feat: variable passing #51
Description
Variable Passing
Currently, the Large Language Models (LLM) does the heavy lifting of copying data observed by an observation (tool call). This is error prone due to the nature of LLMs. The LLM might predict a different value that was actually observed and there is no guarantee the observed value will end up correctly in the final message. Furthermore, this is error prone to the LLMs changing their capabilities. What might work today won't work tomorrow.
Task
- Discover the capabilities of template/text for variable parsing inside the LLM call. Result should be a description if variable parsing is possible and how good LLMs can do this. E.g. instead of passing the observation value to the LLM, the LLM only knows the variable $OBSERVATION and must generate the $OBSERVATION variable whenever it is needed.
- Implement an exp pkg that does the parsing in the folder
pkg/x
Please discuss any ideas in the chat.
See other
No resources found.