Why it Matters
Brand Safety Framework
Preventing Abuse
Safety
- Botpress provides a robust feature set that ensure AI-driven interactions are safe, controlled, and compliant.
- Ensuring the safety of AI-driven interactions is a core priority for our team, and for bot builders. The Botpress platform is designed to empower developers while maintaining strict safety standards, giving you the tools needed to build reliable and controlled AI solutions. This overview covers the key safety aspects of Botpress, including constraint management, data privacy, access control, and mitigation of AI-related risks like hallucinations.
Key Safety Features
- Policy Agent: Control AI behavior with predefined rules and constraints to ensure responses align with your organization’s standards. The Policy Agent acts as a gatekeeper, enforcing guardrails and mitigating hallucinations by filtering outputs against customizable policies.
- Access Control & Permissions: Secure your environment with granular access control. Define roles and permissions to regulate who can view, edit, and deploy AI models, ensuring that only authorized users can make changes.
- Data Privacy & Security: Botpress prioritizes data security with encryption, anonymization, and secure API connections. Our platform helps you comply with privacy regulations, safeguarding sensitive information across all interactions.
- Monitoring & Analytics: Gain insights into AI behavior and performance through monitoring and analytics tools. Track usage patterns, detect anomalies, and adjust policies in real time to maintain optimal safety levels.
- Error Mitigation & Recovery: Implement fallback mechanisms that catch errors or misbehaviour, redirecting interactions to human agents or predefined safe states. This ensures continuity and reliability even when the AI is uncertain or encounters edge cases.
Why Safety Matters
- Building trust in AI systems starts with safety. By implementing robust safety measures, Botpress ensures your AI solutions aren’t only effective but also secure and compliant with industry standards. Our commitment to safety helps you deploy AI with confidence, knowing that you are in control of every interaction.
Preventing Abuse
Brand Safety Framework
Preventing Abuse
- At Botpress, we prioritize the security and integrity of your bots by implementing a range of robust measures to counteract fraud and malicious usage. Our strategy includes:
Rate Limits Team
- Rate limits serve as a fundamental safeguard for your bot’s resources. Each bot is assigned a maximum rate of messages it can process per second. For users on the “Team” plan, a higher rate limit is applied. Enterprise customers can further customize these limits to align with their specific requirements.Rate limiting prevents the excessive influx of messages, ensuring that your bot operates efficiently and consistently. This is an essential component of our abuse prevention strategy.
Allowed Origins Team
- Allowed Origins is a feature that lets you explicitly control where your bot is allowed to run by specifying a whitelist of approved domains. When deploying your bot on a website, you include only the domains that you trust, and the system automatically blocks any attempts to embed the bot from unapproved sources. This straightforward, security-first approach not only minimizes the risk of unauthorized use but also aligns with best practices for maintaining a secure, controlled deployment environment.
- Web Application Firewall (WAF)
- Our Web Application Firewall, often referred to as WAF, is a powerful layer of defense against various cyber threats. WAF can protect your bot from Distributed Denial of Service (DDoS) attacks, SQL injection, cross-site scripting, and many other malicious activities.By using WAF, Botpress takes advantage of AWS’s robust security mechanisms, which are detailed in AWS WAF Documentation. This means that your bot is shielded from a wide range of web-based threats, allowing it to perform at its best without disruption.
Brand Safety Framework
Brand Safety Framework
Brand Safety Framework
- Control AI responses, enforce brand-aligned messaging, and prevent harmful outputs.
- The Brand Safety Framework in Botpress is a comprehensive suite of features designed to provide transparency and control over AI-driven decisions, ensuring brand safety across all interactions. Similar to a car’s safety features—such as airbags and crash collision ratings—the framework isn’t a single toggle but a holistic approach that includes several powerful tools designed to give you visibility into and control over an LLM’s actions. Botpress empowers you with unparalleled insight into the logic of LLMs and equips you with the controls to adjust and refine decision-making processes, securing brand integrity in every interaction.
LLM Transparency
- Traditional LLMs and providers often operate as opaque black boxes, making it challenging to understand the reasoning behind their decisions. Botpress’s BSF addresses this challenge by exposing the decision-making pathways of LLMs, allowing users to see why certain outputs were generated and how they align with brand standards. The core principle is simple: when you can understand and influence an LLM’s decisions, you can protect your brand.
Key Features
Policy Agent
The Policy Agent is a rules-based mechanism that controls the outputs of LLMs by enforcing constraints and guardrails. By configuring policies, users can dictate acceptable response behaviors, prevent harmful outputs, and ensure that AI actions align with brand values.
Usage
- Define constraints, including YAML or JSON configurations.
- Set gates to filter out undesirable responses based on specific keywords, patterns, or context cues.
- Identify policies or behaviors your agent should adhere to during conversations.
RAG Safety
Retrieval-Augmented Generation (RAG) Safety provides transparency over the data selection process used by LLMs. It reveals which data chunks were identified as relevant to a query and which were ultimately submitted to the LLM. Users have the ability to review, approve, or modify these selections, ensuring that only the most appropriate data informs the AI’s output.- Usage
- Access a visual interface that displays selected chunks with contextual relevance scores.
- Modify data chunk selections manually or set up automated rules to refine data retrieval criteria.
- Log and audit the data flow to track the decisions made during the query process.
LLM Inspector
The LLM Inspector is a tool that breaks down the decision-making process of the LLM, showing why it selected certain actions or responses. By dissecting the LLM’s internal logic, users can gain insights into the decision pathways and make informed adjustments to align AI behavior with brand standards.
Usage
- Inspect decision trees that outline response logic and action selection.
- View confidence scores, logic paths, and influence factors contributing to specific outputs.
- Adjust weighting and influence factors to steer future decision-making processes
HITL
HITL, or Human-in-the-Loop, allows you to participate in an AI agent’s conversation directly from the Botpress dashboard. This feature offers a critical layer of oversight and control when using LLMs in customer- or user-facing interactions. It was designed to ensure that sensitive or high-risk conversations can be escalated to human agents, helping maintain brand integrity, mitigate potential issues, and protect user privacy.
Usage
- You can use HITL to define rules or “gates” within a conversation that determine when a human agent should intervene. For example, if the conversation involves sensitive personal data, triggers specific legal language, or moves into territory that may require human expertise, the conversation is automatically handed off to a human agent for further handling.You can configure the system to identify specific keywords, scenarios, or actions that require human oversight. This level of customizability ensures that only relevant conversations are escalated, preserving efficiency while safeguarding privacy.By integrating HITL into your LLM-driven solutions, you ensure your brand remains compliant with privacy standards, while retaining the flexibility and efficiency of AI-driven interactions.This feature allows businesses to have the best of both worlds: leveraging cutting-edge AI to handle the bulk of interactions, while having the safety net of human oversight in sensitive situations. HITL enhances trust by allowing businesses to demonstrate that they’re taking proactive steps to ensure conversations are handled responsibly. This creates confidence in the platform’s ability to handle customer interactions without compromising privacy or brand safety.