Concepts

Pipeline, not a single script

LitCurate is a sequence of stages. Each stage:

1. Reads artifacts from earlier stages
2. Writes new artifacts under runs/<RUN_ID>/artifacts/
3. Records status in a SQLite ledger (run.db)

You can run one stage at a time, resume after failures, or force a re-run.

Config over code

A config YAML defines the research goal, search settings, filter criteria, LLM providers, schemas, and prompts. Changing domain usually means editing YAML and schema/prompt files — not Python.

Schemas and envelopes

Extraction writes envelopes: schema metadata, model name, timestamp, validation result, and a payload matching your schema. The export stage merges successful papers into database.json.

Runs are ephemeral

Everything under runs/ is runtime output (PDFs, markdown, extractions, databases). It is not part of the published package. Treat experimental campaigns as disposable usage of LitCurate.

Integration model

LitCurate  →  database.json  →  your app / database / notebook

Downstream code should consume the export (or per-paper envelopes) using field names from your schemas.