Find answers from the community

Updated 3 months ago

When converting a function into a tool

When converting a function into a tool via FunctionTool() is there a way include any extra prompt along the tool? My tool returns a list of links like: ['<a href="localhost/article/98">The US is proposing ...</a>'], problem is, when the response is synthetised, the LLM removes the markup and returns: [The US is proposing ...](localhost/article/98)

12 comments

LLogan M

hmm, you could try adding extra description to the docstring of the function, or pass in the description manually 🤔

Or maybe you can also add to the system prompt of the agent?

LLogan M

ah you tried the description already. System prompt may help then

i.e agent = OpenAIAgent.from_tools(..., system_prompt="my system prompt")

vvedtam

I thought about the system prompt, but I have multiple tools, so on a per tool basis would be best

vvedtam

will try to include the instruction as a comment in the Fn body

LLogan M

It could be generic too. Something like "When you use a tool, maintain the formatting of the tool output"

vvedtam

thanks, will try bot. So I need to inject it somehow, there's no documented approach for it right now from what I understand

vvedtam

o.O
Including it into the funciton body did the trick:

Plain Text

def latest_news_headlines():
  """Get the latest news headlines. Keep the HTML markup when returning the results."""
...

vvedtam

wondering why having the same instruction in the description didn't have the same effect

vvedtam

Plain Text

latest_news_tool = FunctionTool.from_defaults(
  fn=latest_news_headlines,
  name="latest_news",
  description="Get the latest news headlines. Keep the HTML markup when returning the results.",
)

LLogan M

Hmmm, that is a little suspect haha. I'm not sure either 🤔 Something must be different internally between how the two are presented to the LLM

vvedtam

aham, will dive into the logs

vvedtam

thanks!

Add a reply