Setting up Ohmventory on a Mac

Ohmventory consists of a database, a web search, the server and the interface. All four run in Docker. The language model that recognises photos and suggests projects runs alongside in LM Studio, directly on the Mac, because there it can use the graphics memory. At the end you reach Ohmventory in the browser on the Mac and from your phone on the same Wi-Fi.

What you need

• a Mac with Apple Silicon (M1 or later) and macOS 14 or later
• memory: 32 GB for both models; with 16 GB it works with the small model alone (see step 2)
• about 30 GB of free space for models and Docker images
• an internet connection for the downloads and the web search

1. Install the tools

Open Terminal. Git comes with Apple's developer tools:

xcode-select --install

Then download two programs and install them like any other Mac app:

• Docker Desktop in the version for Apple Silicon. Start it once and wait until the menu bar says "Docker Desktop is running".
• LM Studio. Start it once so that it creates its model folder.

LM Studio comes with the command-line tool lms. So that Terminal finds it:

~/.lmstudio/bin/lms bootstrap

Then open a new Terminal window. This shows you that everything is there:

git --version
docker version
lms version

2. Get the models

Ohmventory uses two models: a small one for the project assistant and the web search, and a larger one for photo recognition.

lms get lmstudio-community/gemma-4-e4b-it-mlx

For the larger model, search LM Studio's Discover tab for gemma-4-26b-a4b-it-qat and download the MLX version (about 16 GB). lms ls then shows the names of both models; Ohmventory needs exactly these names. If they differ from gemma-4-e4b-it-mlx and gemma-4-26b-a4b-it-qat-mlx, you enter them in step 3.

Mac with 16 GB: leave out the large model. The small one can recognise images too, just a little less accurately.

Now start LM Studio's server and load the small model with the context Ohmventory sizes its requests for:

lms server start
lms load gemma-4-e4b-it-mlx --context-length 8192

You can check it like this; the answer lists the models:

curl http://localhost:1234/v1/models

LM Studio loads the large model by itself as soon as the first photo recognition arrives. So that it does not start with too small a context, set the context length of both models to 8192 under My Models in LM Studio.

3. Get Ohmventory and configure it

git clone https://github.com/Marcel-B/Ohmventory.git ~/Ohmventory
cd ~/Ohmventory
cp .env.example .env

The file .env holds the settings. The following lines set random passwords, connect Ohmventory to LM Studio on the Mac and enter the address at which your phone reaches the Mac:

sed -i '' \
  -e "s|^POSTGRES_PASSWORD=.*|POSTGRES_PASSWORD=$(openssl rand -hex 16)|" \
  -e "s|^SEARXNG_SECRET=.*|SEARXNG_SECRET=$(openssl rand -hex 32)|" \
  -e "s|^LLM_GATE_URL=.*|LLM_GATE_URL=http://host.docker.internal:1234|" \
  -e "s|^PUBLIC_BASE_URL=.*|PUBLIC_BASE_URL=http://$(ipconfig getifaddr en0):8080|" \
  .env
grep -E '^(LLM_GATE_URL|PUBLIC_BASE_URL)=' .env

If PUBLIC_BASE_URL does not show an address like http://192.168.….:8080, the Mac is not on the network through Wi-Fi. Then use ipconfig getifaddr en1 and enter the address by hand (open -e .env).

What the settings mean:

• LLM_GATE_URL is the address of the language model. Ohmventory is built to share the model with other applications through a queue. On your own Mac it talks to LM Studio directly; from inside Docker, host.docker.internal is the Mac itself.
• LLM_MODEL and LLM_VISION_MODEL are the model names from lms ls. Leave them empty if they match step 2. On a Mac with 16 GB, set LLM_VISION_MODEL=gemma-4-e4b-it-mlx.
• PUBLIC_BASE_URL goes into the QR labels. If the Mac's address changes, print the labels again, or give the Mac a fixed address in your router.

4. Start

docker compose up -d --build

The first time, Docker builds the server and the interface, which takes a few minutes. Then you can see whether everything is running:

docker compose ps
curl http://localhost:8080/api/health

All four services show "running" (the database "healthy"), and the second line answers {"status":"ok"}. The server's log may contain a warning that the LLM gate cannot be reached; the queue only exists in a larger installation, and on the Mac you do not need it.

If a port is already taken, a service does not start. You can move the interface to another port with WEB_PORT=8081 in .env; the database needs port 5432 and the web search port 8888, so stop any PostgreSQL running locally.

5. The first component

Open http://localhost:8080.

1. Under Container, create a storage place, for example your sorting box. It gets a code like T-001, and you can print its label with the QR code straight away.
2. Under Erfassen (record), upload a photo of a component or breakout board. The first recognition takes longer because LM Studio loads the model first. You confirm the suggestion, enter the quantity and choose the container.
3. In the Assistent, ask for example "Can I build a thermometer?". Whatever is missing goes into the shopping list.

The interface is in German for now; more languages are on the roadmap.

If recognition does not work, look at the log under Developer in LM Studio to see whether the request arrives, and compare the model names with lms ls. Ohmventory's own messages are shown by docker compose logs -f api.

6. From your phone

On the same Wi-Fi, open the address from PUBLIC_BASE_URL on your phone, that is http://<address of the Mac>:8080. If macOS asks whether Docker may accept incoming connections, allow it. From then on, a QR label on a box opens its contents directly, and under Erfassen the phone takes the photo with its camera.

The interface has no login of its own. So do not put it on the internet through port forwarding.

7. Update and stop

You get a new version like this; stock and photos live in Docker volumes and are kept:

cd ~/Ohmventory
git pull
docker compose up -d --build

docker compose stop stops Ohmventory, docker compose start starts it again. LM Studio has to be running for recognition and the assistant to answer; everything else works without it. If you clean up with docker compose down -v, that also deletes the database and the photos.

How Ohmventory runs on a Proxmox server with automatic updates is described in the project's README.