Agent chat credit issue

💬 Agents
1

Hello Retool community.

I've been trying to customize my own agent for the demo app I'm working on.

I added an Agent Chat and configured this message:

{
  "messages": [ 
    { 
      "role": "user", 
      "content": The question you have been asked is {{ agentChat3.agentInputs }}.The patient you need to evaluate is {{ idSelectedPatient_Analytics.value }}. Read that data, evaluate it, and answer
    } 
  ], 
  "action": "invoke" 
}

Instead of returning an answer, it started printing over and over "thinking". I thought it was something I missed in the configuration, but it was actually queueing the same petition, until I ran out of free tokens.

Is there something I can do now / something I missed?

Agent chat component configuration:

image
image1762×774 106 KB

Agent instructions and tools
image
image1068×783 78.3 KB

Tool configuration to read the data from the database
image
image992×677 31.1 KB

Tool function logic
image
image1134×624 40.2 KB

Logs displaying the right output, but all of them triggered every 5~ seconds
image
image1920×995 229 KB

This consummed all my credits, and I didn't get an output in the Agent Chat, only the "thinking" message as I said before.

Thanks a lot

2

For what is worth, I believe that the reason for this is the following:

  • The agent chat component is linked to a query, in this case your agentChat4_query1
  • As designed, if you use agentChat4_query1.agentInputs, it would contain the message, and the invoke property, like:

image
image327×199 10.4 KB

  • once it triggers, it will start polling, and the chart action property will change to getLogs:

    image
    image318×153 9.06 KB

  • It will keep doing it until it gets the results for that agentRunId

  • What is happening with what you're doing, is that the query's data is not changing to getLogs. It is actually triggering the query every time because it is trying to poll, but actually is triggering invoke each time

Not sure if there is a way of capturing the promise in a transformer and link that to the action property in your message setting, so that you can dynamically set whether it is invoke or getLogs

Hope that makes sense

1 Like
3

Hey @MiguelOrtiz, thanks for your answer.

Maybe there is an easier way to do it?

What I wanted was to ask the AI Agent for the patient I was inspecting (saved in the global variable idSelectedPatient_Analytics) and get the results, but it always made me type the patientID, because it couldn't read the global variable that contains it.

That's why I wrote both the agentChat3.agentInputs and the idSelectedPatient_Analytics in the message.

4

Hey @ArturoGonzalez,

As of now, there isn't an easy way, the only work around I can think of, similarly to what was done here is to update a table in your database (rather than setting a global variable) and then add a tool to the agent that fetches the id from that table, and instruct the agent accordingly.

It's not great, but you can get it to work.

5

Thanks a lot @MiguelOrtiz !

I logged in just to ask if that was a possibility haha.

Sadly, there is no way to test this until next month due to the query getting stuck in the invoke state, so I'll need to wait a few days in order to confirm this and share my workaround too.

1 Like