litschema¶
Turn a folder of PDFs into a dataset you can defend.
litschema extracts structured data from scientific papers into a schema you define in LinkML. Each value records where it came from, how it was derived, how likely it is to be right, and whether a person checked it:
- Cited lines. Every value points at the lines of the paper it came from, and validation rejects citations to lines that don't exist.
- Provenance. Each extraction run is immutable and records the hashes of its inputs (PDF, prepared text, schema, prompt) and the model that produced it.
- An independent grader. A second model checks each value against its evidence and gives the probability that the value is correct.
- Human review. You verify, correct, or remove each value in a local web app. Your review is stored as plain JSON beside the run it reviews.
Everything runs on your machine. Models run through your own Claude Code install. The only other network calls are optional DOI lookups.
Alpha software
File formats can change before 1.0. Each release lists its breaking changes in the changelog.