AEDIN (AgroEcological Database of Interactions) is an open agroecological knowledge base for academic researchers and AI consumer applications. It extracts atomic ecological claims — pest, pathogen, beneficial-insect, pollinator, mycorrhizal, soil, and crop-trait relationships — from open-access scientific literature using a large-language-model extraction pipeline, then verifies each claim through a multi-critic AI consensus: two independent specialty critics (an agroecologist plus a domain expert) must agree on plausibility, with no implausible verdict, before a claim is promoted to public visibility.
Verified claims are published at aedin.io with verbatim source quotes, page citations, and the full multi-critic verdict trail. Vocabulary is aligned with the Global Biotic Interactions (GloBI) Relations Ontology so claims can be pushed back to the public GloBI corpus.
- 🌐 Public site: aedin.io
- 🗺️ Project map & methodology: .okf/ — the Open Knowledge Format bundle documenting architecture, pipelines, datasets, and durable decisions
- 🏛️ Architecture overview: ARCHITECTURE.md
- 🧬 Data model / interaction vocabulary: docs/vocabularies.md
- 📚 Data sources & attribution: docs/data-sources-attribution-policy.md
- 📄 Data license (claims, metadata, verdicts): CC BY 4.0
- 💻 Code license: MIT
- 📖 Citation: CITATION.cff — GitHub renders a "Cite this repository" button
- An open-access PDF is chunked; a large-language-model extracts atomic claims with verbatim quotes and page citations.
- A first-pass critic (Haiku-tier) classifies each claim as plausible / implausible / uncertain / out-of-scope.
- Two independent specialty critics — an agroecologist plus a domain expert (entomologist, plant-pathologist, soil-scientist, or horticulturist) — re-evaluate.
- Only claims with ≥2 plausible verdicts and zero implausible verdicts are promoted to public visibility.
The critic prompts themselves are open — see .claude/agents/.
aedin/
├── backend/ Node.js/Express API + LLM extraction & multi-critic pipeline (CommonJS)
├── web/ Public static site — Astro 5 + Tailwind 4 + better-sqlite3
├── .okf/ Open Knowledge Format bundle — the navigable project map
└── docs/ Data model, methodology, and source-attribution reference
See ARCHITECTURE.md for a fuller overview.
# Backend (API server, port 3001)
cd backend && npm install && npm start
# Web (static site, port 3000)
cd web && npm install && npm run devYou'll need a recent Node.js (≥ 20). The backend reads a curated SQLite corpus (backend/aedin.sqlite). The raw GloBI source (backend/globi.sqlite, large) is rebuilt from the public GloBI dump via npm run sync-global. Both database files are generated locally and are not tracked in git.
If you use AEDIN's data, claims, or methodology in your research, please cite it. The recommended citation is in CITATION.cff; GitHub renders a "Cite this repository" button you can use directly. Two citable Zenodo DOIs are available — the software archive (10.5281/zenodo.21367715) and the verified corpus dataset (10.5281/zenodo.21385401).
- AEDIN's own metadata (claim records, entity records, multi-critic verdicts, aggregate analysis) — CC BY 4.0. Cite AEDIN when reusing.
- Verbatim source quotes shown on individual claim pages retain the license of the original source publication. Each quote displays its source license badge alongside it.
- Site code (extraction pipeline, public web site, administrative tooling) — MIT.
The public web surface at aedin.io is live. The multi-critic AI extraction pipeline continues to ingest open-access literature. Funding is being sought to support next development phases — human-reviewer onboarding, corpus expansion beyond open-access PDFs, and an MCP server interface so other agentic systems can query verified claims with full provenance.
Important provenance note: Claims published here are AI-vouched by multi-critic consensus but not yet human-verified. Each claim links to its verbatim source quote — verify against the source before publishing or citing in your own work.