Safety and optimisation

NSWEduChat incorporates several layers of security and optimisation features. These layers work together to make sure our AI chatbot is not just intelligent but also safe, ethical, and a higher quality experience than free-to-use Generative AI.

Below, we explain how the different layers contributing to NSWEduChat work.

How NSWEduChat works

System prompt

The system prompt is the first layer of providing the unique NSWEduChat experience. The system prompt provides the context and boundaries of the platform. The system prompt allows the underlying AI model to know it is operating in NSW Classrooms, and it is this layer that ensures students are accessing an education-focused assistant.

Jailbreak prevention

Jailbreak prevention scans user messages to ensure they are not trying to bypass established protections. This ensures that users cannot trick the system into giving answers it shouldn’t.

Profanity filtering

This layer operates twice in every chat interaction. It ensures that all language going into the bot and coming out is free from profanity and harmful language.

Prompt injection prevention

Prompt injection involves altering and taking advantage of an LLM’s content moderation rules by inputting specific prompts. Deploying specific solutions to address this threat provides another level of protection, ensuring all content returned is safe and aligned with NSWEduChat’s role as an educational assistant.

Auto-prompt engineer

Prompt engineering is the tweaking of prompts or questions you ask the AI to help guide its responses and ensure it keeps the conversation on the right track. The Auto-prompt Engineer level works to achieve two outcomes. First, it ensures NSWEduChat provides the most helpful responses while users learn how to interact most effectively with generative AI systems. Second, valuable feedback is provided to the user to boost their prompting skill level.

Semantic content filter

This level checks that the content the user provides is free from anything inappropriate. This checks the broader meaning of the message for a wider scope of harmful content beyond just profanity. If harmful content is detected, this layer will step in and stop the interaction from going any further.

Orchestrator

The Orchestrator layer allows NSWEduChat to utilise different knowledge sources and generative AI technologies. This ensures that the most effective and efficient model and knowledge source is being used for each request. The orchestrator streamlines this process without requiring any user input. NSWEduChat transparently advises the user which model was used to answer their question.

Constitutional alignment

The Constitutional Alignment layer is the final step in ensuring NSWEduChat is safe and ethical. It acts somewhat as the AI’s moral compass and ensures everything the AI says aligns with the Department’s values. This layer is a prompt that incorporates the NSW Department of Education’s Values in Schools.

This page has recently been modified to reflect the latest information.

Category:

  • DoE

Business Unit:

  • Information Technology
Return to top of page Back to top