llama3.1 is not stable to generate the executable commands via ros2 ai exec

[fujitatomoya]

Issue Description

Compared to OpenAI gpt-4o ( or gpt-4), llama3.1 (via ollama) is so unstable to generate the executable command based on the user request. the generated answer is kinda off, and even it does not make really sense showed as following.

root@tomoyafujita:~/ros2_ws/colcon_ws# unset OPENAI_API_KEY
root@tomoyafujita:~/ros2_ws/colcon_ws# export OPENAI_MODEL_NAME=llama3.1
root@tomoyafujita:~/ros2_ws/colcon_ws# export OPENAI_ENDPOINT=http://localhost:11434/v1
root@tomoyafujita:~/ros2_ws/colcon_ws# ros2 ai exec "give me all topics" --dry-run
Command Candidate: 'ros2 topic list /clock /tf /tf_static /parameter_events /rosout /rosout_agg /rosgraph/initial_node_config /topic_info /param_changed /class_loader/class_list /cmd_vel /odom /imu_data /joint_states'
root@tomoyafujita:~/ros2_ws/colcon_ws# ros2 ai exec "give me all topics" --dry-run
Command Candidate: 'ros2 topic list /clock'
root@tomoyafujita:~/ros2_ws/colcon_ws# ros2 ai exec "give me all topics" --dry-run
Command Candidate: 'ros2 topic list /clock '
root@tomoyafujita:~/ros2_ws/colcon_ws# ros2 ai exec "give me all topics" --dry-run
Command Candidate: 'ros2 topic list /clock /cmd_vel /image /joint_states /topic /type_support_msgs/string__multiarray____1_5 /rosout / rosgraph /clock /parameter_events /time'

AI model can be really different and network size could be significantly different, since local LLM llama3.1 uses 4.7GB. even though i am not sure how much we can adjust the parameter, it would be worth to try Ollama Modelfile.

Consideration

Originally i though System Role configuration was missing in llama3.1 against gpt-4o but according to https://www.llama.com/docs/model-cards-and-prompt-formats/llama3_1#supported-roles and https://ollama.com/blog/openai-compatibility, both supports the same system roles for the chat completion.