I am using adk 1.4.0, and I have over 10+ skills. My understanding is, the skills are first selected, then its L2 information is loaded. Then it should use this instruction to perform the actual steps.
However, what I notice is that, the skill instructions are serving almost like an information, not really a required steps, and the LLM is hallucinating.
I ended up reading the skill instruction, and promote that as System Prompt. Then the LLM is producing much better response.
am I doing right thing or is there a way I am missing. First approach is the skill instruction returning as part of tool loop and the instruction is available as tool response.
The second approach is, I am hijacking the response from the tool call, and promoting that as system prompt, for subsequent calls.
I am using adk 1.4.0, and I have over 10+ skills. My understanding is, the skills are first selected, then its L2 information is loaded. Then it should use this instruction to perform the actual steps.
However, what I notice is that, the skill instructions are serving almost like an information, not really a required steps, and the LLM is hallucinating.
I ended up reading the skill instruction, and promote that as System Prompt. Then the LLM is producing much better response.
am I doing right thing or is there a way I am missing. First approach is the skill instruction returning as part of tool loop and the instruction is available as tool response.
The second approach is, I am hijacking the response from the tool call, and promoting that as system prompt, for subsequent calls.