Run actions to retrieve data before the LLM begins reasoning.
Place action calls at the top of your reasoning instructions to fetch data before the prompt is constructed. This ensures the LLM has access to current, accurate information when generating responses.
Fetching before reasoning ensures the LLM has accurate, current data. For example, you can store the output of an action in a variable and use it to personalize instructions. Or you can create a filter based on the variable to refine the prompt that’s sent to the LLM. Actions inside reasoning instructions execute before the prompt is sent to the LLM.
Pattern Example: Look up the user’s current order before the conversation starts so the agent can greet them with their order status and personalized recommendations.
1reasoning:
2 instructions: ->
3
4 # Check if data has been fetched
5 if @variables.order_summary == "":
6
7 # If not, fetch data with an action
8 # (and store results in a variable)
9 run @actions.lookup_current_order
10 with member_email=@variables.member_email
11 set @variables.order_summary=@outputs.order_summary
12
13 # Reference the variable in the prompt
14 | Refer to the user by name {!@variables.member_name}.
15 Show them their current order summary: {!@variables.order_summary}.The pattern:
Fetch data and immediately check it to determine what options to present.
1reasoning:
2 instructions: ->
3 if @variables.order_summary == "":
4 run @actions.lookup_current_order
5 with member_email=@variables.member_email
6 set @variables.order_summary=@outputs.order_summary
7
8 | If user wants to make a return:
9 if @variables.order_summary.days_since_order <= 60:
10 set @variables.return_eligibility = true
11 | Offer to process return using {!@actions.create_return}.
12 else:
13 | Politely explain the return period has expired.if @variables.data == "") before making a call to fetch data to avoid running actions unnecessarily.
j