Scientific papers, distilled.

Distils is an open source Python pipeline for scientific knowledge extraction. It turns research PDFs into standalone facts, equations, visual findings, named entities, and a shared glossary.

Use its structured JSON to build retrieval-augmented generation (RAG), GraphRAG, semantic search, and scientific knowledge graphs. The use cases explain how its records and entities fit into each application and what you need to add.

You run the pipeline with a GROBID service and an OpenAI-compatible model endpoint that supports images.

GROBID also preserves bibliographic citation markers and internal references by page, alongside the facts, equations, and visual findings extracted from that page. This gives applications a starting point for exploring citations and scientific content together; see citations and references.

Why standalone records?

A common RAG approach splits a paper into text chunks and retrieves the chunks that look most relevant to a query. A sentence inside one chunk may only make sense with a definition, experimental condition, or named object introduced in another. Increasing chunk overlap helps at nearby boundaries, but it does not recover context introduced several pages earlier.

Raw chunks also preserve the paper's prose, including transitions and repeated explanations. Keyword similarity can therefore surface a passage that looks relevant while providing little reusable scientific information. Distils separates standalone scientific content from citation and internal-reference markers, retaining the markers alongside the content at page level.

Distils replaces raw passages with short scientific statements that name their subject, preserve essential conditions, and remain understandable outside the source paper. It reads each page with a compact record of what has already been established, then extracts knowledge that can stand on its own. Instead of carrying previous pages in full, Distils carries accepted facts and a glossary that grows as new paper-specific terms are introduced.

Distils covers PDF processing and knowledge extraction. Your application supplies embeddings, search indexes, entity resolution, graph construction, and answer generation.

Start here

• Explore RAG, GraphRAG, and search use cases — Plan how to turn the extracted records into a research application.
• Run one paper — Install the pipeline and process the seven-page GQA example.
• Understand the workflow — Follow the three model passes and their shared context.
• Inspect published results — Browse four papers and their complete outputs without an API key.
• Read the output contract — See the records, artifacts, and completion states.

Open source

Distils is licensed under the Apache License 2.0. The source code and GitHub documentation are available to read and build on. Paper-derived examples retain their separate source licenses and attribution.

Models can omit qualifications or misread mathematics and figures. A complete result describes execution, not scientific accuracy. Visual records contain descriptions, not figure crops.