Transcript
Transcript: The SchoolGPT Search Tool, by the Canada School of Public Service
[00:00:00 Text on screen: Artificial Intelligence-Powered Projects in the Government of Canada; The SchoolGPT Search Tool, by the Canada School of Public Service.]
[00:00:16 Text on screen: What challenge did we address?]
[00:00:19 The host appears full screen.]
Host: The Canada School of Public Service is the Government of Canada's common learning provider. Many of you have likely used our catalogue search to find the courses you need.
There is a wealth of learning content available, but for a long time our catalogue relied primarily on traditional keyword search. That works well when learners already know what exact search terms they need to use, but it is much less effective when they want to describe a topic, skill, or problem in plain language. This was not just a usability issue. It directly affected how effectively we delivered on our learning mandate.
Put simply, keyword search depends on knowing the right label in advance, while conversational search allows people to describe what they need in their own words.
At CSPS, helping people access training, when and where they need it, is central to our mandate, so addressing that gap mattered. That's why, in 2025, we launched the SchoolGPT pilot.
As conversational AI tools became more common in the workplace, we anticipated that learners would expect this kind of natural interaction in our platforms too. In many ways, the journey became about building a new pathway to learning, while also discovering and reflecting on how people actually wanted to use it.
[00:01:41 Text on screen: How we went about it.]
[00:01:45 Host appears full screen.]
Host: The vision was to make finding learning materials more natural and intuitive. In early 2023, the AI technical environment was still emerging and evolving, with no stable tools or standards to rely on. We had to build much of that infrastructure ourselves. In many respects, we were building the road before the pavement, and signs, and rules had fully taken shape.
Our early proof-of-concept used open-source tools and commercial AI services. These helped us understand what was possible, but they were not viable in a government context because they did not meet Government of Canada data residency and security requirements. This led to a key decision: moving into a secure environment aligned with enterprise requirements.
In government, a tool is not ready simply because it works. It also has to be safe, governable, and built for the right environment.
[00:02:40 Text on screen: Options explored and our solution.]
[00:02:45 Host appears full screen.]
Host: A key turning point came when secure cloud services were assessed to support Protected B workloads, allowing us to move forward responsibly with the proper safeguards in place.
We implemented controls such as content filtering, masking personal information, and working closely with IT security on identity, access, and network protections. These are guardrails and safety checks that make the system usable in a government context.
Instead of relying on a single system to do everything, we moved toward a coordinated model where different components handle different tasks. Like a road crew, where mapping, paving, and traffic control are handled by different teams. This allowed us to add features more quickly, adapt as the technology matured, and avoid redesigning the whole system.
[00:03:35 Text on screen: How we implemented it.]
[00:03:40 Host appears full screen.]
Host: From the start, we designed SchoolGPT to be modular and adaptable. In simple terms, one part retrieves information, one-part reasons over it, and one-part coordinates how the system responds. This matters because each part can evolve independently, allowing us to adapt to new models over time without major rework.
Security and governance remained central. We aligned with Government of Canada requirements and ensured the system operated safely within our environment. A key practical benefit of this approach was flexibility. We could improve components over time, upgrade capabilities, and expand data sources without starting from scratch.
In effect, we built a pathway that could be improved over time rather than rebuild with each change.
[00:04:29 Text on screen: What happened and looking ahead.]
[00:04:34 Host appears full screen.]
Host: As the ecosystem evolved, so too did our approach.
[00:04:38 The CSPS Learning catalogue's search page appears. Text appears in the search bar: "managing people". Several search results appear as links to courses offered by CSPS.
Host: We moved from basic keyword search toward a system that better understands context and actively retrieves the right information before responding,
The mouse clicks on the SchoolGPT link. The SchoolGPT tool page appears. Text appears in the search bar: I am new to managing people, what are the key skills I need? The tool replies with a list of suggestions, examples, and CSPS training module recommendations.]
Host: rather than relying only on general training. This improves relevance, accuracy, and trust. It's like a guide who checks the map before giving directions, rather than relying only on memory.
By 2025, several key elements came together: a secure environment, the right controls, and a more mature technology landscape. This allowed SchoolGPT to move from experimentation into production.
[00:05:14 Host appears full screen.]
Host: Looking ahead, we are exploring more advanced capabilities that move beyond answering questions and toward helping users solve problems. The focus remains the same: increasing capability while maintaining strong governance and oversight.
[00:05:29 Text on screen: Lessons learned.]
[00:05:34 Our host appears full screen.]
Host: We piloted SchoolGPT alongside the traditional search bar in early 2025. Despite extensive testing, real user behaviour revealed something important. We expected course searches. Instead, users were asking broader, conceptual questions, rather than asking "What courses can I take on digital competencies?"
[00:05:55 The SchoolGPT tool page appears. Text appears in the search bar, as described. The Chatbot replies with a definition of digital competencies, a list of key competency areas, and CSPS training module recommendations.]
Host: They asked questions like "What are the digital competencies?"
This highlighted a key lesson: people do not use conversational tools like search boxes. They use them more like guides.
This insight shaped our next steps, expanding the knowledge base and improving how information is retrieved.
[00:06:15 Our host appears full screen.]
Host: Another important lesson was that breaking tasks into smaller, coordinated components produced better results than trying to handle everything in one place.
Ultimately, real users will always reveal needs that design alone cannot predict. Getting tools into users' hands early is essential.
If there is one takeaway: start small, try before you buy, and stay curious. For public institutions, this is how you balance innovation with responsibility.
[00:06:45 CSPS animated logo appears. Text on screen: canada.ca/school.]
[00:06:51 The Government of Canada wordmark appears.]
