OpenAI

Avatar
Jürgen Jürgenson
  • updated
  • Under review

It's year 2023, the AI bots will take over the "world" so here is a template for it for bOS too...Kodu-Http (OpenAI) (1.2).bos

Right now it works just as a outdated google search with better logic or smth. Somehow the image generator crashes the http driver. Remember everything is in beta.

To get it working you need to make a account for OpenAI and generate your API key.

1. insert it here - for both completions and image generator commands YOUR_KEY.

Image 4432

2. You can ask the AI your question and it should respond. Made a TTS for it's  answers too so its extra creepy.

Image 4434

FYI it's not as good as the web page bot - for longer answers sometimes it doesn't give out the full info and I've seen that some of codes may go missing. In theory you can make it give you suggestions automatically too. If it gives you a BadRequest error see if the input is all in one line. It does not like it when some text is on a new line.

Happy new year!

Replies 20
Avatar
0
Jürgen Jürgenson

Also sharing my prompt for easier translations for bOS. I use Gemma3:12b with Ollama/OWUI and it works quite well. 10 indexes take about 1-2 minutes.

Setting: Temperature: 0 (Less errors)

num_gpu (Ollama): 256 (depends on your system)


Prompt, change it as needed for you - I needed translations for EN, ET and RU: 

Translate the <text> values in the following XML into English (EN), Estonian (ET), and Russian (RU) without altering the XML structure or order. Ensure each <table1> entry includes translations while preserving the original tags and layout. Do not skip any entries. Maintain the exact XML format and follow the provided template for each entry. If the text is already in English, copy it into the <en> field.  Please dont add <newdataset> and </newdataset> if its not pasted in as the template.
Translation template:

<table1>
<index>0</index>
<text>Are you sure?</text>
<en>Are you sure?</en>
<et>Oled sa kindel?</et>
<ru>Вы уверены?</ru>
</table1>
<table1>
<index>1</index>
<text>OFF</text>
<en>OFF</en>
<et>VÄLJAS</et>
<ru>ВЫКЛ</ru>
</table1>
<table1>
<index>2</index>
<text>ON</text>
<en>ON</en>
<et>SEES</et>
<ru>ВКЛ</ru>
</table1>
Avatar
0
Jürgen Jürgenson

Http Ollama-deepseek-r1 (0.06).bos

Setting Up Open-WebUI for Ollama Integration (With/Without Apache Proxy)

I recently integrated Open-WebUI with Ollama to create an enhanced user interface with a streamlined experience. The system is fully operational for core functions, including file uploads and text-to-speech (TTS), though microphone input (speech-to-text) remains unresolved.

1. Installation Overview

  • Open-WebUI Setup:
    Installed locally via pip (non-Docker) on the primary server, running on port 8080.
    Note: Multiple installation methods exist – this pip-based approach was chosen for simplicity.
  • Reverse Proxy Configuration:
    A Raspberry Pi running Apache handles HTTPS traffic forwarding to Open-WebUI.
    Why? Modern Android WebView blocks HTTP connections – HTTPS via proxy ensures mobile compatibility.

2. Accessing the Interface

3. Functional Features

  • Ollama Chat Interface: Full functionality
  • File Uploads:
    Successfully integrated for image analysis
    Limitation on API: Base64 encoding still required for API-based image processing
  • Text-to-Speech (TTS): Operational

4. Current Limitations

  • Microphone Input (Speech-to-Text):
    Non-functional
  • Capture/camera input does not work.
    Anyways it looks a lot better this way and does not have bOS quirks.

Image 5262

Avatar
0
Jürgen Jürgenson

Http Ollama-deepseek-r1 (0.05).bos

Added Vision Support
The feature currently works with BakLLaVA and LLaVA models. To use it, you’ll first need to download/pull one of these models.

Image Format Requirements
Images must be provided in base64 format rather than as URLs hopefully it will change in upcoming updates. However, during testing in PS, links appeared to work inconsistently—this discrepancy between direct testing and API behavior is still under investigation.

Current Focus
I’m exploring methods to implement base64 encoding exclusively within bOS.

Avatar
1
Jürgen Jürgenson

Http (OpenAI) (1.7) for public.bos

Made few changes. Added some more fine tuning options. Also the new gpt4.5 model.

Avatar
0
Jürgen Jürgenson
Quote from Jürgen Jürgenson

So there's a new kid on the block. More info on its API can be found here: https://github.com/ollama/ollama/blob/main/docs/api.md#api
I made a quick and dirty template for it, do what ever you like with it. And it is running locally and it is able to communicate with bOS. Deepseek-r1 template v0.01

Deepseek Template for bOS: The Ultimate Guide (or at least a "decent" one)

Introduction
Congratulations! You've just acquired the legendary Deepseek template for bOS. This isn't just any template—it's the template. It seeks, it finds, and sometimes, it even works on the first try! (No guarantees, though.)

Getting Started

  1. Open bOS and load the Deepseek template. If you don’t know how to do that, maybe ask Google or your friendly neighborhood smart home specialist.
  2. Once loaded, take a deep breath. Everything looks complicated, but don't worry—only about 73.5% of it actually matters.

Installing Ollama on Windows (So Deepseek Can Actually Work)

  1. Download the Ollama
    installer for Windows from the official site. If you end up on a shady website offering "free bonus tools," reconsider your choices.
  2. Run the installer and follow the prompts. If something goes wrong, blame Windows Update.
  3. Once installed, open a terminal (Command Prompt or PowerShell) and run:

    ollama pull deepseek-r1

    This will download Deepseek-r1 or what ever model you choose. Yes, it might take a while. No, you don’t need to stare at the progress bar (but we know you will).

  4. After it's downloaded, start the model with:

    ollama run deepseek-r1

    If all goes well, you should see some text output instead of errors. Congrats, you’re halfway there!

Configuring Deepseek in bOS

  1. Import the Template – Go to bOS config and import the template like a pro.
  2. Adjust the IP Address – If Ollama is running locally, keep it on localhost and default port 11434 If it's on another machine, change the IP in the settings. (No, 192.168.1.1 is not always the right answer.)
  3. Check if it shows as connected.
  4. Set Parameters – Adjust the values as needed Select the correct model you downloaded, insert a prompt and hit send. If something breaks, remember the sacred IT mantra: turn it off and on again.
  5. Test It! – Click the send button and watch in awe as Deepseek does its thing (or sits there doing nothing—both are possible).

Troubleshooting (a.k.a. “Why is this not working?”)

  • Nothing happens? Check if the device is actually powered on. It sounds obvious, but you'd be surprised.
  • Weird values appear? Congratulations! You've just discovered the joy of debugging. Start by verifying the input data, then take a coffee break—this will help maintain sanity.
  • Everything’s broken? Reload the template and pretend it never happened.

Final Thoughts
You are now the proud user of the Deepseek template. Will it change your life? Probably not. But it might make bOS a little bit smarter, and that’s a win in our book.

Good luck, and may your smart home always obey your commands (eventually). Tutorial made by it's competitor ChatGPT

Http Ollama-deepseek-r1 (0.02).bos 

Fixed some errors and made a main control panel. Also added model Pull request. Still a lot to do.

Avatar
0
Jürgen Jürgenson
Quote from ComfortClick Support

Hello Jurgen,

Ah i see what's going on.. So the longer the response, the more time we wait. For general chatting, it's quite slow at this point, even in PowerShell... To get a full response, in some cases we wait a minute or two to finish everything..
Same goes in bOS.. If the response is very long, we wait till it finishes.. then post the response.. Which in some cases feels like it's not working.. but a minute or so later, we get a response.

For what it is at this point, it seems to work quite well! Just a bit slow :)

Best regards.

Maybe you can look into the http streaming. If streaming=true it would be faster. Here is a same prompt for a joke made 4 times. Times are different even if the prompt is the same. 

llama_new_context_with_model: KV self size = 384.00 MiB, K (f16): 192.00 MiB, V (f16): 192.00 MiB

llama_new_context_with_model: CUDA_Host output buffer size = 0.60 MiB

llama_new_context_with_model: CUDA0 compute buffer size = 307.00 MiB

llama_new_context_with_model: CUDA_Host compute buffer size = 14.01 MiB

llama_new_context_with_model: graph nodes = 1686

llama_new_context_with_model: graph splits = 2

time=2025-01-30T13:40:57.000+02:00 level=INFO source=server.go:594 msg="llama runner started in 3.26 seconds"

[GIN] 2025/01/30 - 13:40:58 | 200 | 3m41s | 192.168.1.100 | POST "/api/chat" I think this was a bug as i had it paused in PS - it calculated both jokes and sent them out together. Usually for a small joke it takes 5-7s

[GIN] 2025/01/30 - 13:40:59 | 200 | 2m20s | 192.168.1.100 | POST "/api/chat"

[GIN] 2025/01/30 - 13:41:02 | 200 | 18.1981495s | 192.168.1.100 | POST "/api/chat"

[GIN] 2025/01/30 - 13:43:50 | 200 | 52.7484304s | 192.168.1.100 | POST "/api/chat"

Avatar
0
Jürgen Jürgenson
Quote from ComfortClick Support

Hello Jurgen,


So right out of the bat, when i started sending out commands, the HTTP driver would automatically disconnect, i think it had issues with the model number selection list. If i ran the /api/tags, to get the model number i could get the response for the model number to be : deepseek-r1:latest.. So i created a new string with deepseek-r1:latest as value and use this in the Send command to always try to use the "latest model".

Which seemed fix the issue and all the commands seem to work ok now. It does take quite some time to get an answer (for me it's like 30s.. How's it like on your end?


Best regards.

Yep I added the model later on, didnt update the template in here tho(Edit button went missing by that time). I've only tested deepseek-r1:14b and 9b so far. I haven't timed the responses yet but it isnt that bad under 5 minutes usually(20,76s without stream with 14b) it would show its thinking part sooner if the streaming flag would be true. Ollama is running on a 5950x and 3080ti right now so it is what it is. If you run it without an GPU then it will be a lot slower.

Avatar
0
ComfortClick Support

Hello Jurgen,

Ah i see what's going on.. So the longer the response, the more time we wait. For general chatting, it's quite slow at this point, even in PowerShell... To get a full response, in some cases we wait a minute or two to finish everything..
Same goes in bOS.. If the response is very long, we wait till it finishes.. then post the response.. Which in some cases feels like it's not working.. but a minute or so later, we get a response.

For what it is at this point, it seems to work quite well! Just a bit slow :)

Best regards.

Avatar
0
ComfortClick Support

Hello Jurgen,


So right out of the bat, when i started sending out commands, the HTTP driver would automatically disconnect, i think it had issues with the model number selection list. If i ran the /api/tags, to get the model number i could get the response for the model number to be : deepseek-r1:latest.. So i created a new string with deepseek-r1:latest as value and use this in the Send command to always try to use the "latest model".

Which seemed fix the issue and all the commands seem to work ok now. It does take quite some time to get an answer (for me it's like 30s.. How's it like on your end?


Best regards.

Avatar
0
Jürgen Jürgenson
Quote from ComfortClick Support

Hello Jurgen,

as always, great work on the integration! Will test it out internally as well and post the example (with your credit of course) to our website for all to use!


Best regards, Support team

It's not done yet, so it's still too raw to share, but I'm happy to let anyone contribute, edit, and test it. I even forgot to add the DeepSeek-R1 model to the list, so there's still a lot of work to do. But I did include all the other DeepSeek models, making it easy to switch between them for testing.

What I love about it is that it runs locally, and it's surprisingly good at answering my stupid questions.

Also, I couldn't get chat streaming to work. It would be awesome if users could see the bot's "Thinking" process, as there's a lot of back-and-forth reasoning happening, and it's fun to watch. Is there a way to get streaming to work with the HTTP driver so it stays connected for about five minutes?


Customer support portal powered by UserEcho