#Any ideas on that. It's a major show

1 messages · Page 1 of 1 (latest)

ebon stone
#

Hey @grave turtle creating a 🧵

#

What is the use-case for this? WebHook tools are currently limited by a timeout to prevent long-running actions - curious to know what tools you need require a longer timeout so I can bring this to my team for review.

grave turtle
#

I created my own back end team of agents (calendar agent, sales analyst, task management) and the way to trigger them is via a webhook. That webhook usually returns the info in 30 seconds to up to 2 mins if it's a complex sales query

#

The system works perfectly and I added eleven labs as a voice interaction front end. Works well for any query that takes less than 15 s to return an answer

#

The back end always returns the right answer but since it timed out for the conversation agent , it doesn't return the output

#

Maybe we can collab together and you have a look at my system . Any good tool will need some time to process specially of its complex, without allowing that time , you would be limiting tool usage to very basic requests. Even a perplexity search sometimes takes more than 15s to return an output

#

I think it should be a field that the user can customize depending on the use case . Maybe you can even set it for individual webhooks in the webhook setting .

#

I'm a heavy user so I can definitely share a lot of feedback. When it triggers the webhook it goes silent and only when it receives a webhook response will it acknowledge calling on the webhook for some reason

ebon stone
#

This is great context thank you for the detail!

#

Makes a lot of sense to have tool timeout be configurable.

#

If you have any more ideas to share as you are a heavy-user please feel free to shoot me an e-mail: louis@elevenlabs.io

grave turtle
#

Lemme know if you would want to pilot test this feature. I would be happy to test it out on my Elevenlabs account !

left sluice
#

Hey, I have the same timeout issue. Indeed, it would be good to be able to configure it.
Do you know what is the current value? Thanks!!

#

On the same note, as some requests can take longer, that would be great to configure some "filling words" while the query is sent. Not sure how complex it is on your side
I've tried to add to my prompt Before calling the API say: “Let me check that for you.” (or using the tool name), but the sentence is said when the response is received.

grave turtle
grave turtle
#

as if interupted

ebon stone
#

Thanks all for the feedback! I have aggregated and shared this with the team. Better tool calling Prosody is certainly on our roadmap

#

We will be increasing tool latency once we have worked on the filling words piece.

Do you happen to have a conversation ID which shows off

  1. Before calling the tool, say let me check that for you but this didn't work as the sentence is. said when the tool response is received.
  2. The sentence is sometimes cut in the middle when sending or receiving Webhook response (as if interrupted)

I will then look into this for you

left sluice
#

Hi @ebon stone .
Sure. conversation_id: 2wfh6eBgcXNqSTsQbeJf
at 0:10, the question is fully asked and the api request is sent
at 0:30 you hear "let me check that for you" then directly the response

grave turtle
#

Awesome, let me know if you need another example and I will dig for some

ebon stone
#

Thanks for this! We're working on improving this piece

grave turtle
#

Happy New Year guys. Hope you had a good one. One more thing I noticed. The number of tools you could add seems to be limited. Can we open this up. At the end of the day , it will all depend on the model that is used to call on these tools. Or maybe increase the limit to 15?

grave turtle
#

Hey @ebon stone any luck with your team on these fixes?

ebon stone
#

We've upped the tool limit to 15!

#

@grave turtle There's a few parameters at play here, I built a demo for a Stripe conference last week and have been learning about tool calling prosody myself. Wanted to share my learnings with you:

Tool level prompts:

  • I would isolate the description of the tool & parameters to only define what it does, not how it should respond after the tool is called. Keep these descriptions atomic and standalone.

System-level prompts:

  • This is the best place to describe workflows and patterns, for example when tool X is called say Y. If tool X triggers an event which takes a while you should be able to get it to describe what it's doing.

I have attached an example system prompt I worked on yesterday!

đź“Ž message.txt
#

These are all client-tools (not ideal) but the tool calling flow was pretty good.

#

In terms of timeouts, we have decided to keep the max tool-timeout as is for the time being.

What you can do is have 1 tool trigger the long-running action (for example in my case above it's sendCheckoutSession which sends the user a link to pay in stripe & have another tool (getCustomerDetails) which can be triggered after some time to get the result of the first call.

#

I hope this makes sense!

grave turtle