AI4A is an AI assistant for agroforestry. It answers questions using retrieval-augmented generation (RAG) over a curated knowledge base - research from the EURAF community on Zenodo plus a catalogue of agroforestry tools, datasets, and projects. Built with Next.js.

It's already live at digitaf.eu.

AI4A uses retrieval-augmented generation (RAG): instead of answering from the model's general memory, it first finds the most relevant passages in its knowledge base, then asks Mistral to write an answer grounded in them — with citations back to the sources.

flowchart LR
    Q(["Your question"]) --> E["Turn the question into<br/>a vector (HuggingFace)"]
    E --> S["Search the knowledge base<br/>for the closest passages"]
    KB[("Knowledge base<br/>Zenodo research + catalogue")] --> S
    S --> M["Mistral writes an answer<br/>grounded in those passages"]
    M --> A(["Answer + source citations"])

The knowledge base ships pre-built in public/data/, so a fresh clone works out of the box. See INTEGRATION.md for the full architecture, the embed widget, and self-hosting.

API keys for self-hosting (bring your own: free to start, billed to your accounts):

Add the assistant to any site with a single tag:

<script src="https://digitafmagicbox.vercel.app/widget.js" defer></script>

A chat button appears in the bottom-right corner. Set the title, colour, suggested questions, theme, and more via data-* attributes — see INTEGRATION.md for the full widget reference.

digitafmagicbox.vercel.app is the hosted app that serves the widget and API — use it (or your own instance's domain) in the script. digitaf.eu is the end-user site.

Running it locally is quick — clone, install, add your keys, run:

git clone <repository-url> digitaf
cd digitaf
npm install
cp .env.example .env.local      # then add your two API keys (see "What you need")
npm run dev                     # → http://localhost:3000

Open http://localhost:3000 and start asking. (Both keys matter: without MISTRAL_API_KEY the chat can't reply, and without HUGGINGFACE_API_KEY it answers from general knowledge only, since the knowledge-base search can't run.)

For production (any Node host):

npm run build && npm run start

Set MISTRAL_API_KEY and HUGGINGFACE_API_KEY as environment variables in your host — don't commit .env.local. Full hosting guide: INTEGRATION.md → Hosting Your Own Instance.

Without the keys the app builds and the UI loads, but the chat won't answer — MISTRAL_API_KEY is required for replies and HUGGINGFACE_API_KEY for knowledge-base search. Add both before real use.

Pre-built embeddings ship in public/data, so the app works out of the box. To rebuild or refresh from live sources:

npm run build:all-embeddings    # Zenodo research + catalogue

Rebuilding needs HUGGINGFACE_API_KEY, and the embedding model must match between build time and query time — see INTEGRATION.md → How Embeddings Work before changing anything. A GitHub Actions workflow (monthly-update.yml) can refresh the data automatically each month. For custom data sources and the API reference, see INTEGRATION.md.

MIT — see LICENSE, which covers this project's code. The knowledge base bundled in public/data/ is generated from Zenodo and catalogue sources that keep their own licenses and attribution. DigitAF is developed as part of the DigitAF initiative under the EURAF community.

Developer: Miska Nevalainen — miska.nevalainen@efi.int