Potential use cases for Distils

Explore applications you could build from the scientific knowledge extracted by Distils: tools to retrieve, compare, and connect research papers. The use cases below describe potential applications of its outputs, each requiring additional development.

Distils handles PDF processing and knowledge extraction, producing standalone facts, equations, visual findings, glossary definitions, and named entities in structured JSON. It also groups GROBID's bibliographic citation markers and internal references with each page's records. Your application adds the indexing, relationship modeling, retrieval, or answer generation needed for the chosen use case.

RAG over scientific papers

Retrieval-augmented generation (RAG) uses information retrieved from your papers to help a model answer a question. Distils supplies standalone scientific statements with explicit subjects, essential conditions, and source pages, giving the model context it can use beyond the original passage.

Index these records, retrieve those relevant to a question, and keep the paper and page alongside the answer. This approach is related to proposition-based retrieval, which uses individual statements as retrieval units; see Dense X Retrieval. Distils provides the extracted content; your application handles search and answer generation. The output reference covers the record mapping.

Semantic search

For scientific literature search, index the facts and the meanings of equations and visual findings. Embeddings can support semantic search; a keyword index can support exact scientific terms. Combining the two gives you a starting point for hybrid search.

For example, a query about sharing key and value heads could retrieve the standalone record about grouped-query attention shown in the output format. Your interface could show the statement, paper, and source page together. Distils provides the content; you implement and evaluate the search index and ranking.

Scientific knowledge graphs

Named entity extraction identifies methods, systems, datasets, and scientific concepts. The output separates a paper's main subjects from useful background entities, providing candidates for connecting papers through shared concepts.

You can build a graph with paper and entity nodes, then create edges from each paper to the entities listed on its pages. Resolve aliases and ambiguous names across documents before merging nodes. This supports literature discovery and grouping papers by a common method or dataset.

For example, an application could link papers mentioning grouped-query attention to a shared method node. Preserve whether the entity was selected as main or secondary and the pages that support the association. A mention indicates an association with the paper; it does not by itself establish a relation such as “improves on” or “outperforms”.

Entities are attached to pages, not individual facts. Distils does not produce typed relationships between entities, resolve identities across papers, or construct a citation graph. Those steps require additional processing.

GraphRAG

GraphRAG uses a knowledge graph to help retrieve context for answer generation. Microsoft's GraphRAG implementation extracts entities and relationships, builds graph communities, and summarizes them for retrieval.

Distils can supply scientific statements and candidate entities for a GraphRAG application. To build on them, you need to resolve entities, extract or define relationships, construct the graph, and implement the retrieval and answer stages.

The GQA example lets you inspect what is available before designing that integration: extracted facts, named methods, glossary entries, and source pages.

Citation graphs and reference navigation

Explore a page's scientific content alongside its references to other works and to figures, tables, or equations within the paper. Distils groups these GROBID markers with the page's facts, equations, and visual findings, giving a literature exploration tool both the extracted knowledge and its surrounding references.

An application could build a graph from papers to their pages, and from each page to its scientific records and reference markers. By parsing the saved GROBID XML and resolving reference targets, it could extend that graph to cited works and referenced objects. Readers could then find pages citing a particular paper and inspect the findings discussed on those pages.

These are associations through a shared source page. Distils does not determine which citation supports an individual fact, or link a reference to a specific extracted visual or equation. Claim-to-citation alignment requires additional analysis of the source context. See the reference fields and saved GROBID XML for what is available.

Scientific literature discovery

Build a literature exploration interface around normalized entities and extracted findings. A reader could start from a method, discover related papers, then compare the statements, equations, or visual findings on their source pages.

The distinction between main_entities and secondary_entities gives you a way to separate papers centered on a concept from papers that mention it as background. Glossary definitions can help explain unfamiliar terms. Your application still needs entity resolution, linking to any shared concept catalog, and ranking rules for the collection.

For literature review support, retain each statement's experimental conditions, quantities, and uncertainty. Similar wording does not establish agreement between papers, and an omitted finding is not evidence that a paper lacks it. Comparison tools should let readers inspect the original sources.

Scientific PDF to JSON

The same outputs can populate a scientific knowledge base or a literature exploration tool. Because the extraction model reads page images, it can capture scientific content in equations, figures, and tables as well as prose. Equation records contain LaTeX and an interpretation; visual records contain findings in text, rather than image crops or a full reconstruction of every table cell.

Hypergraphs for scientific knowledge

Hypergraphs are a more exploratory application of Distils' outputs. A scientific result often involves a method, a dataset, a metric, and experimental conditions together. A hypergraph can represent this as one shared relation: a hyperedge connecting several entities, with their roles, the result, and its conditions attached.

For example, consider the illustrative statement: “Method A improves metric M on dataset D under condition C.” A hyperedge could group the method, dataset, and metric while keeping the condition and finding with that specific experiment. This can help a research application retrieve the full relationship when comparing results across papers.

Distils' standalone facts and equation interpretations can be starting points for these hyperedges because they aim to preserve each claim's subject and essential conditions. Your application still needs to identify the entities involved in each statement, resolve their names, and construct the relations: Distils attaches entities to pages, not individual facts, and does not generate hyperedges.

Hypergraphs can also support retrieval-augmented generation, as explored in HyperGraphRAG. Their value for a Distils corpus would need to be evaluated against simpler retrieval approaches.

Start with the quick start to process a paper, or inspect the published results and output contract before building an importer.