A second machine for search and faces

The server finds faces and makes the vectors behind visual search itself. Another computer with a GPU can share that work by running Lumiverb's AI worker. This page sets one up on a Mac or a Linux machine and adds it to your server.

What the AI worker is#

The AI worker holds no library, no account and no files. The server sends it each photo's preview (and frames of videos, for faces) and it sends back what it found. It uses the same models as the server, so a photo comes out the same whichever machine did it. It uses an NVIDIA GPU, or a Mac's own GPU, when there is one.

It does visual search and faces only. Descriptions, text in photos and transcripts are other AI machines' work: see AI machines.

The first photos it gets download its models once, into its own data folder.

On a Mac#

Install Lumiverb with Homebrew, as for a server:

brew trust --tap bubbafat/lumiverb

brew install bubbafat/lumiverb/lumiverb

Don't start its service (brew services start lumiverb): that's a whole server. Instead, install the worker as a service of its own:

lumiverb-server ai-worker --env "$(brew --prefix)/etc/lumiverb/env" --install

It starts now and again at login, and restarts itself after brew upgrade. Its log is ai-worker/worker.log in the install's data folder (DATA_DIR in the env file). To take it away:

lumiverb-server ai-worker --env "$(brew --prefix)/etc/lumiverb/env" --uninstall

On Linux#

On a Linux machine with Lumiverb installed, run the worker as a systemd unit. Create /etc/systemd/system/lumiverb-ai-worker.service with:

[Unit]
Description=Lumiverb AI worker (search and faces for another Lumiverb)
After=network-online.target

[Service]
User=lumiverb
EnvironmentFile=/etc/lumiverb/env
ExecStart=/opt/lumiverb/.venv/bin/lumiverb-server ai-worker
Restart=always
RestartSec=10

[Install]
WantedBy=multi-user.target

Then start it, now and at every boot:

sudo systemctl daemon-reload

sudo systemctl enable --now lumiverb-ai-worker

Its log:

journalctl -u lumiverb-ai-worker -f

Its key#

Every request to the worker needs its key, so only your Lumiverb can use it. The worker makes one the first time and keeps it. On a Mac:

lumiverb-server ai-worker --env "$(brew --prefix)/etc/lumiverb/env" --show-key

On Linux, once the unit has started, it's in the data folder (DATA_DIR in /etc/lumiverb/env, /var/lib/lumiverb unless you chose another):

sudo cat /var/lib/lumiverb/ai-worker/key

Add it to your server#

As an admin, in the Lumiverb it should help:

1. Settings › AI › Add machine. Leave Where at Custom URL.
2. Name: anything, such as "Gaming PC".
3. Endpoint URL: http:// and that computer's address, then :6366/v1, such as http://192.168.1.30:6366/v1.
4. API key (optional): the worker's key.
5. Click Connect. It says "Connected: offers ViT-B-32, buffalo_l".
6. Under Does, tick Visual search (CLIP) and Faces, then Add.

The machine's row then says "Online · checked …", or "Offline:" and why. Lumiverb checks every machine every minute.

The server's own Built in machine keeps doing visual search and faces too, and work is spread across both. To leave them to the worker, click Edit on Built in and untick them under Does.

Requests at once#

Each machine works on one photo at a time on its GPU. Requests at once in Settings › AI (2 to begin with) is how many the server sends it together, so the next is ready when one finishes. The worker itself takes 2 at once unless started with --at-once N.

Its network#

The worker listens on port 6366 on every network interface, over plain HTTP. --port and --host change that; on a Mac, give them with --install so the service keeps them. Its key keeps others from using it, but the previews it's sent cross your network unencrypted, as they do to an Ollama machine. Run it only on a network you trust.

If Connect can't reach it, check that the worker is running (its log), that the address and port are right, and that nothing on that computer blocks port 6366.