Listen: Query Intent via Retrieval Augmentation and Model Distillation

Understanding what people are searching for online can be tricky, because search queries are often short and vague. To solve this, researchers developed a system called QUILL, which uses large language models to better classify search intent.

QUILL relies on retrieval-augmented generation. This process looks up relevant web pages and adds their titles and web addresses to the search query for extra context. While this extra information makes the model much smarter, it also makes it slower and more expensive to run.

To keep things fast, the researchers designed a unique, two-stage distillation process. First, they distill a massive, context-rich model called the Professor into a smaller Teacher model. Then, they distill that Teacher into an even smaller Student model. This final Student model is highly efficient and ready for real-world applications, yet it retains most of the performance gains of the larger models.

The researchers also discovered that when adding context, more is not always better. Stacking too many features leads to diminishing returns. Web addresses, or URLs, actually provide the most consistent and valuable context on their own, even more than page titles. For search engine optimization, or SEO, this is incredibly practical. It means you can build highly effective tools using just queries and primary URLs, simplifying your data pipelines while keeping your systems fast and accurate.