Ollama

What it's good for#

Online AI chat services see everything you type into them. Ollama lets you download an open model, such as one of the Llama, Mistral, Gemma or Qwen families, and run it yourself. The answers come from your server, and nothing you ask goes anywhere else.

LibrePortal's Ollama app is really two parts. Ollama holds the models and does the work, with an API but no page of its own. Open WebUI is the chat page in front of it, with conversations, several accounts and model management. The Web Interface button opens Open WebUI.

Signing in#

Open WebUI normally makes the first person who signs up its administrator, which on a server anyone can reach would be whoever gets there first. LibrePortal doesn't leave that to chance. It creates the admin account itself before the page is ever served, and sign-up is then switched off.

The account signs in with the email address admin@ollama.invalid (it doesn't need to be a real one) and a password made for you. Both are in the Login Details shown when the install finishes; press Welcome on the app's page to see them again. Change the password inside Open WebUI if you like, but note that LibrePortal keeps showing the one it made, so keep the new one somewhere yourself.

If you installed Ollama with 0.4.6 or earlier, your Open WebUI kept its accounts and chats inside a container that is replaced on every update, so it came back empty each time, ready for the next visitor to claim. From 0.4.7, whenever Ollama starts and finds Open WebUI with no accounts at all, LibrePortal makes the admin account for it straight away (it waits up to two minutes for Open WebUI to answer), with the password shown in Login Details. That closes the door, but the chats are still lost at each update, and sign-up stays switched on.

To fix it for good, press Update Configuration on Ollama's Config tab. That sets Ollama up again from LibrePortal's current version of it, which gives Open WebUI a folder of its own on the server and hands it the admin account at its first start, with sign-up off. Your models stay where they are. Chats made since the last update are lost this one last time, as they would have been at the next update anyway; from then on everything is kept.

Getting your first model#

A new install has no models, so the chat page has nothing to talk to yet. The quickest way to get one is from the chat page itself:

Admins can also download, update and delete models from Open WebUI's admin settings, under its Ollama connection. Models live on the server's disk in Ollama's folder, so check you have the space before downloading several big ones.

Adding people#

Sign-up is off, so people don't make their own accounts. As the admin, add them from the users section of Open WebUI's Admin Panel, or turn sign-up back on in its settings if you'd rather people registered themselves. Open WebUI puts a new sign-up in a pending state by default, so nobody can chat until an admin approves them. Each person gets their own chats.

Using the models from other apps#

Ollama's API is on its own port, shown as Ollama API on the Config tab's Network section. It's the standard Ollama API, and it also answers the OpenAI-style requests many tools understand, so apps and editor plugins that ask for an "Ollama URL" or an "OpenAI-compatible endpoint" can use your server. Give them your server's address and that port, for example http://192.168.1.50:5300.

Open WebUI's page, on the other hand, is public to begin with, because it has its own sign-in. If only your household uses it, making it private too costs nothing and hides the sign-in page from everyone else; use the VPN when you're out. With a domain and Traefik, the chat page is at ollama.<your domain>; the API stays on its port.

Good to know#

• Backups include your models. A snapshot copies all of Ollama's files, and downloaded models can be many gigabytes, so backups get much bigger once you have a few. Your chats and accounts in Open WebUI are backed up too. To check: decide whether Ollama's models folder should be marked as a cache (CACHEDIR.TAG) like LibreTranslate's, since models can be downloaded again
• Updates are automatic for both parts, with a backup first. Both follow their projects' release channels rather than a pinned version, so Open WebUI in particular changes quite often.
• One per server. Every app that uses it points at one address, and a second copy would only duplicate the models.
• Memory is taken while a model is loaded. Ollama keeps a model in memory for a few minutes after it last answered, then lets it go, so the server's memory use rises and falls as people chat.