01In one sentence#
You need a Linux server you rent (a VPS) with Docker, a web address you own, and either an AI provider key or the free local model. The installer does the rest in one command. No, it cannot be installed on shared hosting, cPanel, a WordPress host or a website builder: those cannot run Docker.
- A server running Ubuntu 22.04, Ubuntu 24.04 or Debian 12, with the sizes below.
- A web address, such as
chat.sharmafurniture.in, with a DNS record pointing at the server. - An AI answer source: a key from an AI provider (paid by use), or the free local model that runs on your own server (slower).
02Server size#
The engine itself is light. The weight is the search model it runs for you, and the free local model if you turn that on. The AI model that writes answers normally runs at your AI provider, not on your server.
| Setup | Minimum | Recommended |
|---|---|---|
| Without the free local model | 2 vCPU, 4 GB memory, 40 GB SSD (the installer adds 2 GB of swap) | 4 vCPU, 8 GB memory, 80 GB SSD |
| With the free local model (advice, to be re-measured) | 4 vCPU, 12 GB memory, 80 GB SSD | 8 vCPU, 16 GB memory, 100 GB SSD |
| A server with a GPU | Not offered by the installer. Follow the deploy guide by hand. | |
- Measured: the default stack used about 2.1 GB of memory on a CPU server with 270,000 passages of text stored. That is where the 4 GB minimum comes from.
- Advice, not measured: the local-model rows. The free local model needs about 6 GB of free memory for a 4-billion-parameter model, on top of the rest. We will publish measured numbers after testing on real servers.
- A “4 GB” server often shows as 3.8 GB. The installer allows for that.
- The disk fills with the images, the search model, your customers’ uploaded files and backups. Rent a disk you can grow.
03Operating system#
Ubuntu 22.04 LTS, Ubuntu 24.04 LTS or Debian 12, 64-bit Intel or AMD. A 64-bit ARM server may work but is not tested yet. The installer stops with a clear message on anything else; you can try it anyway with --force-os, at your own risk.
04Web address and network#
- A web address you control. A domain or a subdomain such as
chat.sharmafurniture.in, with an A record that points to the server’s address. Everything (the dashboard, the engine and the chat window) is served from this one address. - Ports 80 and 443 open at your provider, and not used by another program on the server (an old web server such as Apache or Nginx must be stopped). They are how the free HTTPS certificate is issued and how visitors reach you.
- Outgoing internet access. The server downloads Docker, the product’s images, the search model and the certificate; and it talks to your AI provider. The product sends nothing about your businesses, visitors or documents to us.
- A development address is not a live server.
localhost, a private address or a name ending.localor.testis treated as a development copy and cannot get an HTTPS certificate. The installer refuses these.
05Accounts and costs that are not in the price#
- The server you rent (a VPS), paid to your provider each month.
- An AI provider for the answers, paid by use, unless you choose the free local model (slower, no AI bill).
- An e-mail sending service (SMTP) if you want invitations and password resets sent from your own address. Without it, the first sign-in link is printed on screen by the installer.
- A Meta app for WhatsApp, only if you offer WhatsApp. Meta’s own review and fees apply.
- Support covers the product working as described. It does not cover installing it on your server, your server’s setup or your provider; this page and the deploy guide are there so you do not need that.
06Windows and macOS#
Windows and macOS are for trying the product on your own computer, with Docker Desktop; follow the step-by-step guide. A live server, the kind your customers’ websites talk to, runs on Linux.
07Install in one command#
Clone the repository on the server (or upload a release zip and unzip it), and run:
cd cob-rag-engine
sudo bash deploy/install.shIt asks two things: your web address and your e-mail. Then it checks the server, installs Docker if it is missing, sets everything up, gets the HTTPS certificate and prints the link you sign in with. Run it again at any time: it changes nothing that already works. Updates, backups and restores are one command each; see the deploy guide.