BeeAgent with granite LLM uses BeeAssistantPrompt vs. GraniteBeeAssistantPrompt

[michael-desmond]

Describe the bug
I was noticing that BeeAgent when configured with a granite llm uses the standard BeeAssistantPrompt which conflicts with the granite system prompt and causes the llm to perform poorly. The offending code is here:

bee-agent-framework/src/agents/bee/agent.ts

Line 133 in fd942ae

text: (this.input.templates?.assistant ?? BeeAssistantPrompt).render({

A fairly simple fix is to conditionally choose a default AssistantPrompt before rendering i.e.

const template = (this.input.llm.modelId.includes("granite"))?GraniteBeeAssistantPrompt: BeeAssistantPrompt;

But this is not the most elegant solution imo because it introduces more ad hoc checks into the agent.

Ideally we could move the rendering of the assistant message into the runner hierarchy. The renders method seems to be the most appropriate spot.

https://github.com/i-am-bee/bee-agent-framework/blob/fd942ae0116893b9452530c305e0053bb83f5578/src/agents/bee/runners/default/runner.ts#L256C17-L256C28

However the renders method is introduced on the DefaultRunner as opposed to the BaseRunner which is the type maintained by the BeeAgent.

To Reproduce
Steps to reproduce the behavior:

1. Create a BeeAgent with granite and breakpoint line 130.
2. The BeeAssistantPrompt is applied when rendering the assistant prompt.

Expected behavior
The GraniteBeeAssistantPrompt should be applied when rendering the assistant prompt.

Set-up:

• Bee version: 0.0.47
• Model provider: Ollama

[michael-desmond]

@Tomas2D Can u give an opinion on this one when you get a chance.

[Tomas2D]

Good catch!

Fixed in #223

[Tomas2D]

Released in v0.0.48