ИИ-агент двойного назначения

В наши дни каждый разработчик, наверняка, пробовал вайбкодить, а некоторые идут дальше и заводят себе целых ИИ агентов. Однако отовсюду доносятся новости о том, как какой-то AI агент удалил базу данных со всеми бэкапами. Поэтому давайте посмотрим исходных код проектов, которые так или иначе связаны с агентной разработкой.

https://habr.com/ru/companies/pvs-studio/articles/1041056/

#go #golang #localai #photoprism #агентная_разработка #tau

أطلق تطبيق PhotoPrism تحديثاً جديداً ينقل ميزة التعرف على الوجوه بالكامل إلى نظام يعتمد على ONNX، مما يتيح مرونة أكبر في استخدام نماذج الذكاء الاصطناعي. كما تم إعادة تصميم شريط المعلومات الجانبي لتسهيل إدارة البيانات الوصفية، وإضافة دعم لتسريع ترميز الفيديو عبر تقنية Vulkan وقراءة صور HEIC وAVIF وملفات PSD. يتضمن التحديث أيضاً تحسينات في تصفية الملصقات، ودعم ميزة السحب والإفلات لرفع الملفات، وترجمات جديدة للواجهة بالاستعانة بأدوات الترجمة الآلية.

#PhotoPrism #ONNX

Tornare da un viaggio significa quasi sempre ritrovarsi con una quantità ingestibile di foto. Nel caso di Lisbona, il problema non era tanto archiviare gli scatti, quanto riuscire a estrarne una ventina davvero condivisibile: belle, sì, ma anche varie e capaci di raccontare l’esperienza nel suo insieme. PhotoPrism offriva già un’ottima base grazie a geolocalizzazione, riconoscimento facciale, label e strumenti di organizzazione, ma non aveva ancora un modo per comporre automaticamente un album con “le foto più belle” e soprattutto con sufficiente varietà.

Da qui è nata l’idea di un selezionatore AI: una piccola applicazione Java che usa PhotoPrism per recuperare le miniature delle immagini e Ollama per far lavorare due modelli AI, uno multimodale per assegnare un punteggio estetico e produrre una descrizione oggettiva, e un secondo modello testuale per raggruppare semanticamente le foto e selezionarle con più equilibrio.

Il problema vero non era la qualità

Il primo prototipo faceva una cosa molto semplice: prendere le foto da PhotoPrism, inviarle a un modello multimodale su Ollama e chiedere un voto estetico da 1 a 100 insieme a una breve descrizione. Sulla carta sembrava sufficiente, ma in pratica produceva una selezione monotona: immagini molto belle singolarmente, ma spesso troppo simili tra loro.

Era il classico caso in cui un ranking puro ottimizza la qualità locale ma non la copertura narrativa. Se cinque foto dello stesso scorcio o dello stesso momento ricevono voti alti, un algoritmo ingenuo tende a sceglierle tutte. Per costruire un album da condividere, invece, non basta premiare le immagini migliori: bisogna anche evitare la ripetizione.

[…]

WriteFreely + PhotoPrism

Współpraca z Immichem pod kątem wrzucania jakoś z automatu zdjęć z udostępnianych albumów szła jak po grudzie. [...]

https://blog.argilus.online/czach/writefreely-photoprism

Using Photoprism and Immich in Tandem Across Three Machines

I have an HP laptop running Photoprism, and two Raspberry Pi 5 running Immich. One of the Pi devices is running Immich, solo, and is made accessible via Tailscale when I am on the move. Photoprism runs on the HP laptop and is not available at all times. When the laptop sleeps, Photoprism is not available. This means that I have to consciously decide to ingest the photos to photoprism.

Propagating from Photoprism to Immich

I setup Photosync on my phone to upload the most recent photos and organise them by year and month. This is easy to forget, so I consciously set that up. The reason for this decision is two-fold. The first reason is that by organising photos in this manner it is easy to mirror them to multiple drives, as well as to make them accessible to Immich, and Photoprism on the second Pi I mentioned earlier.

At the time when I set this up I was using the Ubuntu machine daily so every morning I could upload the photos with photosync.

Woupsie Daisies

Recently I allowed the HP laptop to run out of battery, and when it rebooted everything seemed fine, so I went to sync between devices and that’s when the “woopsie daisisies” incident occured. The “Woopsie Daisies Moment” was both catastrophic, but luckily recoverable. The SD card that I use within the laptop didn’t mount properly but I didn’t notice. When I told rsync to sync from the HP SD card to the external hard drive locally, and to Immich on the Pi it deleted all the files on the Mac spare drive. I suddenly noticed that Nextcloud had no more photos so I killed the sync within seconds. When a delete command is sent it goes for a while before the kill command was received.

The impact was that my secondary backup was wiped, and that my cloud backup, as a result was also in danger. When I realised I unplugged drives, safely, to isolate them, until I understood what happened.

Drive not Mounted

When I looked at the mounted drives on the HP machine I checked to see if the external hard drive was mounted. It was, but the photos were gone. It occured a few weeks ago so I don’t remember if the Pi5 was also wiped, but I don’t think it was. I then realised that the primary drive that I was using had not mounted after the dead battery incident. I mounted the drive, and the photos were intact.

A Sync Shell Script

To avoid this mistake from occuring again I made two key changes. The first is that I removed –delete from the rsync commands. If a file is not present rsync will not delete anything. Accidental deletion is almost instanteneous. The second change was to make sure that the shell script checked that the drives are mounted before doing anything. If the drive is not mounted it aborts the transfer.

Now I sync from the local SD card to the local external hard drive. When that’s done I sync the external hard drive to the Pi5 currently running Immich before then syncing from Linux to MacOS and my kDrive cloud instance. Within two or three minutes I have things backed up across the HP machine, the MacOS machine, a Pi5, and Infomaniak.

The Standalone Immich Backup

I have an entirely seperate workflow for my second Pi5. I synchronise to this Pi5 via the Immich app directly. In theory, within seconds of taking a photo it can be synced and indexed by this Pi5. If the Photoprism and Immich suffers an issue such as a describe above, this one is entirely seperate and isolated. A mistake with one worfkflow only affects one workflow.

The Tertiary Backup

The primary worflow goes from Photosync to Photoprism, and then Immich across multple machines and drives. The secondary worflow is from the Immich iPhone app to the Immich instance on a seperate Pi. The tertiary backup is kDrive. All of the photos are automatically backed up by kDrive in the year/month folder structure, so if the primary and secondary workflow fail, I can still revert to the tertiary solution.

Each library uses the same structure therefore migration data from one library to the other should be seamless. This gives me the opportunity to continue learning, and if I make a mistake I have several recovery solutions.

And Finally

The key mistake was not to check that the volume was available. If I had done that, then I would not have run rsync. I have removed that risk by replacing manual operation with a shell script. The second mistake was to have a delete flat in an rsync workflow by default. I removed it. The third mistake was to allow the battery on the laptop to run flat. That’s why when the machine was rebooted it did not mount correctly.

If you see that files are missing, the first thing to do is check that the volume is mounted. If I had done that, then I would have seen it wasn’t and I could have mounted it, and the issue would have been resolved within seconds.

The current workflow is now more resilient, and automated.

The "convenience" of Google Photos often comes at the cost of your privacy. If you're tired of big tech scanning your family memories to build a data profile, it's time to switch.

PhotoPrism uses AI to tag your photos and recognize faces locally on your own instance. Same smart features, zero data mining.

Host it on PikaPods from $6.20/mo (sponsor features included): https://pikapods.com/pods?run=photoprism

#OpenSource #Privacy #PhotoPrism