NUS AI Sense Maker turns 150,000 digitised collections into a conversational research tool

The examples reported for the corpus show why ordinary keyword search may not be enough. Users may encounter material including Singapore’s 1881 Chinese-language daily Lat Pau, the 1817 first printed edition of Sir Stamford Raffles’ The History of Java, Tang-Dynasty literary works and, once digitised and ingested, Wang Sha’s original musical compositions and scripts.

How an AI Sense Maker search works

A user begins with a question written in everyday language rather than a list of exact database keywords. The platform offers fast, iterative and deeper search modes, allowing users to choose between a quick orientation and a more extensive exploration.

The system is designed to return several layers of help:

1. A structured topic overview that gives the user an initial map of the subject.
2. Related concepts that broaden or refine the original question.
3. References and further reading drawn from the NUS repositories.
4. Citations that point users back to the underlying material.
5. A knowledge graph showing connections among concepts, disciplines and sources.

For example, a question about why some Singapore neighbourhoods feel hotter could lead to connections among building density, vegetation, construction materials and water features. The graph may then relate those ideas to urban ecology, environmental science, atmospheric physics and microclimatology.

This makes the platform most useful as a research-orientation layer: it can help a student see which terms, disciplines and sources may matter before deciding what to read closely.

Why the knowledge graph matters

Traditional search often returns a ranked list of documents. A knowledge graph can add a different view by showing how the ideas in those documents relate to one another.

AI Sense Maker’s graph is based on a managed taxonomy. The taxonomy was initially informed by Wikidata, then narrowed and refined for NUS’s collections. That curation matters because the usefulness of a graph depends not only on the language model, but also on how concepts are defined and connected within the collection being searched.

For interdisciplinary projects, this could help reveal relevant vocabulary or neighbouring fields that a user would not have included in the first query. The resulting links are leads for investigation, not proof that every suggested relationship is academically sound.

GPT-4.1, retrieval-augmented generation and source guard rails

The platform uses OpenAI’s GPT-4.1 together with retrieval-augmented generation, or RAG. In a RAG system, the model first retrieves relevant material from a defined source collection and then uses that material to construct an answer. AI Sense Maker is designed to ground its responses in retrieved NUS sources and attach citations so users can inspect them.

Its guard rails are also designed to keep the tool focused on the university’s academic content: the platform is not intended to search the public internet to assemble its answers. That source boundary may reduce unsupported output, but it does not remove the need to verify the cited documents, interpret primary material and check whether the answer fairly represents the evidence.

GPT-4.1 is the model named in the reporting about AI Sense Maker. OpenAI describes GPT-4.1 as a model family with improvements in instruction following and long-context comprehension. Those general model capabilities should not be confused with independent evidence that AI Sense Maker will produce error-free research results.

Current limitations and planned upgrades

The first release has important limits:

• It supports one-off queries, so it does not yet retain conversational context across a sequence of follow-up questions.
• It can search only material that has already been digitised and incorporated into the relevant NUS repositories.
• It can help identify sources and connections, but it cannot assess every scholarly argument or formulate a final research case for the user.

NUS is developing chat-history memory for longer research conversations and voice-to-text transcription as planned enhancements.

These constraints define the right workflow. A user can treat AI Sense Maker as a starting point for discovery, then open the cited sources, compare interpretations and build the final analysis independently.

Could it reduce research discovery time?

NUS estimates that the initial discovery and orientation stage of research, which can take about two weeks, could be reduced to roughly 30 minutes or less with AI Sense Maker.

That figure is an institutional estimate, not an independently validated productivity result. The more defensible expectation is that the platform may reduce time spent on weak search leads and help users assemble an initial, cited set of sources more quickly.

The likely benefit is therefore front-loaded: faster topic scoping, better vocabulary for subsequent searches, and earlier visibility into connections across disciplines and collections. The reading, source criticism and argument-building that follow remain human research work.