[Feature] Unified Context and Filename Search

[haoenz]

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

• Yes

The feature

I would like to suggest supporting context and filename search simultaneously in the search bar.

The ranking strategy for fuzzy filename matches might require further discussion, but for exact filename matches, assigning them the highest priority should be reasonable, right?

There are times when I know exactly the name of the file I'm looking for, yet I still have to repeatedly go through the mechanical steps of Open Menu → Switch Search Type → Search.
This process feels a bit frustrating and inefficient.

Would it be possible to streamline this experience by allowing both search types to be performed together?

Platform

• Server
• Web
• Mobile

[DankMemeGuy]

I think would be a good feature. I have had similar issues where it would be useful to have a unified search. My only concern is that while I think this may be useful to many people, the issue may be the compute required. I have noticed that even right now doing 1 'type' of search takes 20-45 seconds, so if you add that up over the 4 types, then you significantly add wait time, which may be a non-starter to folks.

That being said, I think this would be a useful feature, but I don't think it should be a default setting, unless it can somehow be optimized for those with lower spec'd servers.

[fhuhne]

May I ask what your specs are? I’m running on a 2017 Core i3 and get results in half a second … To me this sounds like your server is bottlenecked for ML workloads.

[DankMemeGuy]

May I ask what your specs are? I’m running on a 2017 Core i3 and get results in half a second … To me this sounds like your server is bottlenecked for ML workloads.

The main immich server is hosted on an 1821+ synology (64gb ram), and the ML is offloaded to two 3060 GPUs. there are quite a number of images though, and the main db stored on synology is on HDD not SSD so no doubt that impacts performance.

[fhuhne]

I knew Synology's don't have strong CPUs, but I didn't expect them to be this slow with ML. The DB on HDD probably also plays it's role. Thanks for sharing the insight! 20+ seconds does sound very inconvenient in which case I can totally understand your argument.

[DankMemeGuy]

solved the slow issue: #28229

I think would still be good to be cognizant of querying times, but I think it's probably OK

[SuperEllipsis]

I would love to expand this to include all the 4 different types (Context, Filename, Description, OCR), as I feel it would make my searches much more efficient and relevant, if allowed to search both context search and OCR as one query, or context and description as one query.

As an example, I am looking for a black SUV in a driveway with a specific number plate contaning "AB123". The "Black SUV in a driveway" could be the context search, while having the "AB123" as an OCR search and having that partially match the number plate

[justFreddy]

This would be perferct. Just select type if you want and by default all search types together.

[mbuzina]

I would like to have a single text input allowing me to search for everything, ML learned features, file- / directory name, descritpion, OCR, location city/state/country, person names, tags, camera make/model, lens info, album names, star rating, dates (year, month, day, day of the week), etc.

So a search like: photos taken with iphone in ireland showing sofia and max taken on a thursday would just "magically" work (note: add voice input on mobile for extra credits)

I have not looked at the ML implementation, but switching to a multi-model model and adding all textual information into the learning process might help here. If given all metadata it might even recognize things like slowmo video by itself.

[mbuzina]

I guess the best way to do this would be a to add an additional embedding for each asset, where the metadata is stored. Transform the metadata into a sentence like this:

Photo directory/filename taken on 1. January 2022 with a Canon EOS 500D and the Canon EF 50mm lens in Paris, île-de-France, France showing "Person 1, Person 2, ...", tagged holiday, city trip. It is contained in the albums "a", "b" and "c", rated 4 stars and is described as "description". Text in the image is "Text from OCR" and Metadata Details are "...."

When performing a smart search, combine both embeddings with a factor which even could be surfaced to the user.

If combining on search is too slow, story a normalized weighted sum of the embeddings.

--> I learned a few things from ChatGPT to come up with this.

• Added OCR Text to the senctence

[bigntallmike]

This might be easier to accomplish with a search within search results -- if we could save a search result list and then search within that, we could stack searches.

[Vogeluff]

My approach would be to add a filename filter in the extended search options to thereby combine the smart search with the filename search:

something like this:

seems like the easiest approach. stacking searches inside each other seems convoluted and somehow having the normal smart search also respect file names sounds like a nightmare in detection and would also always run both searches which would harm performance.

[Vogeluff]

I've created a draft PR with use of ai to visualize better what my approach/goal would be:
#27988

i hope someone who is capable of making these more technical changes sees this and perhaps sees the benefit of merging the currently split up search into one smart search with additional filters. i think the current approach is rather unintuitive and not very flexible.

@haoenz would this also solve your problem?

i also made a little video demo of my approach. as i said: way above my skill level where i feel comfortable working on it properly, but from the UX point of view i'd be happy to help with this.

demovideo.of.advanced.unified.search.immich.mp4

i just noticed i never showed that properly but the toggle in the search bar still works as it does now for quick searching those aspects. also you can use in the popup just the filename or ocr or desc fields to search only for that aspect aswell.

[manne01]

Transform the metadata into a sentence like this

Please always consider to involve tokens that can be extracted from the assets folder paths (especially in external libraries) as a high prio search information source. For example mylib/2026/05-vacation-paris-with-max-and-marie/day1-louvre