What is pager?

PAGER is a research project that helps AI language models give better, more complete answers when they need to pull information from a large document collection. The core problem it tackles is that standard "retrieval-augmented" systems (where an AI looks up documents before answering) often grab scattered, loosely related text. Instead, PAGER organizes what it finds into a structured "knowledge page" so the AI has a clean, coherent context to work from. The approach works in stages. When given a question, a language model first builds a cognitive outline, a set of labeled slots, each representing a different dimension the answer should cover. Then, for each slot, the system searches a document collection (in this case, Wikipedia) and fills in relevant material. It repeats this retrieval-and-refinement process iteratively until every slot is populated. The finished "page" is then handed to the language model as context for generating its final answer. This is aimed primarily at researchers working on retrieval-augmented generation, or anyone building question-answering systems over large knowledge bases. A concrete use case: if someone asks a complex factual question, a standard system might return a few paragraphs that partially relate. PAGER instead structures the retrieved evidence by topic, making it easier for the model to synthesize a well-rounded answer. The README reports that it outperforms baseline methods across multiple benchmark tasks and several large language models. The project is built as a multi-step pipeline with distinct scripts for each stage: encoding the Wikipedia corpus into embeddings, building a search index, constructing the outline, filling in the page, and finally running inference and evaluation. It relies on specific models, Qwen3 and Llama 3.1, and uses a GPU-based serving setup for both embedding and generation. The setup is fairly involved and clearly intended for users comfortable with reproducing research experiments rather than casual application use.