For semantic search we used rapidly evolving open-source technology including Solr, Python and community-developed language models. We demonstrated how this technology could be leveraged in a safe way, with privacy and trust fundamental to the approach.

We used a locally hosted language model, meaning the language model is stored and used on the Head to Health platform and that no data leaves the website. This ensures user privacy and that no search data can be harvested by third parties. This approach also ensures safety and consistency. SaaS language models are regularly updated which can introduce risks, as these updates might change how the models behave. Instead, our locally hosted model is a snapshot that can be updated as part of intentionally designed and tested changes.

We also used Natural Language Processing (NLP) rather than generative AI. While the two are related, there are some clear distinctions between them. Generative AI is the ability to generate new content or data, including text, while NLP focuses on understanding and processing human language.

The language model is used only to classify, retrieve and rank existing content; no new content or data is generated. This way we avoid AI hallucinations, which is when a language model perceives patterns that are non-existent and creates nonsensical or inaccurate outputs, which can be especially harmful in sensitive areas. For example, the model may incorrectly classify content and present results that are safe but perhaps out of context – such as confusing 'ADD' (the abbreviation for attention deficit disorder) with 'addition'.

How AI can be used in a sensitive area to help people express their needs

Improving mental health and wellbeing outcomes for all Australians