[Feature] Disable smart-search should disable contextual search from web UI

[charlotron]

I have searched the existing feature requests, both open and closed, to make sure this is not a duplicate request.

• Yes

The feature

Context:
I have tried to make smart search working on a little raspberry pi5 (8gb version) without success, trying multiple changes (modifying threads, timeouts, changing models..) to realise contextual search using ML in raspberry pi 5 is erratic on this system.

Due to this I have been forced to disable smart search to prevent rpi5 hangouts due to the pikes of cpu/ram usage of the machine-learning service.

As long as I have disabled the smart-search feature I am supposed to not to try to find using contextual search, but disabling that option does not make "disappear" this kind of search.

Request:
Disabling smart-search should make disappear "contextual search" and set "search by name" feature as the default search option
Maybe this should also happen in mobile apps.

Platform

• Server
• Web
• Mobile

[mertalev]

Not to dismiss your feature request (which I think is valid), but could you describe the issue you were facing with ML? 8GB should be comfortably enough if you're using the default model or a similarly lightweight model.

[charlotron]

I am facing problems with the smart search, the rpi5 takes all the cpu for itselfs, overheats and fully degradates the experience overall, and if you are transcoding videos or something that requires computation capacity. In some cases the rpi5 gets frozen and I have to restart it.

Besides the response of the api is not very good too many false positives, besides searching another keyword returns allways the same results which I cannot understand.

The only stable configuration in docker is:
immich-machine-learning:
image: ghcr.io/immich-app/immich-machine-learning:${IMMICH_VERSION:-release}
user: "0"
environment:
MACHINE_LEARNING_WORKER_TIMEOUT: 300
MACHINE_LEARNING_WORKERS: 1
MACHINE_LEARNING_MODEL_INTRA_OP_THREADS: 1
MACHINE_LEARNING_MODEL_INTER_OP_THREADS: 1
MACHINE_LEARNING_REQUEST_THREADS: 1
IMMICH_LOG_LEVEL: DEBUG
#MACHINE_LEARNING_PRELOAD__CLIP: nllb-clip-base-siglip__v1
#MACHINE_LEARNING_PRELOAD__FACIAL_RECOGNITION: buffalo_s
volumes:
- /etc/localtime:/etc/localtime:ro
- ${ML_CACHE_LOCATION}:/cache
- ${DOCKER_IMMICH_LOCATION}/ml_start.sh:/ml_start
entrypoint: /ml_start
#tty: true # Habilitar la interfaz de terminal
#stdin_open: true # Abrir stdin para interactividad
healthcheck:
test: ["CMD-SHELL", "pidof python || exit 1"]
interval: 1m
timeout: 10s
retries: 3
start_period: 30s
deploy:
restart_policy:
condition: any
resources:
limits:
cpus: '1.00'
#memory: '4G'
networks:
- overlay_network

Limiting as much as I can the ML instance, if I comment the environment variables that limit threads, etc.. I got better results but the rpi frozes easily and the container is getting killed and everything fall down in cascade.

I suppose that would be pretty nice if the ML would process the images extracting categories, and later it will be searched this categories by simple sql queries. (ie: detect a dog in an image, then search for dogs by the categories of an image and returns them, also those categories could be listed)

PDT: sorry for my english

[keesjebal]

I have the same issue on my Synology DS223 with 2GB of ram. Tried many things, but even with the default model and 1 single Smart Search job, the NAS is stuck with a CPU at 100% when running the face recognition (1 job) and the Smart Search job together. Immich is completely unavailable at that point. Since I don't use Smart Search I decided to disable it. This makes Immich still usable when importing new foto's, but I'm also loosing the duplicate-detecion which is based on Smart Search.

I don't think that there is a model available that will run smoothly on a DS223 with 2GB of ram.

I really like the idea to make the "search by name" the default when Smart Search is disabled.

[charlotron]

I think this solution is powerfull but you need a decent hardware to run it, DS family of synology has a very low amount of cpu power and 2gb of ram to run multiple docker containers would probably be not enough. ML models also requieres a good amount of ram as defined in immich smart search documentation https://immich.app/docs/features/searching/ (if you go down it is shown the model requirements just for running the model, and you have to run the rest of the application containers)