Enhanced Model Management and SystemPrompt Integration in OllamaAgent

[cnupy]

PR Summary

Feature 1: Support for Multiple Models

Description

OllamaAgent can now handle and switch between multiple models seamlessly, listing available models dynamically via an API

Implementation Details

• Modified the agent to include a model selection mechanism that allows it to iterate through predefined models.
• Introduced a model command to manage the switching between models based on available models.

Feature 2: Integration of SystemPrompt

Description

It is possible to configure SystemPrompt with the selected model into OllamaAgent, providing users with a toolkit for task-oriented conversations.

Implementation Details

• Modified the agent to accommodate Ollama's SystemPrompt functionality.
• Introduced a system-prompt command to manage the setting.

Feature 3: Support for Predefined ModelConfigs sets

Description

OllamaAgent now supports the use of predefined model configurations, allowing users to easily switch between different models and system prompts based on specific requirements.

Implementation Details

• Introduced a ModelConfig Record to encapsulate data for each configuration.
• Introduced a config command to manage the switching between configuration based on available predefined sets.

PR Context

This Pull Request enhances the functionality of OllamaAgent by introducing improved model management. By allowing seamless switching between multiple models and supporting predefined configurations, it offers users greater flexibility. The addition of SystemPrompt functionality empowers users with a more task-oriented conversational toolkit, streamlining interactions. These updates not only enhance user experience but also improve the agent's adaptability to diverse requirements. This makes OllamaAgent a more versatile and powerful tool overall.

[daxian-dbw]

@kborowinski @cnupy After upgrading to the v0.5.13 of Ollama, the OllamaAgent stops working even from the main branch ... After sending a query, it's just spinning forever, and after cancel the request, the server side displays the following record:

I moved to OllamaSharp v5.1.4, but it's the same result. Any idea what could be the problem?

[kborowinski]

@daxian-dbw I'm on latest AIShell build from main and Ollama v0.5.13 with llama3.1:8b-instruct-q5_K_M and it works. What model are you running? Has anything changed on your computer where you run AIShell like drivers? How your config looks like?

And this is different computer, graphic card and model phi4-mini:3.8b-q4_K_M, and it works as well:

Try to stop all Ollama processes and then start it again:

Get-Process ollama* | Stop-Process -Force

[cnupy]

Everything's working fine for me as well. I'm using Ollama 0.5.4

/llm/ollama# ./ollama -v
ggml_sycl_init: found 1 SYCL devices:
ollama version is 0.5.4-ipexllm-20250310

[daxian-dbw]

Sorry for the delay!
The ollama agent from the main branch works for me now (I have no idea why it didn't weeks before. I guess something specific to my system). I left some comments, but haven't finished reviewing the Command.cs file. Will finish it tomorrow.

[daxian-dbw]

Done with my 2nd pass of review :)

[daxian-dbw]

@cnupy Thanks for the quick updates. Now only those 4 remaining comments are relevant (I left follow-up comments). All rest looks good.

[cnupy]

@cnupy Thanks for the quick updates. Now only those 4 remaining comments are relevant (I left follow-up comments). All rest looks good.

I've resolved the four remaining comments. Additionally, I renamed the Configs property to Presets for better clarity. I've also carried out some additional refactoring to properly implement the running model checking. Furthermore, I moved and optimized PerformSelfcheck from OllamaAgent to Settings so that it can be utilized in both locations. Let me know if there's anything else you'd like adjusted.

[daxian-dbw]

@cnupy To be more efficient, I went ahead and made some refactoring directly. The main changs are:

1. Update PerformSelfcheck to account for RunningConfig check as well, while making it possible to do endpoint check only too. With this change, we will only need a call to PerformSelfcheck, and then we can directly use _settings.RunningConfig.ModelName.
2. Update EnsureModelsInitialized to not do the endpoint check when the passed-in host is null.
3. Use async actions for commands where they are applicable.

Please review my changes and see if you have any concerns. Thanks!

[daxian-dbw]

BTW, the Ollama server writes out a warning about using the context field:

[GIN] 2025/04/23 - 13:05:47 | 200 | 9.6578815s | 127.0.0.1 | POST "/api/generate"
time=2025-04-23T13:06:59.988-07:00 level=WARN source=routes.go:274 msg="the context field is deprecated and will be removed in a future version of Ollama"

So, the ollama agent code will need to be updated accordingly to avoid breaking (tracked by #375).

[cnupy]

Thanks for the effort. Everything looks good to me! I just have a small suggestion to make the code logic easier to follow.

[cnupy]

the context field is deprecated and will be removed in a future version of Ollama

Yes, the correct way is to use the chat endpoint instead of the generate endpoint (ChatAsync method in the OllamaSharp client), as mentioned previously.

[daxian-dbw]

@cnupy Thank you for the contribution!