Self-hosting an AI chatbot with RagLeap Core: what you run, what you need and the limits

RagLeap Core is open-source software (MIT) that you run yourself, on your own computer or server. The hosted RagLeap service that used to be reachable at this address has been retired. This short guide describes what you run, what you need, where your data goes, and what has not been verified. For the exact steps, use the install documentation.

What you run

One installer command checks that Docker is present, downloads the project, asks whether you want Gemini, Ollama or to skip the AI choice for now, generates an API key and secrets for you, and starts everything:

curl -fsSL https://raw.githubusercontent.com/ragleap/ragleap-core/main/install.sh | bash

The stack is a PostgreSQL database with pgvector, the app, a worker, a voice service, Neo4j for the knowledge graph, Redis, a sandbox for code, and Ollama for local models. Every port it publishes on your computer is bound to 127.0.0.1, so it is reachable only from that machine. A ragleap command wraps the daily tasks: launch, stop, status, logs, update and key (with key --rotate to issue a new API key). Windows has a PowerShell installer; see the limits below.

Which AI it uses

RagLeap Core is bring-your-own-key: there is no system-provided key for any provider. Gemini is the default. The supported chat providers (set with LLM_PROVIDER) include Anthropic, OpenAI, Mistral, Groq, a local Ollama model and any OpenAI-compatible endpoint; the README lists all 20. Embeddings use Gemini by default and can be switched with EMBEDDING_PROVIDER.

Where your data goes

Your documents, conversations and settings are stored on your own infrastructure. The text sent to get an answer or an embedding goes to the AI providers you configure; a local model through Ollama keeps those steps on your machine.

Approvals

Autonomy is off by default. In semi mode an AI employee asks before it acts; in full mode it does not, and full mode is opt-in. Roles detected as sensitive are held at semi even if you choose full.

Limits you should know

Tested by the maintainer on one Windows laptop and on Linux servers. macOS, Linux desktops and the Ollama option on Windows are not verified yet.
Local models on a CPU are slower and less accurate than Gemini.
No named users or roles yet, and no independent security review yet.
The compose file still ships a public default password for the graph database (Neo4j listens on localhost only). A random password per install is planned.

See the roadmap page for what is planned, and the guides for more detail.