RAGApp

Last reviewed Sep 13, 2026 · 820 words

RAGApp ships with no authentication of any kind. The admin panel at /admin, where you paste an API key and upload the documents you want to chat with, is open to anyone who can reach the port, and the project's own README says so and tells you to protect it yourself. That is the one fact to carry away from this page: RAGApp takes about 5 minutes to stand up, and it is only safe once something in front of it (a reverse proxy with basic auth, a Tailscale-only bind, a firewall rule) gates the admin path. Put that in place before you upload anything you would mind a stranger reading.

What the container actually contains

RAGApp comes from the LlamaIndex team, and it is best understood as a LlamaIndex chat engine with a settings page bolted on. One image serves three things: an admin UI for choosing a model provider and dropping in files, a chat UI for end users, and an OpenAI-style API endpoint so other software can call the same agent. There is no separate database, no Redis and no external vector store; the index lives on disk inside the container, so mount a volume or you will be re-uploading after every recreate.

docker run -d --name ragapp -p 8000:8000 \
  -v ./ragapp-data:/app/data \
  ragapp/ragapp

Open http://host:8000/admin, pick a provider, upload documents, and the chat is at /. The catalogue lists it at 4,440 stars, Apache-2.0, TypeScript, difficulty Easy, and about 1 GB of RAM before any local model is involved. That figure is honest for the app itself; the model is a separate bill.

The "agentic" part of the name means you can switch on tools alongside retrieval: web search, a code interpreter, a document generator. In practice nearly everyone I know who runs it uses it as plain RAG, asking questions of a folder of PDFs, and the RAG vs fine-tuning piece explains why that is usually the right instinct.

Local models work, with two caveats

The provider list covers the hosted names plus Ollama, which is where self-hosters go. Point the base URL at your Ollama host (http://host.docker.internal:11434 on Docker Desktop, the LAN IP on a Linux box) and pull two models, not one: RAGApp needs an embedding model such as nomic-embed-text as well as a chat model, and the admin page will fail its connection test without both.

Caveat one is speed. Embedding a few hundred pages on CPU takes minutes, not seconds, and every re-upload repeats it. Caveat two is quality. Retrieval with a 7B or 8B chat model answers noticeably worse than the same documents behind a hosted frontier model, and RAGApp gives you few knobs to compensate: a system prompt, a model choice, and not much else. That is the "limited customization" in the catalogue's cons, and it is accurate.

Where it stops being enough

RAGApp has no users. One admin panel, one document set, one chat. There is no per-person history stored server-side, no way to give the sales team one corpus and engineering another, no roles and, as covered above, no login. It is also a young project, first released in 2024, and its commit cadence has been quieter than its star count suggests, so do not assume a bug you hit will be fixed on any schedule.

If you need any of the missing pieces, the bigger projects among AI chat UIs are the safer pick and I would say so plainly. Open WebUI has accounts, groups, built-in RAG with a document library, and native Ollama support. AnythingLLM organises documents into workspaces with their own permissions, which is exactly the multi-corpus case RAGApp cannot do.

Where it fits anyway

Two uses justify it. The first is the prototype week: a team wants to know whether "chat with our docs" is worth building, and RAGApp lets a non-engineer configure the whole thing from a browser in an afternoon. The second is the embedded backend: you want an agent endpoint inside some other application and do not care about the bundled UI at all. In both cases its single-container simplicity is the feature, and its lack of multi-user is irrelevant.

What I'd do

Run RAGApp for the prototype behind Caddy with basic auth on /admin, reachable over Tailscale only, with Ollama on a GPU box if you have one and a hosted model if you do not. Give it 2 weeks. If the idea sticks and more than one person needs it, move to Open WebUI for general chat plus RAG, or AnythingLLM if separate document workspaces matter, and retire RAGApp rather than trying to harden it into a service it was never designed to be.