.register_validator(
    ContentFilterValidator(forbidden_words=["confidential", "secret"]),
    ValidationPoint.PRE_OUTPUT
)

# Create Bumpers callback
bumpers_callback = BumpersLangChainCallback(validation_engine=validation_engine)

# Add to your LangChain agent
agent_executor = AgentExecutor(
    agent=agent,
    tools=tools,
    callbacks=[bumpers_callback]  # Add Bumpers protection
)

# Use normally - Bumpers will automatically protect your agent
result = agent_executor.invoke({"input": "What's 2 + 2?"})

🛡️ Available Validators

Action Control

• ActionWhitelistValidator: Restrict which actions an agent can perform
• ResourceValidator: Prevent resource-intensive operations
• RateLimitValidator: Control action frequency

Content Safety

• ContentFilterValidator: Block sensitive or inappropriate content
• LLMSafetyValidator: Use LLMs to validate content safety
• PatternValidator: Block content matching specific patterns

Semantic Control

• SemanticDriftValidator: Keep responses on topic
• GoalFulfillmentValidator: Ensure answers address the original question
• TemporalRelevanceValidator: Validate time-based responses

Chain Control

• ChainLengthValidator: Limit action chain length
• RedundancyLoopingValidator: Detect and prevent action loops

🔍 Monitoring & Analytics

Track agent behavior and get alerts:

from bumpers.monitoring.monitor import BumpersMonitor
from bumpers.monitoring.conditions import create_high_failure_rate_condition

def alert_handler(message):
    print(f"🚨 Alert: {message}")

# Create monitor
monitor = BumpersMonitor(
    logger=logger,
    alert_handlers=[alert_handler]
)

# Add conditions
monitor.add_condition(
    create_high_failure_rate_condition(threshold=0.3)
)

# Start monitoring
monitor.start()

📊 Analytics

Analyze agent behavior:

from bumpers.analytics.analyzer import BumpersAnalyzer

analyzer = BumpersAnalyzer(logger)

# Get validation statistics
stats = analyzer.get_validation_stats()
print("Validation Stats:", stats)

# Get intervention summary
summary = analyzer.get_intervention_summary()
print("Interventions:", summary)

🎯 Fail Strategies

Customize how validators handle failures:

from bumpers.validators.base import FailStrategy

validator = ContentFilterValidator(
    forbidden_words=["confidential"],
    fail_strategy=FailStrategy.STOP  # Immediately halt execution
)

validator = RateLimitValidator(
    max_actions_per_minute=10,
    fail_strategy=FailStrategy.RAISE_ERROR  # Raise exception but allow catching
)

validator = PatternValidator(
    patterns=[r"sensitive\s+data"],
    fail_strategy=FailStrategy.LOG_ONLY  # Just log the violation
)

🔧 Custom Validators

Create your own validators:

from bumpers.validators.base import BaseValidator
from bumpers.core.engine import ValidationResult, ValidationPoint

class CustomValidator(BaseValidator):
    def __init__(self, name: str = "custom"):
        super().__init__(name)

    def validate(self, context: dict) -> ValidationResult:
        # Implement your validation logic
        return ValidationResult(
            passed=True,
            message="Validation passed",
            validator_name=self.name,
            validation_point=ValidationPoint.PRE_ACTION,
            context=context
        )

🌟 Star History