Quickstart

I3K RAG Enterprise Community is a single binary. It starts its own vector database and its own inference engine, so there is nothing to compose, nothing to build, and no container runtime to install.

This guide takes you from a downloaded tarball to your first cited answer.

Requirements

• OS: Linux (x86_64 or ARM64) or 64-bit Windows
• RAM: 16 GB minimum, 32 GB comfortable
• Storage: 30 GB free — the first run fetches roughly 12 GB of models and components
• GPU: an NVIDIA GPU is detected and used automatically. CPU-only works too, more slowly
• Network: needed for the first run only

Install

Download the tarball for your platform from the releases page. The archive's contents are flat — the binary, .env.example and frontend/dist/ — with no wrapping folder, so create a directory first:

mkdir i3k-rag-engine
tar -xzf i3k-rag-engine-linux-x86_64.tar.gz -C i3k-rag-engine
cd i3k-rag-engine

Exactly one setting has no default: the secret that signs session tokens.

cat > .env <<'EOF'
AUTH__JWT_SECRET=change-this-to-a-long-random-string
EOF

Then run it:

./i3k-rag-engine

What happens on the first run

The engine downloads what it needs and checks each component against a sha256 pinned in its manifest. A mismatch aborts the run rather than continuing with an unverified file.

• The vector database (Qdrant) — fetched and started by the engine itself
• The inference engine (eullm) — fetched, started and kept up to date automatically
• The chat model and the BAAI/bge-m3 embedding model
• OCR language data for Italian and English

That is about 12 GB in total, so allow time for it. Every later start is immediate, because components are only re-fetched when missing or failing their checksum.

When it is ready your browser opens at http://localhost:8000, and the initial admin password is printed in the log:

Set AUTH__ADMIN_DEFAULT_PASSWORD in .env beforehand if you would rather choose it yourself.

First upload

Log in as admin and upload a document. Supported formats are PDF, DOCX, XLSX, HTML, TXT, Markdown and CSV.

Scanned PDFs need no special handling: when a page yields too little text it is detected as a scan, rasterised and passed through OCR in Italian and English.

On upload the engine runs four stages:

1. Extraction — text is pulled out of the file, via OCR where needed.
2. Chunking — the text is split into overlapping chunks.
3. Embedding — each chunk is encoded with BAAI/bge-m3 (1024 dimensions, 100+ languages), in-process.
4. Indexing — vectors are written to Qdrant with the metadata used for role filtering.

First question

Ask a question from the chat interface. The engine embeds the question the same way, retrieves the closest passages, and streams an answer token by token with each source shown.

Because the embedding model is multilingual, the language of the question does not have to match the language of the document: an Italian question can find its answer inside an English contract.

Conversations are kept, and history is carried into follow-up questions.

Upgrading

DATA__DIR defaults to the directory holding the executable, and every component is fetched only when missing or failing its pinned checksum. So extracting a new release over an existing install — same directory, overwriting the binary, .env.example and frontend/dist/ — reuses everything already downloaded there. Your own .env is never touched, because the tarball only ever ships .env.example.

Extracting into a fresh empty directory instead starts every download over, since nothing at that new path exists yet.