New Providers, Bug Fixes & Improvements

[valentinfrlch]

1.5.1 Release Notes

⚠️ Please read the breaking changes before updating!

This update adds native support for OpenRouter and Azure, fixes an issue with Google Gemini, adds support for GPT-5 models and adds Dutch and Polish translations.

Contributors

A huge thank you to our contributors @TheRealFalseReality, @NikNikovsky, @Minionguyjpro, @meceonzo—and of course everyone who helped test and provided feedback for the beta versions!

⚠️ Breaking Changes

• Timeline snapshots have moved from /www to /media. add the following to your configuration.yaml:

  # Add llmvision /media folder
  homeassistant:
    media_dirs:
      media: /media
      llmvision: /config/media/llmvision

  💡 Make sure to create the /media and the /llmvision inside it before restarting!
  💡 Update LLM Vision Card to the latest version! Otherwise snapshots won't show correctly.

• max_tokens is now displayed as a box type number selector rather than a slider. This change accommodates newer models that use thinking parameters counting toward max_tokens. Unlike the limited slider, the box allows for inputting much larger numbers.

  💡If you notice empty responses, consider increasing max_tokens.

Integration

✨ Features

• New Provider: LLM Vision now supports OpenRouter natively. (@valentinfrlch)
• New Provider: LLM Vision now supports Azure natively (Support Azure Open AI #64, Added Azure OpenAI provider #103, Refactor providers into classes, Add Azure #144). (@valentinfrlch)
• Support for Polish language: Added 🇵🇱 Polish translations. Thank you @NikNikovsky!
• Support for Dutch language: Added 🇳🇱 Dutch translations. Thank you @Minionguyjpro!

🔧 Improvements & Fixes

• Google Gemini fix: Fixed an infinite retry loop in the Google Gemini provider and corrected an issue where the fallback provider wasn't being used properly. (Gemini endless retry loop #398, (Gemini) Resource has been exhausted (e.g. check quota) (image_analyzer) #262), (@valentinfrlch)
• Support for GPT-5 models: Remove temperature and top_p from request when a gpt-5 model is used (GPT-5 not support temperature #437), (@valentinfrlch)
• Moved Timeline snapshots: Snapshots are now stored securely behind Home Assistant authentication in /media. LLM Vision will attempt to migrate previously captured snapshots to the new folder.
• Remove max_tokens limit: max_tokens is now displayed as a box type number selector rather than a slider. This change allows for larger token limits, as thinking tokens count toward the max_tokens limit. (@valentinfrlch)
• Gemini key logging: Fixed a security issue that logged the Google API key in clear text (Debug logs are sharing (at least) Gemini API keys #334), (@valentinfrlch)
• Missing import: Add aiofile package to imports (aiofile package declaration missing in manifest.json #404), (@valentinfrlch)
• Sanitize IP address: Ignore any protocol input for the IP address field for Ollama and Open WebUI providers. Thank you @meceonzo!
• Fallback provider: Fixed a retry logic bug that caused an exception when no fallback provider was configured. (v1.5.1 beta - AttributeError: 'NoneType' object has no attribute 'get' #461)
• Title generation: Fixed an issue that caused an exception when generate_title was not set. (UnboundLocalError when generate_title is false (default) in service calls #463)

Blueprint

Update the blueprint by re-importing it from the Blueprint settings page in Home Assistant.

https://github.com/valentinfrlch/ha-llmvision/blob/main/blueprints/event_summary.yaml

A huge thank you to @TheRealFalseReality for maintaining the blueprint!

✨ Features

• See snapshot quick action: Notification now shows button to preview the snapshot in a browser.

🔧 Improvements & Fixes

• Multiple Cameras: Fixed an issue that displayed the wrong camera (Blueprint v1.5.0 shows wrong camera #403)
• Live Preview notification (iOS only): Preview mode can now be customized before and after analysis.
• Time format: Remove leading zero from 12-hour format.
• New media folder: Changed snapshot location from /www to /media/llmvision/snapshots (LLMVision 1.5.1 Beta 3 and Blueprint 1.5.1 problem with snapshot-folder #457)

---

This discussion was created from the release New Providers, Bug Fixes & Improvements.

[RoToR47]

Does this update add Gemini 2.5-flash support?

[Vitani]

A breaking change, but only a patch number bump?!

[zav13r]

Adding the configuration.yaml text here prevents home assistant from accessing other media. How do we ensure this is in addition to existing media (wallpapers, pictures etc) in the media folder rather than replacing them?

EDIT: think I found a solution. If you have not manually specified media directories previously in configuration.yaml, you also need to add a second entry to include the default media folder:

homeassistant:
  media_dirs:
    llmvision: /config/media/llmvision
    media: /media

[tabascoz]

HI, i have a similar problem with the new directory. As part of IOS alert, as per below:

actions:
        - action: URI
          title: See Snapshot
          uri: /media/llmvision/snapshots/xxxxxx.jpg

I receive the message in IOS but when i click the action it opens the companion app with a 401- unauthorzied, i also receve a log of Login attempted failed. How should I set the URL to have the screenshot visible?

[TheRealFalseReality]

the media location that you have provided is what it should be, but for now, the actual folder is /config/media/llmvision/snapshots/xxxxxx.jpg

[tabascoz]

thanks, but i still get 401 unauthorized. Are you able to use message actions and open the screenshot?

I was able before the move from www folder, but i know www is unauthenticated.

Struggling to learn how I get it to show, I did a research and HA generates temporary tokens for ex when you click on a camera snapshot in the button "download snapshot" and it generates a url link like that: blob:https://xxx.xxx.org:8123/a2e28xxx3-491xxxxx2-d66xxx/xxx8 this is a temporary token and i am able to open it for some time in another browser before it expires.

[TheRealFalseReality]

I think it has something to do with HA not letting you see what's in that folder. Are you using HA OS or Docker? It should work using HAOS but other install types you gotta do something extra.

[tabascoz]

none of them, I'm using bare metal (gentoo). I think i can live without it. Before it used to work when it was under www. I may create a hard link from snapshot to a folder inside www and I can circumvent the issue. Thanks

[danielszilagyi]

I tried to switch my blueprint based automation to use gpt-5-mini model via OpenAI integration but I only get this: I dont see the description or image details Please paste the text or describe whats in the images object and location and Ill create the title in the format Object seen at location under 6 words Examples Person seen at front door Dog seen at backyard
If I switch back to gpt-4o-mini everything is back to normal. There is no error in the logs. Other requests to the model via HA work, only the image analysis broke. I set max_tokens to max (300) but no change.

What could the issue be?

[valentinfrlch]

The gpt-5 models are reasoning models. They require significantly more tokens to "think". You can increase max_tokens, but you need to type the number (not use the slider). Just set it to a very large number (e.g. 10'000). I did some tests and the actual token consumption seems to be about 2500/call.

[danielszilagyi]

You can increase max_tokens, but you need to type the number (not use the slider). Just set it to a very large number (e.g. 10'000)

Thanks, the max value in the blueprint made me think I cannot adjust it to a higher value from the textbox.

[deergitseason]

The gpt-5 models are reasoning models. They require significantly more tokens to "think". You can increase max_tokens, but you need to type the number (not use the slider). Just set it to a very large number (e.g. 10'000). I did some tests and the actual token consumption seems to be about 2500/call.

The documentation currently recommends gpt-5-mini as the best combination of cost and performance, however it seems to require significantly more tokens than when I use gpt-4o-mini. I am able to get a usable result with 30 max tokens, however the lowest I can set gpt-5-mini to while still getting a result is 400.

Is there something to this equation I'm missing? (highly probable since I'm not very knowledgeable about tokens)

[valentinfrlch]

@deergitseason You're not missing anything. The benchmark may be slightly misleading.
It is true that gpt-5-mini is the best price/performance per 1 million tokens. However, since it uses more tokens than gpt-4-mini, it's difficult to compare these two models.

I recommend experimenting with different models. If you're happy with gpt-4-mini's responses, just stick with that.