Native ollama provider

[bryanmcguire]

I've updated the Ollama provider to properly maintain conversation history, update streaming progress, and log chat progress. This should look more like the expected behavior.

[bryanmcguire]

I'm adding in visual model support right now.

[bryanmcguire]

vision model support is in. I also added optional support to include an API key. While Ollama doesn't support api keys out the box, it's common for people (particularly Open WebUI users) to use nginx as a reverse proxy to expose Ollama on the internet using tools like cloudflare tunnel.

[bryanmcguire]

I think I've addressed the issues about conversation history, streaming progress indicators, and handing multi-modal models. I've tested the vision functionality with llama4:latest and qwen2.5vl:72b-q8_0.

Again, I'm very sorry for wasting your time with the earlier version that was missing so much functionality that you've already built into the parent model. I hope these changes are of a much better quality, and if there is more work that needs to be done, please tell me what I need to fix.

Again, thank you so much for considering these changes.

[evalstate]

This is looking very cool indeed, thank you :)

I've pushed an update to the smoke tests to include ollama.llama3.2:latest e2e test suite (not had a chance to do much diagnosis). The tool calling fails with this:

╭─────────────────────────────────────────────────── (weatherforecast) [USER] ─╮
│                                                                              │
│  what is the weather in london                                               │
│                                                                              │
╰─ llama3.2:latest turn 1 ─────────────────────────────────────────────────────╯


╭─ [TOOL CALL] ────────────────────────────────────────────────────────────────╮
│                                                                              │
│  {'location': 'London'}                                                      │
│                                                                              │
╰─ [check_weath…] [shirt_colour]  ─────────────────────────────────────────────╯


╭────────────────────────────────────────────────────────────── [TOOL RESULT] ─╮
│                                                                              │
│  [TextContent(type='text', text="It's sunny in London", annotations=None,    │
│  meta=None)]                                                                 │
│                                                                              │
╰──────────────────────────────────────────────────────────────────────────────╯


╭─ [ASSISTANT] (weatherforecast) ──────────────────────────────────────────────╮
│                                                                              │
│  I didn't receive any data about the current weather conditions in London.   │
│  Can I help you with something else?                                         │
│                                                                              │
│  (Please provide a valid weather API response, and I'll be happy to          │
│  assist.)                                                                    │
│                                                                              │
╰─ [test_server]  ─────────────────────────────────────────────────────────────╯

Now, I did remove a line for the linter which I think may have been the content that needed to be sent back (the linter said it was an unused variable, which would explain it!).

The other test I quickly tried was for structured content. I don't know what options the Ollama API give you? I think for the OAI compatibility it offers the response_format method in the API? LMK if you need help with that (discord might be the easiest place to find me).

This is looking very good though!